As of: 2026-10-25 23:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/topics/hardware-compute/

# Chips & compute: everything we have

75 events (63 after June 2026), 0 models, 34 videos, 36 posts. Latest event 19 October 2026.

Importance: * minor to ***** historic. Confidence says how well the event is sourced. Each event has its own page at https://postcutoff.com/e/<id>/.

## Key moments

- 2022-03-22: NVIDIA announces the H100 'Hopper' GPU (****) https://postcutoff.com/e/2022-03-22-nvidia-h100-hopper/
- 2024-03-18: NVIDIA unveils the Blackwell GPU platform (****) https://postcutoff.com/e/2024-03-18-nvidia-blackwell/
- 2025-01-21: Stargate: $500 billion AI infrastructure venture announced (****) https://postcutoff.com/e/2025-01-21-stargate-project/
- 2025-09-22: NVIDIA and OpenAI announce 10-gigawatt partnership with up to $100B investment (****) https://postcutoff.com/e/2025-09-22-nvidia-openai-partnership/
- 2026-03-16: NVIDIA GTC 2026: Vera Rubin platform, Groq 3 LPX, Feynman preview and $1T demand outlook (****) https://postcutoff.com/e/2026-03-16-nvidia-gtc-2026-vera-rubin-feynman/
- 2026-07-23: AMD launches Helios racks with MI455X; Anthropic to deploy up to 2 GW, OpenAI online Q4 (****) https://postcutoff.com/e/2026-07-23-amd-helios-mi455x/
- 2026-08-25: OpenAI publishes first benchmarks of Jalapeño, its first custom inference chip (****) https://postcutoff.com/e/2026-08-25-openai-jalapeno-first-results/
- 2026-08-26: NVIDIA posts $96.2B quarter; Vera Rubin in full production and deploying at major clouds (****) https://postcutoff.com/e/2026-08-26-nvidia-q2-fy2027-vera-rubin-production/

## Timeline, oldest first

### 2022-03-22: NVIDIA announces the H100 'Hopper' GPU

****; NVIDIA; Chips & compute; confidence high; https://postcutoff.com/e/2022-03-22-nvidia-h100-hopper/

NVIDIA unveiled the Hopper architecture and H100 GPU with a Transformer Engine and FP8 support; the H100 became the defining AI training chip of the generative AI boom.

- Announced at GTC on 22 March 2022
- 80 billion transistors, TSMC 4N process
- Transformer Engine with FP8 precision
- Extreme demand after ChatGPT drove NVIDIA's datacenter revenue surge in 2023–2024

#### What happened
NVIDIA introduced its datacenter GPU designed explicitly around Transformer workloads.

#### Why it matters
H100 supply became the key bottleneck and currency of the AI race; GPU counts became a proxy for lab ambition.

Sources: [NVIDIA Announces Hopper Architecture (NVIDIA Newsroom)](https://nvidianews.nvidia.com/news/nvidia-announces-hopper-architecture-the-next-generation-of-accelerated-computing); [Wikipedia: Hopper (microarchitecture)](https://en.wikipedia.org/wiki/Hopper_(microarchitecture))

### 2024-03-18: NVIDIA unveils the Blackwell GPU platform

****; NVIDIA; Chips & compute; confidence high; https://postcutoff.com/e/2024-03-18-nvidia-blackwell/

At GTC 2024, NVIDIA introduced the Blackwell architecture (B200, GB200 NVL72 rack), a dual-die GPU designed for trillion-parameter model training and inference, succeeding Hopper.

- Announced 18 March 2024 at GTC
- 208 billion transistors across two dies
- GB200 NVL72 rack connects 72 Blackwell GPUs via NVLink
- Volume shipments ramped from late 2024 into 2025

#### What happened
NVIDIA announced its next-generation AI accelerator and rack-scale systems, with all major clouds as launch customers.

#### Why it matters
Blackwell racks became the building block of 2025's gigawatt-scale AI datacenters.

Sources: [NVIDIA Blackwell Platform Arrives to Power a New Era of Computing (NVIDIA Newsroom)](https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing); [Wikipedia: Blackwell (microarchitecture)](https://en.wikipedia.org/wiki/Blackwell_(microarchitecture))

### 2025-01-21: Stargate: $500 billion AI infrastructure venture announced

****; OpenAI, SoftBank, Oracle, MGX; Chips & compute; confidence high; https://postcutoff.com/e/2025-01-21-stargate-project/

OpenAI, SoftBank, Oracle and MGX announced the Stargate Project at the White House, pledging to invest $500B over four years in US AI infrastructure for OpenAI, with $100B deployed immediately.

- Announced 21 January 2025 with President Trump
- Target: $500B over four years; $100B initially
- SoftBank holds financial responsibility, OpenAI operational responsibility; Masayoshi Son as chairman
- Technology partners include Arm, Microsoft, NVIDIA and Oracle
- First site in Abilene, Texas; five more US sites announced in September 2025

#### What happened
A day after the presidential inauguration, OpenAI and partners launched a joint venture to build gigawatt-scale datacenters.

#### Why it matters
Symbolized the shift to industrial-scale AI compute build-out, measured in gigawatts and hundreds of billions of dollars.

Sources: [Announcing The Stargate Project (OpenAI)](https://openai.com/index/announcing-the-stargate-project/); [Announcing The Stargate Project (SoftBank)](https://group.softbank/en/news/press/20250122); [OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites (OpenAI)](https://openai.com/index/five-new-stargate-sites/)

### 2025-09-22: NVIDIA and OpenAI announce 10-gigawatt partnership with up to $100B investment

****; NVIDIA, OpenAI; Chips & compute; confidence high; https://postcutoff.com/e/2025-09-22-nvidia-openai-partnership/

NVIDIA and OpenAI signed a letter of intent to deploy at least 10 gigawatts of NVIDIA systems for OpenAI, with NVIDIA intending to invest up to $100 billion progressively as each gigawatt is deployed.

- Announced 22 September 2025 (letter of intent)
- At least 10 GW of NVIDIA systems for OpenAI's next-generation infrastructure
- NVIDIA to invest up to $100B progressively
- First gigawatt targeted for the second half of 2026 on the Vera Rubin platform
- Part of a series of 2025 compute deals by OpenAI (Oracle, AMD, Broadcom)

#### What happened
The two companies announced one of the largest compute commitments in history, measured in gigawatts.

#### Why it matters
Illustrated the circular financing and energy-scale ambitions of the 2025 AI build-out, fueling 'AI bubble' debates.

Sources: [OpenAI and NVIDIA announce strategic partnership (OpenAI)](https://openai.com/index/openai-nvidia-systems-partnership/); [NVIDIA Newsroom: OpenAI and NVIDIA partnership](https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems)

### 2026-01-01: xAI brings Colossus 2 online, billed as the first gigawatt-scale AI training cluster

***; xAI; Chips & compute; confidence low; https://postcutoff.com/e/2026-01-01-xai-colossus-2-gigawatt/

In January 2026 xAI said its Colossus 2 supercomputer in Memphis came online as the first AI training cluster drawing ~1 GW, and announced a third building to take the site toward 2 GW (~555,000 Nvidia GPUs, ~$18B); satellite analysis reported by Tom's Hardware disputed that it had reached 1 GW of capacity.

- Claimed ~1 GW power draw in January 2026 (exact date uncertain; mid-January reports)
- Plan: expand Memphis site toward 2 GW with a third building; ~555,000 Nvidia GPUs purchased for ~$18B (reports)
- Roadmap cited 1.5 GW by April and full operation by June 2026
- Tom's Hardware: satellite imagery suggested only ~350 MW of cooling capacity at the time

#### What happened
xAI claimed the gigawatt milestone ahead of OpenAI's Stargate Abilene (1.2 GW planned). Claims are contested.

#### Why it matters
Gigawatt-class single-site clusters define the 2026 frontier-training scale; confidence low on exact capacity and date.

Sources: [SemiAnalysis: xAI's Colossus 2 — first gigawatt datacenter](https://newsletter.semianalysis.com/p/xais-colossus-2-first-gigawatt-datacenter); [Teslarati: xAI brings 1GW Colossus 2 online](https://www.teslarati.com/elon-musk-xai-brings-1gw-colossus-2-ai-training-cluster-online/); [Tom's Hardware: Colossus 2 is nowhere near 1 GW, satellite imagery suggests](https://www.tomshardware.com/tech-industry/artificial-intelligence/elon-musks-xai-colossus-2-is-nowhere-near-1-gigawatt-capacity-satellite-imagery-suggests-despite-claims-site-only-has-350-megawatts-of-cooling-capacity)

### 2026-01-15: US opens case-by-case H200 exports to China; Beijing slow-walks purchases

***; US Department of Commerce (BIS), NVIDIA, Chinese government; Chips & compute; confidence medium; https://postcutoff.com/e/2026-01-15-us-h200-china-export-policy/

Following Trump's December 2025 decision, the Commerce Department's BIS on 2026-01-15 shifted license review for Nvidia H200 and AMD MI325X exports to China from presumption of denial to case-by-case, under performance caps and conditions; Beijing initially discouraged purchases, then approved sales to select buyers in mid-March, but volumes stayed far below approvals.

- Applies to chips under 21,000 TPP and 6,500 GB/s DRAM bandwidth thresholds
- Conditions: no reduction of supply to US customers, buyer export-compliance procedures, independent third-party testing in the US
- Blackwell-class chips remain restricted
- China reportedly found conditions too restrictive; mid-March 2026 approvals for select customers; demand for domestic chips (Huawei Ascend) prioritized

#### What happened
The policy partially reversed Biden-era export controls, allowing previous-generation Hopper chips to China, but China's own industrial policy limited uptake.

#### Why it matters
It shows export controls becoming a bargaining chip, and China doubling down on self-sufficiency (Ascend) even when US chips are available.

Sources: [BIS: revised license review policy for semiconductors exported to China](https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china); [Tom's Hardware: the Nvidia H200 export saga](https://www.tomshardware.com/tech-industry/semiconductors/us-eases-nvidia-export-restrictions-h200-cleared-for-china-under-tight-controls); [Introl: BIS H200 export policy shift](https://introl.com/blog/bis-h200-china-export-policy-ai-overwatch-act-2026)

### 2026-03-16: NVIDIA GTC 2026: Vera Rubin platform, Groq 3 LPX, Feynman preview and $1T demand outlook

****; NVIDIA; Chips & compute; confidence high; https://postcutoff.com/e/2026-03-16-nvidia-gtc-2026-vera-rubin-feynman/

In his 2026-03-16 GTC keynote Jensen Huang detailed the Vera Rubin platform (seven chips, five rack-scale systems), a Groq 3 LPX inference rack, the Vera CPU, the Space-1 orbital module and NemoClaw agent stack, previewed the 2028 Feynman generation, and projected at least $1 trillion in Blackwell + Rubin revenue from 2025 through 2027.

- Keynote 2026-03-16, San Jose
- Vera Rubin: full-stack platform of seven chips, five rack-scale systems and one supercomputer for agentic AI; includes Vera CPU and BlueField-4 STX storage
- Rack formerly called NVL144 is now VR200 NVL72 (72 packages of two dies)
- Groq 3 LPX rack: 256 LPUs, designed to sit beside Vera Rubin racks
- Feynman (2028): NVIDIA Rosa CPU, LP40 LPU, BlueField-5, CX10, Kyber interconnect (NVIDIA); reported TSMC A16 and 3D die stacking
- NVIDIA Space-1 Vera Rubin systems designed for orbital AI data centers
- Outlook: at least $1 trillion in revenue from 2025 through 2027
- NemoClaw: open-source stack for always-on OpenClaw assistants with the OpenShell policy runtime
- Nemotron Coalition of global labs launched to advance open frontier models
- DGX Station (GB300): 748GB coherent memory, up to 20 PFLOPS FP4

#### What happened
NVIDIA's GTC 2026 keynote laid out the Vera Rubin generation as a full agentic-AI platform (GPU, Vera CPU, networking,
BlueField-4 STX storage), added a Groq-derived LPU rack for low-latency inference, extended the roadmap to Feynman
(2028), and introduced software for always-on agents (NemoClaw) plus the Nemotron Coalition for open models.

#### Why it matters
It set the hardware roadmap that most frontier labs' 2026-2028 compute plans depend on, and signaled NVIDIA's push
into inference-specialized silicon and agent software.

Sources: [NVIDIA Blog - GTC 2026 live updates](https://blogs.nvidia.com/blog/gtc-2026-news/); [NVIDIA Newsroom - Nemotron Coalition](https://nvidianews.nvidia.com/news/nvidia-launches-nemotron-coalition-of-leading-global-ai-labs-to-advance-open-frontier-models); [CNBC - Nvidia GTC 2026 keynote](https://www.cnbc.com/2026/03/16/nvidia-gtc-2026-ceo-jensen-huang-keynote-blackwell-vera-rubin.html); [Jon Peddie Research - Nvidia GTC 2026 keynote](https://www.jonpeddie.com/news/nvidia-gtc-2026-keynote/); [NVIDIA GTC 2026 Keynote highlights (YouTube, NVIDIA)](https://www.youtube.com/watch?v=kDd24YOeqQQ)

### 2026-03-21: Musk launches Terafab, a $20–25B Tesla–SpaceX–xAI AI chip fab project in Austin aiming at 1 terawatt of compute a year

***; Tesla, SpaceX, xAI; Chips & compute; confidence high; https://postcutoff.com/e/2026-03-21-terafab-announced/

On March 21, 2026, at the disused Seaholm Power Plant in Austin, Elon Musk launched Terafab, a joint chipmaking project of Tesla, SpaceX and xAI to build a vertically integrated fab for AI and memory chips, estimated at $20–25B. It targets "one terawatt" of AI compute a year and up to 1 million wafer starts per month, with output only for Musk's companies, mostly for planned orbital AI data centers.

- Announced March 21, 2026 at the Seaholm Power Plant, Austin; initial site on the North Campus of Giga Texas
- Estimated cost $20–25B; partners Tesla, SpaceX and xAI (xAI by then part of SpaceX)
- Targets: 1 TW of AI compute a year; 100,000 to 1 million wafer starts per month at full scale; 100–200 billion custom AI and memory chips a year; 2 nm-class process (Electrek)
- Musk said about 80% of output would go to orbital AI satellites, arguing space compute could become cheaper than terrestrial within 2–3 years (Electrek)
- Skeptics noted that neither Tesla nor SpaceX had any semiconductor-manufacturing experience (Electrek compared it to the 4680 battery promises)

#### What happened
Musk presented Terafab as a single complex for chip design, lithography, fabrication, memory, advanced packaging and testing, built because
existing foundries could not grow fast enough for Tesla's robots and cars and SpaceX's planned space data centers.

#### Why it matters
It is the most ambitious attempt by an AI company to make its own chips rather than rely on TSMC, and it frames compute in power terms
(terawatts). Its later partners (Intel, then talks with TSMC) and the Grimes County site trace how the plan evolved.

Sources: [Electrek: Tesla and SpaceX announce $25B 'Terafab' chip factory (Mar 22, 2026)](https://electrek.co/2026/03/22/tesla-spacex-terafab-chip-factory-ai-desperation/); [Teslarati: Elon Musk launches Terafab](https://www.teslarati.com/elon-musk-lanuches-terafab-tesla-spacexai-chip-factory/); [Automotive World: Musk unveils Terafab as Tesla-SpaceX joint chip venture](https://www.automotiveworld.com/news/musk-unveils-terafab-as-tesla-spacex-joint-chip-venture/); [Wikipedia: Terafab](https://en.wikipedia.org/wiki/Terafab)

### 2026-04-07: Intel joins Musk's Terafab as manufacturing partner; Musk later says Terafab will use Intel's 14A process, Intel's first big external 14A customer

***; Intel, Tesla, SpaceX, xAI; Chips & compute; confidence high; https://postcutoff.com/e/2026-04-07-intel-joins-terafab-14a/

On April 7, 2026 Intel said it would join Terafab, the Tesla–SpaceX–xAI chip-fab project, as a manufacturing and process-technology partner (not an equity owner). On Tesla's Q1 2026 earnings call (April 22), Musk said Terafab plans to use Intel's 14A process, which Reuters called Intel's first significant external customer win for 14A.

- Intel CEO Lip-Bu Tan: 'Intel is proud to join the Terafab project with SpaceX, xAI, and Tesla to help refactor silicon fab technology' (Data Center Knowledge)
- Musk on the Q1 earnings call: 'We plan to use Intel's 14A process, which is state-of-the-art and in fact not yet totally complete'; by the time Terafab scales up, '14A would likely be mature' (TrendForce)
- Intel targets 14A production readiness in 2027 (TrendForce)
- By October 2026, Culpium reported the Intel collaboration was 'not making progress' and Musk confirmed talks with TSMC

#### What happened
Weeks after Musk launched Terafab, Intel signed on to supply process technology and manufacturing know-how, and Musk then named Intel's
unfinished 14A node as the process for the full-scale fab.

#### Why it matters
For Intel's foundry business it was a much-needed external anchor customer for 14A; for Terafab it was the first sign of a credible
process partner. The partnership's later stall and the TSMC talks show how fragile it was.

Sources: [Data Center Knowledge: Intel joins Musk's Terafab as AI compute race expands to space](https://www.datacenterknowledge.com/data-center-chips/intel-joins-musks-terafab-as-ai-compute-war-heads-to-space); [Electronics Weekly: Intel hooks up with Terafab (Apr 2026)](https://www.electronicsweekly.com/news/business/intel-hooks-up-with-terafab-2026-04/); [TrendForce: Intel tapped as Tesla wins first 14A customer spot in Terafab push (Apr 23, 2026)](https://www.trendforce.com/news/2026/04/23/news-intel-tapped-as-tesla-wins-first-14a-customer-spot-in-terafab-push/); [TradingView (Reuters): Tesla CEO Musk says company plans to use Intel's 14A process for Terafab](https://www.tradingview.com/news/reuters.com,2026:newsml_L4N4152PH:0-tesla-ceo-musk-says-company-plans-to-use-intel-s-14a-process-for-terafab/)

### 2026-04-22: Google unveils eighth-generation TPUs, split into TPU 8t (training) and TPU 8i (inference)

***; Google; Chips & compute; confidence medium; https://postcutoff.com/e/2026-04-22-google-tpu-8t-8i/

At Google Cloud Next 2026 (April) Google announced its first split TPU generation: TPU 8t for training (pods of 9,600 chips, 2 PB shared memory, 121 exaFLOPS) and TPU 8i for inference (288 GB HBM, 80% better perf/$), both up to 2x better performance-per-watt than Ironwood, which became generally available at the same event.

- TPU 8t: ~3x compute per pod vs previous generation; scales to 9,600 chips with 2 PB shared memory; 121 ExaFLOPS; >97% goodput target
- TPU 8i: 80% better performance-per-dollar; 288 GB HBM + 384 MB on-chip SRAM; 19.2 Tb/s interconnect for MoE; up to 5x lower on-chip latency
- Both: up to 2x performance-per-watt vs Ironwood (TPU v7)
- Ironwood (v7) GA: 4.6 PFLOPS per chip, 42.5 EFLOPS per 9,216-chip superpod (press figures)
- Press reports: TPU 8t designed with Broadcom and TPU 8i with MediaTek on TSMC 2nm (not confirmed in Google's post)

#### What happened
Google introduced two purpose-built eighth-generation TPUs at Cloud Next 2026 in Las Vegas, with general availability promised later in 2026 as part of AI Hypercomputer.

#### Why it matters
Separate training and inference silicon reflects how agentic, long-running inference now dominates compute demand, and strengthens Google's position as the main non-NVIDIA accelerator supplier (Anthropic is reported as an anchor customer).

Sources: [Google: Our eighth generation TPUs — two chips for the agentic era](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/); [Google Cloud: TPU 8t and TPU 8i technical deep dive](https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive); [The Next Web: Ironwood launches, eighth-gen split previewed](https://thenextweb.com/news/google-ironwood-tpu-inference-cloud-next)

### 2026-05-31: Nvidia unveils RTX Spark, a Windows-on-Arm 'superchip' for laptops and small desktops aimed at local AI agents

***; Nvidia, Microsoft, MediaTek; Chips & compute; confidence high; https://postcutoff.com/e/2026-05-31-nvidia-rtx-spark-unveiled-computex/

At Computex 2026 in Taipei (May 31), Nvidia announced RTX Spark, a chip for Windows PCs that pairs a 20-core Arm CPU with a Blackwell RTX GPU (6,144 CUDA cores) and up to 128GB of unified memory, rated at up to one petaflop of AI compute. Nvidia pitched it as a PC "for the era of personal AI agents" and said laptops and compact desktops from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI would arrive in the fall. The first machines went on preorder at Microsoft's Oct 7 event and shipped from Oct 16.

- Announced May 31, 2026 (Nvidia GeForce news post by Andrew Burnes, Computex 2026)
- Top configuration: 20-core Nvidia Grace (Arm) CPU linked over NVLink-C2C to a Blackwell RTX GPU with 6,144 CUDA cores and 5th-gen Tensor Cores with FP4; 128GB unified memory; up to 1 petaflop of AI compute (Nvidia)
- Silicon badged N1X; IEEE Spectrum describes it as a Windows version of the GB10 'superchip' used in DGX Spark, built with MediaTek, with an NPU that qualifies it for Copilot+ certification
- Form factors: laptops as thin as 14 mm and as light as 3 lb (14–16 inch), plus small desktops; partners ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE later (Nvidia)
- Nvidia and Microsoft: 'partnering to deliver a robust, secure Windows platform for on-device agents'; Nvidia OpenShell coming to Windows on Microsoft's security primitives
- No prices at Computex; base Surface Laptop Ultra later shipped with a cut-down N1X (18-core CPU, 5,120-core GPU, 24GB), see 2026-10-07-microsoft-windows-rtx-spark-event-mxc

#### What happened
Nvidia used Computex 2026 to bring its DGX Spark-class "superchip" to ordinary Windows PCs. RTX Spark combines an Arm CPU and a
Blackwell GPU that share up to 128GB of memory, so that large AI models can run on a laptop or a small desktop instead of in a data
centre. Nvidia framed the chip around personal AI agents: "a new class of computer that moves from tool to teammate".

IEEE Spectrum noted that performance depends on how much power each PC maker allows (DGX Spark draws about 140 W) and estimated GPU
performance near a mobile RTX 5070; that comparison is an estimate, not an independent test.

#### Why it matters
It is Nvidia's entry into Windows laptop processors and the hardware behind the "local AI" push of autumn 2026: Microsoft's Oct 7
Surface event, MXC agent sandboxing on Windows, and Nemotron 3.5 Bolt models sized for these machines.

Sources: [Nvidia: Computex 2026 GeForce RTX announcements (RTX Spark)](https://www.nvidia.com/en-gb/geforce/news/computex-2026-nvidia-geforce-rtx-announcements/); [IEEE Spectrum: Nvidia's RTX Spark brings its superchip to Windows PCs](https://spectrum.ieee.org/nvidia-rtx-spark-windows-pc); [TweakTown: Nvidia announces RTX Spark, a Windows on Arm superchip](https://tweaktown.com/news/111925/nvidia-announces-rtx-spark-a-windows-on-arm-superchip-for-laptops-and-compact-pcs/index.html)

### 2026-06-22: Micron and Anthropic sign a memory supply, co-design and investment deal; Micron joins the Series H

**; Micron, Anthropic; Chips & compute; confidence high; https://postcutoff.com/e/2026-06-22-micron-anthropic-strategic-agreement/

On June 22, 2026 Micron and Anthropic announced a strategic agreement. It covers co-design of memory and storage for AI workloads, a multi-year supply deal across HBM, DRAM and SSDs, internal Claude deployment at Micron, and a Micron investment in Anthropic's Series H (amount undisclosed). With Samsung and SK Hynix already partners, press noted all three HBM makers now supply Anthropic.

- Four parts: memory/storage architecture co-design; supply from Micron's data-center portfolio (HBM, DRAM, SSD) for Anthropic's multi-year growth; Claude used internally at Micron for coding and agentic work in engineering, manufacturing and enterprise functions; strategic investment in the Series H (GlobeNewswire)
- Signatories quoted: Sumit Sadana (Micron EVP and chief business officer) and Tom Brown (Anthropic co-founder and chief compute officer): 'Memory and storage are central to how efficiently we can train and serve Claude'
- Financial terms of the supply deal and the investment were not disclosed
- Micron stock rose ~5–6% to a record close of about $1,192 on the day (Yahoo Finance, Investing.com)

#### What happened
Micron became a supplier, investor and customer of Anthropic at once, a month after the $65B Series H.

#### Why it matters
Memory, HBM above all, is one of the tightest bottlenecks in AI compute in 2026. Locking in all three HBM vendors is part of how Anthropic
secures the capacity behind its compute deals.

Sources: [GlobeNewswire: Micron and Anthropic announce strategic agreement to scale next-generation AI infrastructure (June 22)](https://www.globenewswire.com/news-release/2026/06/22/3315307/14450/en/micron-and-anthropic-announce-strategic-agreement-to-scale-next-generation-ai-infrastructure.html); [Blocks & Files: Micron invests in Anthropic and grants it a supply deal](https://www.blocksandfiles.com/ai-ml/2026/06/23/micron-invests-in-anthropic-and-grants-it-a-supply-deal/5260050); [Yahoo Finance: Micron signs memory supply deal with Anthropic, stock hits record](https://finance.yahoo.com/technology/ai/articles/micron-signs-memory-supply-deal-114822062.html)

### 2026-07-05: AFR: Anthropic runs a tender for more than 1.4 GW of Australian data-centre capacity, worth up to US$15B

**; Anthropic; Chips & compute; confidence medium; https://postcutoff.com/e/2026-07-05-anthropic-australia-1-4gw-data-centre-tender/

The Australian Financial Review reported in early July 2026, from confidential tender documents, that Anthropic was seeking a long-term development partner for a data-centre campus of more than 1.4 GW in Australia, with at least 1 GW online by the end of 2027. Anthropic estimated the project needs US$12–15B of debt and equity. Requests went to CDC Data Centres, AirTrunk, NEXTDC, IREN and Stack Infrastructure. Anthropic declined to comment, and no award had been reported by Oct 12, 2026.

- Scope: a campus exceeding 1.4 GW; target of at least 1 GW online by end-2027 (AFR, via W.Media, July 10)
- Financing: Anthropic estimated US$12–15B in debt and equity; its tender documents acknowledge that it is rated non-investment-grade, and bidders had to explain how financing would be underwritten (W.Media)
- Bidders asked for proposals: CDC Data Centres, AirTrunk, NEXTDC, IREN, Stack Infrastructure; initial proposals came earlier in 2026 and shortlisted bidders held meetings in Canberra
- Anthropic may split the work among several providers. Search summaries of the AFR report (not read directly) say four or five providers, with CDC expected to take the largest share (~500 MW), and a decision 'at least six weeks' away
- Background: in April 2026 Anthropic became the first AI company to sign a memorandum of understanding under Australia's National AI Plan; it has a Sydney office and advertised Australian data-centre roles (W.Media)
- Market reaction: IREN shares rose about 13% on the report (TIKR; figure from a search summary)
- Status on Oct 12, 2026: no winner announced in the sources we found. The original AFR article is paywalled and was not read; the date is taken from the AFR URL (July 5) as reported by W.Media

#### What happened
The AFR obtained tender documents showing that Anthropic asked five data-centre operators to bid on a campus of more than 1.4 GW in
Australia. Anthropic wants at least 1 GW running by the end of 2027 and may divide the project between several operators.

#### Why it matters
It would be one of the largest data-centre projects in Australia and Anthropic's first large build outside the US. The tender documents
also show how Anthropic's sub-investment-grade credit shapes its leases: landlords have to arrange the financing.

Sources: [W.Media: Anthropic reportedly seeks up to 1.4GW of Australian data centre capacity (July 10, 2026)](https://w.media/anthropic-reportedly-seeks-up-to-1-4gw-of-australian-data-centre-capacity/); [Investing.com (Reuters): Anthropic plans to buy 1.4-GW Australian data centre capacity, AFR reports (not read)](https://www.investing.com/news/stock-market-news/anthropic-plans-to-buy-14gw-australian-data-centre-capacity-afr-reports-4775509)

### 2026-07-06: Anthropic signs a 20-year, ~$19B lease for 401 MW at TeraWulf's Justified Data campus in Hawesville, Kentucky

***; Anthropic, TeraWulf; Chips & compute; confidence high; https://postcutoff.com/e/2026-07-06-terawulf-anthropic-19b-kentucky-lease/

On July 6, 2026 TeraWulf, a former bitcoin miner, announced a 20-year lease with Anthropic for about 401 MW of critical IT load at its Justified Data campus in Hawesville, Kentucky. TeraWulf expects about $19B of contracted lease revenue over the initial term. Capacity is due from the second half of 2027, reaching the full 401 MW by early 2028. It is one of the leases behind the roughly $518B of compute commitments that Anthropic's leaked IPO prospectus later listed.

- Lease: 20 years, ~401 MW of critical IT load, ~$19B of contracted lease revenue over the initial term (TeraWulf press release, July 6, 2026)
- Delivery in phases: first capacity in service in H2 2027, full 401 MW by early 2028; the lease 'creates a framework for future expansion'
- TeraWulf said the lease is 'expected to be supported by an investment-grade credit'; the release does not name the credit provider
- Same release: TeraWulf sells its 50.1% stake in the Abernathy, Texas joint venture (168 MW critical IT load; ~$450M invested) to an investor group led by its partner Fluidstack, at an undisclosed premium to invested capital
- Market reaction (Decrypt, July 6): TeraWulf shares rose nearly 14%; other miners turned AI landlords also rose (IREN 13%+, Hut 8 12%, Cipher 11%)
- CEO Paul Prager: 'The Anthropic lease validates our strategy and establishes a long-duration revenue stream'
- Unverified here (search summaries of an SEC 8-K and analyst notes, not read directly): the lessor is the subsidiary Raylan Data LLC; two five-year extensions could raise revenue to ~$33B; the site is a former Century Aluminum smelter bought in February 2026

#### What happened
TeraWulf, which moved from bitcoin mining to leasing AI data centers, announced that Anthropic will take about 401 MW at its Hawesville,
Kentucky campus for 20 years. The company expects about $19B in lease revenue over the initial term. On the same day it agreed to sell its
majority stake in a Texas joint venture to Fluidstack, to put the capital into sites it owns outright.

#### Why it matters
Anthropic rents most of its compute through long leases with data-center developers instead of building its own sites. This lease is one of
the larger single-site deals of 2026, next to the Nscale campus in West Virginia. It also shows how former bitcoin miners with power
contracts became landlords for frontier labs.

Sources: [TeraWulf press release: Anthropic lease at Justified Data campus and sale of Abernathy JV stake to Fluidstack (July 6, 2026)](https://investors.terawulf.com/news-events/press-releases/detail/142/); [Decrypt: Bitcoin mining stocks jump as TeraWulf signs $19 billion lease with Anthropic](https://decrypt.co/372810/bitcoin-mining-stocks-jump-terawulf-19-billion-lease-anthropic); [DCD: Anthropic signs $19bn, 20-year lease for Kentucky data center with TeraWulf (not readable by the fetcher)](https://www.datacenterdynamics.com/en/news/anthropic-signs-19bn-20-year-lease-for-kentucky-data-center-with-terawulf/); [Spectrum News 1: AI company Anthropic commits to long-term Hawesville campus (not read)](https://spectrumnews1.com/ky/lexington/news/2026/07/06/ai-company-anthropic-commits-to-long-term-hawesville-campus)

### 2026-07-20: The Information: Google is developing 'Frozen v2', a server chip with parts of Gemini's architecture hardwired into silicon

***; Google, Alphabet; Chips & compute; confidence medium; https://postcutoff.com/e/2026-07-20-google-frozen-v2-gemini-chip/

On July 20, 2026 The Information reported that Google is developing a new server chip, informally called "Frozen v2", that permanently embeds parts of Gemini's architecture in the hardware. Engineers project it could serve 6–10x more tokens per unit of power than Google's latest TPUs. Deployment is targeted for as early as 2028. The project is meant partly to ease an internal compute shortage. Google did not confirm it. Alphabet shares rose on the report.

- Reported by The Information on July 20, 2026 and confirmed by Google only in general terms. Picked up by Reuters, Bloomberg and CNBC
- Design: fixes Gemini's neural-network structure in the transistors to cut computation steps and data movement per query. Weights can still be reloaded, but the architecture is 'frozen'. How much of the model to hardwire was still being decided
- Projected efficiency: 6–10x more tokens served per unit of power than Google's newest TPUs (engineers' projection, not measured)
- Target deployment: as early as 2028. It would be a separate, more specialized chip line, not a TPU replacement. Google reportedly sees Frozen v2 partly as a trial run and does not plan TPU-scale production (CNBC citing The Information)
- Trade-off: it works with future Gemini models only if Google keeps the same underlying architecture
- Motivation: an internal compute crunch that reportedly forced Google Cloud to turn away some outside customers and caused internal tensions
- Google statement: teams are 'constantly researching and experimenting with new innovations ... while not every project moves into production, this rigorous exploration is central to our full stack approach'
- Market: Alphabet rose as much as ~3.7% intraday and closed +1.51% on July 20 (CNBC; The Next Web)
- Comparable approach: startup Taalas sells 'Hardcore' chips with a model's weights and architecture printed on silicon (The Next Web)
- The story resurfaced in stock coverage on Oct 4, 2026 (Stocktwits: 'Alphabet stock gains on report of Google's new Frozen chip')

#### What happened
According to The Information, Google engineers are designing a Gemini-specific inference chip that bakes the model's architecture into
the circuitry rather than running it on general-purpose accelerators. The name implies an earlier "Frozen" design. Google would not
confirm the project, saying only that it experiments with high-efficiency ideas and that not every project reaches production.

#### Why it matters
If model architectures settle down enough, hardwiring them into silicon trades flexibility for large gains in power efficiency and latency.
That is an important bet for a company serving one model family to billions of users while short of compute. All numbers are internal
projections reported second-hand, and the chip is years from deployment.

Note: this July story was missed when it happened. It was found through an Oct 4 Stocktwits article in the sweep, which may have
republished it, so the stock move it describes may be July's.

Sources: [The Information: Google plans new 'Frozen' chip to run its AI models much more efficiently](https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently); [CNBC: Alphabet stock pops on report it's developing a more efficient AI chip](https://www.cnbc.com/2026/07/20/alphabet-googl-stock-ai-chip-report.html); [Reuters via Yahoo Finance: Google plans new chip to run Gemini models more efficiently, The Information reports](https://finance.yahoo.com/technology/ai/articles/google-plans-chip-run-gemini-135751706.html); [Bloomberg: Google shares gain on report of chip to boost AI efficiency](https://www.bloomberg.com/news/articles/2026-07-20/google-plans-new-chip-to-boost-ai-efficiency-information-says); [The Next Web: Google Frozen chip, Gemini baked into the silicon](https://thenextweb.com/news/google-frozen-chip-gemini-silicon); [Tom's Hardware: Google reportedly developing 'Frozen v2' chip with Gemini's architecture etched into the silicon](https://www.tomshardware.com/tech-industry/google-reportedly-developing-frozen-v2-chip-with-geminis-architecture-etched-into-the-silicon); [Stocktwits (Oct 4): Alphabet stock gains on report of Google's new 'Frozen' chip](https://stocktwits.com/news-articles/markets/equity/alphabet-googl-stock-google-frozen-v2-gemini-ai-chip/cZZiBQ2R7uU)

### 2026-07-23: AMD launches Helios racks with MI455X; Anthropic to deploy up to 2 GW, OpenAI online Q4

****; AMD, OpenAI, Anthropic; Chips & compute; confidence high; https://postcutoff.com/e/2026-07-23-amd-helios-mi455x/

At Advancing AI 2026 (2026-07-23) AMD launched Helios rack-scale systems (72 Instinct MI455X GPUs + 18 EPYC 'Venice' CPUs) into production, claiming up to 30% more tokens per dollar than the leading competitor; Anthropic announced plans for up to 2 GW of MI455X/Helios, and OpenAI expects its first Helios capacity online in Q4 2026 under its 6 GW AMD deal.

- Helios: 72 MI455X GPUs + 18 6th-gen EPYC 'Venice' CPUs per rack
- MI455X claimed 34x token throughput vs MI355X; Helios 'up to 30% more tokens per dollar' than leading competitor (AMD claims)
- Anthropic: up to 2 GW of MI455X in Helios
- OpenAI: Helios online from Q4 2026; part of 6 GW multi-generation deal starting with 1 GW of MI450-class in H2 2026
- Customers also include Meta, Microsoft, Oracle, HUMAIN; roadmap MI500 (2027), MI600 (2028)

#### What happened
AMD's first rack-scale system answers Nvidia's NVL72 and comes with gigawatt-scale commitments from two of the top three frontier labs.

#### Why it matters
A credible second source of frontier training/inference compute weakens Nvidia's pricing power and diversifies lab supply chains.

Sources: [AMD IR: AAI 2026 — full-stack compute for the agentic AI era](https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era); [TechWire Asia: AMD Advancing AI 2026 highlights](https://techwireasia.com/2026/07/amd-advancing-ai-2026-helios-openai-meta-anthropic/); [Fierce Network: AMD launches full AI stack](https://www.fierce-network.com/cloud/amd-launches-full-stack-ai-compute-agentic-era)

### 2026-08-06: Texas announces SpaceX's $16.8B first phase of Terafab in Grimes County, a 100-million-sq-ft chip campus with 3,000 jobs

***; SpaceX, Tesla, State of Texas; Chips & compute; confidence high; https://postcutoff.com/e/2026-08-06-terafab-grimes-county-16-8b/

On Aug 6, 2026 Governor Greg Abbott announced that SpaceX will build the first phase of Terafab, a vertically integrated chip plant for logic, memory and advanced packaging, in Grimes County, Texas, with over $16.8B of capital investment and 3,000 jobs, backed by a $30M Texas Enterprise Fund grant. County filings had discussed ~$55B initially and up to $119B across all phases.

- First phase: >$16.8B capital investment, 3,000 new jobs, a 100-million-square-foot campus on the site of a former power plant (Gibbons Creek) (Governor's office; Manufacturing Dive)
- Incentives: $30M Texas Enterprise Fund grant; qualified under the JETI program
- Musk: the project is 'the most epic chip-building effort' and will 'produce AI chips at scale for use on Earth and in space'
- Chips for Tesla's Optimus robots and Cybercab and SpaceX's planned space-based data centers
- Earlier county filings (May 2026): ~$55B initial outlay, up to $119B across all phases (TechCrunch via Wikipedia); September state filings list construction from Dec 1, 2026 to end-2028

#### What happened
Terafab's permanent home moved from the Giga Texas campus in Austin to a much larger rural site in Grimes County, with state money behind the
first phase.

#### Why it matters
It turned Terafab from a keynote promise into a sited, state-subsidised project with a construction timeline, the scale that later drew
TSMC's interest.

Sources: [Office of the Texas Governor: Governor Abbott announces SpaceX expansion in Grimes County](https://gov.texas.gov/news/post/governor-abbott-announces-spacex-expansion-in-grimes-county); [Manufacturing Dive: SpaceX to invest $16.8B on first phase of Terafab project](https://www.manufacturingdive.com/news/spacex-to-invest-16b-first-phase-terafab-project-grimes-county-texas/827422/); [KBTX: Gov. Abbott confirms SpaceX expansion in Grimes County (Aug 6, 2026)](https://www.kbtx.com/2026/08/06/gov-abbott-confirms-spacex-expansion-grimes-county/); [Wikipedia: Terafab](https://en.wikipedia.org/wiki/Terafab)

### 2026-08-25: OpenAI publishes first benchmarks of Jalapeño, its first custom inference chip

****; OpenAI; Chips & compute; confidence high; https://postcutoff.com/e/2026-08-25-openai-jalapeno-first-results/

On Aug 25, 2026 OpenAI published the first measured results for Jalapeño, its first in-house AI inference chip. On SemiAnalysis's public InferenceX benchmark, serving GPT-OSS 120B, DeepSeek R1 and Kimi K2.5, OpenAI says Jalapeño did 1.5-1.9x more work per watt at peak throughput and had 1.7-3.6x lower end-to-end latency than the Nvidia GB200/GB300 systems it was compared with. OpenAI says AI helped take the chip from design to tapeout in nine months, and it plans to start deploying Jalapeño in its own data centers by the end of 2026.

- OpenAI's first custom inference chip; first measured results published Aug 25, 2026 (Engineering blog), with a companion essay 'The full stack behind abundant intelligence'
- InferenceX (SemiAnalysis) results, normalized by rated chip power: 1.5-1.9x more peak throughput per watt, 1.7-3.6x lower end-to-end latency, and 2.1-4.1x higher performance on highly interactive workloads vs. the comparison systems (OpenAI)
- GPT-OSS 120B vs GB200: ~1.9x peak mixed tokens/s per kW (85,448 vs 44,960), ~1.7x lower end-to-end latency (1.03 s vs 1.80 s), 1,459 vs 535 tokens/s per user at minimum TBT
- DeepSeek R1 670B (MXFP4) vs GB300: ~1.7x per kW, ~3.6x lower latency; Kimi K2.5 1T (MXFP4) vs GB300: ~1.5x per kW, ~3.4x lower latency
- Rated at 700 W package TDP (GB200 1,200 W, GB300 1,400 W); measured sustained power stayed at or below 550 W on the tested workloads
- AI-assisted design: design to tapeout in nine months; AI also optimized arithmetic circuits. Codex with GPT-Astra brought three open-weight models not in the original plan to high performance within two months; AI-written kernels for selected GPT-OSS attention/MoE blocks ran 1.5-1.8x faster than human-expert kernels
- Deployment in OpenAI's compute infrastructure planned 'by the end of the year'; Gen 2 'deep in development', Gen 3 'taking shape'; OpenAI will keep deploying Nvidia and other accelerators
- Oct 2, 2026 (Tom's Hardware interview with Richard Ho, OpenAI VP and Head of Hardware): Jalapeño is being deployed internally in racks hosted by AMD EPYC 'Turin' CPUs with 1.5 TB of memory per host, not Nvidia Vera or Arm's AGI CPU. Ho called Turin 'pragmatic' and said 'Vera, as a standalone, is a little bit behind on that maturity level'
- More Than Moore interview with Ho (Oct 3): OpenAI used 'raw', not fine-tuned, internal models; one attention implementation went from under 1% to nearly 90% of the memory/compute roofline in ~40 hours, and one optimization saved over 13% die area. First silicon mid-May 2026; scale-up domain of 128 accelerators, full system 2,048; Broadcom is the design partner and does board/rack integration
- Inference-only design: OpenAI describes it purely as an inference chip; press (e.g. Quasa) stresses that it cannot be used to train models
- IEEE Spectrum (Sept 14): project began Oct 2024; first architecture concept to first silicon in under 20 months, first RTL to tape-out in 9 months, first chips back in May 2026; team averaged fewer than 100 people (Richard Ho). LLM-guided optimization cut matrix-unit area ~10% versus optimized human baselines, and lifted DeepSeek's MLA kernel from 0.31% to 88.94% of the theoretical ceiling in ~40 hours. The team started with o3 and later used GPT-6 Astra precursors that wrote Verilog directly
- Tom's Hardware interview with Ho ('This is how AI should be used', in its AI Chip Design week, roundup dated Oct 3): Codex was used across the design; Ho says the nine-month RTL-to-tapeout 'established a new baseline'; Jalapeño is for OpenAI's internal use for now

#### What happened
OpenAI published the first measured performance numbers for **Jalapeño**, its first custom inference chip. It tested the chip on
SemiAnalysis's public **InferenceX** benchmark, which measures the whole process of serving a request, using three public models:
GPT-OSS 120B (against an Nvidia GB200 system), DeepSeek R1 670B and Kimi K2.5 1T (against GB300). OpenAI normalized results by each
accelerator's rated chip power. Jalapeño is rated at 700 W, but OpenAI says it stayed at or below 550 W in practice. On that basis
Jalapeño sat on the Pareto frontier of throughput per watt against latency for all three models. OpenAI says the lead grew further on
its own frontier models in internal tests. Those internal results were not published.

OpenAI says the design keeps model state such as the KV cache local and uses a large network domain, so a whole request stays in one
system. It also says AI shortened the design-to-tapeout cycle to nine months. Codex with GPT-Astra ported three extra open-weight models
in two months, and for some GPT-OSS blocks AI-generated kernels beat human-written ones by 1.5-1.8x. OpenAI says the chip will enable
"ultra-fast-mode inference at efficiencies previously available only in fast mode".

#### Why it matters
This is the first published evidence that OpenAI's own silicon works, and it puts OpenAI next to Google (TPU), Amazon (Trainium/Inferentia)
and Meta (MTIA) as a lab with first-party accelerators. The comparison is OpenAI's own and uses rated rather than measured power for the
Nvidia systems. Critics (e.g. MLQ) note the limits of that comparison and that Jalapeño is an inference-only part. The results support OpenAI's
push for faster serving tiers (Ultrafast, launched at DevDay on Sept 29) and lower cost to serve.

Sources: [OpenAI: Jalapeño's first results show industry-leading speed and efficiency in AI inference](https://openai.com/index/jalapeno-first-results/); [OpenAI: The full stack behind abundant intelligence](https://openai.com/index/the-full-stack-behind-abundant-intelligence/); [OpenAI (Sarah Friar): The Work Now Within Reach (Sept 8, cites Jalapeño results)](https://openai.com/index/the-work-now-within-reach/); [TechCrunch: OpenAI's Jalapeño chip is built for fast inference at scale, benchmarks show](https://techcrunch.com/2026/08/25/openais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show/); [MLQ: Jalapeño shows strong inference gains, but its Nvidia comparison has limits](https://mlq.ai/news/openais-jalapeno-shows-strong-inference-gains-but-its-nvidia-comparison-has-limits/); [Quasa: OpenAI Jalapeño leads InferenceX, but cannot train models](https://quasa.io/insights/openai-s-jalapeno-leads-inferencex-but-it-cannot-train-models); [Tom's Hardware: OpenAI's Jalapeño ASICs are deployed alongside AMD EPYC Turin CPUs as hosts (Oct 2)](https://www.tomshardware.com/pc-components/cpus/openais-jalapeno-asics-are-deployed-alongside-amd-epyc-turin-cpus-as-hosts-hardware-vp-says-nvidias-vera-standalone-is-a-little-bit-behind-on-that-maturity-level); [Tae Kim on X: quoting OpenAI's Jalapeño results](https://x.com/firstadopter/status/2092266927377039838); [More Than Moore: Interview with Richard Ho, OpenAI](https://morethanmoore.substack.com/p/interview-with-richard-ho-openai); [Ian Cutress (More Than Moore) video interview with Richard Ho (YouTube, Sept 30)](https://www.youtube.com/watch?v=8s7uYtCM1bc); [IEEE Spectrum: How OpenAI used its own LLMs to design its Jalapeño chip (Sept 14)](https://spectrum.ieee.org/llms-for-chip-design); [Tom's Hardware: OpenAI head of hardware breaks down the AI-assisted design of Jalapeño](https://www.tomshardware.com/tech-industry/asics/this-is-how-ai-should-be-used-openai-head-of-hardware-breaks-down-the-ai-assisted-design-of-its-jalapeno-asic); [Tom's Hardware: Jalapeño is for OpenAI's internal use, door left open to broader rollout](https://www.tomshardware.com/tech-industry/artificial-intelligence/openais-custom-jalapeno-ai-inference-asic-is-for-openais-internal-use-but-company-leaves-the-door-open-to-broader-rollout-firm-says-it-will-have-its-hands-full-with-jalapeno-for-a-good-long-time)

### 2026-08-26: NVIDIA posts $96.2B quarter; Vera Rubin in full production and deploying at major clouds

****; NVIDIA; Chips & compute; confidence high; https://postcutoff.com/e/2026-08-26-nvidia-q2-fy2027-vera-rubin-production/

NVIDIA's Q2 FY2027 results (2026-08-26) showed revenue of $96.2B (+106% YoY) and data-center revenue of $89.0B, with the Vera Rubin platform in full production and deploying at CoreWeave, Google Cloud, Microsoft Azure, OCI and Nebius; NVIDIA guided the next quarter to $108B.

- Q2 FY2027 revenue: $96.2B, +106% YoY, +18% QoQ
- Data Center revenue: $89.0B, +117% YoY
- Q3 FY2027 outlook: $108.0B +/-2%; gross margin ~74.0%
- Vera Rubin in full production; deploying at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius
- Vera CPU ('first AI agent CPU') rolling out; Spectrum-6 switches arriving at AI factories; Vera BlueField-4 STX announced
- Cosmos 3 launched as an open frontier omnimodel for physical AI
- Jensen Huang: 'AI has reached its inflection point... compute is revenue.'

#### What happened
NVIDIA reported its fiscal Q2 2027 (quarter ending July 2026): revenue $96.2B, more than double a year earlier, and
data-center revenue $89.0B. The company said the **Vera Rubin** platform is in full production and being deployed by
major clouds and neoclouds, alongside the Vera CPU, Spectrum-6 networking and BlueField-4 STX storage.

#### Why it matters
Vera Rubin shipping in volume in H2 2026 is the compute step-change that 2027 frontier models will be trained and served
on; NVIDIA's near-$100B quarter is the clearest financial measure of the AI buildout's scale.

Sources: [NVIDIA Q2 FY2027 press release (SEC 8-K)](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm); [TechPowerUp - Vera Rubin NVL144 servers set for 2026 volume production](https://www.techpowerup.com/342049/nvidia-vera-rubin-nvl144-servers-set-for-2026-volume-production)

### 2026-08-26: Anthropic agrees to pay Nscale $45B over six years for ~460 MW at the Monarch campus in West Virginia

***; Anthropic, Nscale, Nvidia; Chips & compute; confidence high; https://postcutoff.com/e/2026-08-26-anthropic-nscale-45b-monarch-west-virginia/

Bloomberg reported on Aug 26, 2026 that Anthropic agreed to spend $45B over six years renting about 460 MW of AI computing capacity from Nscale at its Monarch Compute Campus in Mason County, West Virginia, with Nvidia Vera Rubin chips coming online from late 2027. Microsoft had signed a letter of intent for the site in March and walked away in the summer. In October, conservationists warned that a grove of about 200 American chestnut trees next to the site could be cleared.

- Deal (Bloomberg via TNW, Aug 26): ~$45B over six years, ~460 MW (roughly what 345,000 US homes draw), first of three buildings; Nvidia Vera Rubin chips from late 2027
- Campus: Monarch Compute Campus, up to 2,250 acres in Mason County, WV, acquired by Nscale in March 2026 via AIP Corp; ~1.35 GW for the first buildings, power runway above 8 GW; full campus plus on-site power plant ~$71B per planning documents, $47B of it chips (TNW)
- Microsoft signed a letter of intent for 1.35 GW in March 2026 and walked away in the summer, leaving Anthropic as tenant (TNW)
- The contract became the largest item in Nscale's ~$103B contracted backlog ahead of its US listing (see 2026-09-25-nscale-convertible-ipo)
- Chestnut grove (Oct 2026): about 200 blight-resistant American chestnuts planted in the 1970s at the former Clements State Tree Nursery in West Columbia, the main seed source of the American Chestnut Foundation's WV chapter, may be cleared as early as Nov 1, 2026, per advocacy sources cited by Futurism and Gadget Review; the project's website says the nursery is not on its land, and neither Nscale nor Anthropic had commented (unresolved)
- TACF's WV chapter newsletter (April 2026) says it had tried for years to learn whether the trees would be affected, with no reply from the governor's or commerce secretary's office
- Second deal in talks (Bloomberg, Oct 19, via Investing.com): Nscale and Anthropic are negotiating another ~$45B agreement for a second Monarch building, again 460 MW of Vera Rubin capacity; it 'could be finalized within days' but terms may change; Anthropic reportedly holds a right of first refusal on the campus's expected third building. Nscale's IPO filing puts existing Anthropic agreements at up to $44.6B. Both companies declined to comment. The talks came a day after Microsoft dropped its Loughton (UK) anchor-tenant deal with Nscale

#### What happened
Anthropic added another multi-year compute lease to a string of 2026 deals (SpaceX, Fluidstack, Lambda, Akamai). Nscale, a two-year-old
London company, took over a West Virginia site first planned for Microsoft and booked Anthropic as its anchor tenant before going public.

#### Why it matters
It is one of the largest single compute contracts of 2026 and a big part of Nscale's backlog. The chestnut grove dispute shows how
local environmental concerns around AI campuses now reach the labs that rent them, even when the lab is only a tenant.

Sources: [Investing.com (Bloomberg): Nscale in talks for $45 billion Anthropic AI deal as IPO looms (Oct 19)](https://www.investing.com/news/stock-market-news/nscale-in-talks-for-45-billion-anthropic-ai-deal-as-ipo-looms-4958181); [Bloomberg: Anthropic to pay Nscale $45 billion for AI computing power (Aug 26, 2026)](https://www.bloomberg.com/news/articles/2026-08-26/anthropic-to-pay-nscale-45-billion-for-ai-computing-power); [TNW: Anthropic signs $45B compute deal with Britain's Nscale](https://thenextweb.com/news/anthropic-nscale-45b-compute-deal-london-ipo); [DatacenterDynamics: Anthropic signs $45bn compute capacity agreement with Nscale (report)](https://www.datacenterdynamics.com/en/news/anthropic-signs-45bn-compute-capacity-agreement-with-nscale-report/); [Futurism: One of America's last remaining chestnut groves could be wiped out for an AI data center (Oct 10)](https://futurism.com/artificial-intelligence/chestnut-grove-last-remaining-anthropic-ai-data-center); [Gadget Review: One of America's last chestnut groves is about to be destroyed for a data center](https://www.gadgetreview.com/?p=417727); [American Chestnut Foundation, WV chapter newsletter (April 2026): Monarch Data Center in Mason County](https://tacf.org/wp-content/uploads/2026/04/WV-Newletter-April-2026.pdf)

### 2026-08-31: Anthropic signs a $35B cloud deal with Nvidia-backed Lambda

***; Anthropic, Lambda, Nvidia; Chips & compute; confidence medium; https://postcutoff.com/e/2026-08-31-anthropic-lambda-35b-cloud-deal/

On Aug 31, 2026 the Wall Street Journal, then Reuters and Bloomberg, reported that Anthropic had signed a cloud-computing deal worth about $35B with Lambda, an Nvidia-backed GPU cloud, to bring Nvidia capacity online for Claude. The contract pushed Lambda's backlog from $15B in June to $50B in September.

- Value: ~$35B (WSJ first; Reuters citing a source familiar, Aug 31, 2026)
- Purpose: bring online Nvidia capacity to meet demand for Claude (Reuters)
- Lambda's backlog: $15B (June 2026) → $50B (September 2026), mostly from the Anthropic contract (TechCrunch, Oct 6)
- Secondary reports (not verified here) link the deal to a ~350 MW Texas site in Nueces County developed by Hut 8 and leased by Nvidia

#### What happened
Anthropic added Lambda to the clouds it rents Nvidia GPUs from, next to Google TPUs, AWS Trainium and other neoclouds. The deal was found
on 2026-10-07 through coverage of Lambda's funding round, five weeks after it was reported, so this entry was written late. The WSJ and
Bloomberg articles are paywalled; the facts come from the Reuters summary and TechCrunch. Details of the site and contract length are not
confirmed here.

#### Why it matters
It is one of Anthropic's largest compute commitments of 2026 and on its own more than tripled Lambda's backlog. That shows how far
frontier labs now pay for the growth of the neocloud companies.

Sources: [Bloomberg: Anthropic Strikes $35 Billion Cloud Deal With Nvidia-Backed Lambda](https://www.bloomberg.com/news/articles/2026-08-31/anthropic-seals-35-billion-cloud-deal-with-nvidia-backed-lambda); [Reuters via Yahoo: Anthropic signs $35 billion cloud deal with Nvidia-backed Lambda, source says](https://www.yahoo.com/news/articles/anthropic-signs-35-billion-cloud-235939268.html); [WSJ: Anthropic signs $35 billion cloud deal backed by Nvidia](https://www.wsj.com/tech/ai/anthropic-signs-35-billion-cloud-deal-backed-by-nvidia-f12622f1); [TechCrunch (Oct 6): AI computing startup Lambda to raise $4B ahead of planned IPO](https://techcrunch.com/2026/10/06/ai-computing-startup-lambda-to-raise-4b-ahead-of-planned-ipo/)

### 2026-09-01: Center for Technology & Statecraft: Chinese fabs stockpiled ~343 ASML immersion DUV scanners, enough to scale Huawei Ascend output; urges a China-wide DUVi ban

***; Center for Technology & Statecraft, ASML, Huawei; Chips & compute; confidence medium; https://postcutoff.com/e/2026-09-01-cts-china-duv-lithography-stockpile/

A September 2026 report by Nicholas Brown and Saif M. Khan of the Center for Technology & Statecraft estimates that Chinese-owned fabs had imported about 343 immersion DUV lithography systems by Q1 2026, roughly 270 of them ASML NXT:1980i tools that can make 7nm-class logic and HBM for Huawei Ascend AI chips, after spending over $13B on NXT:1980i machines in 2024–2025. It recommends replacing partial controls with a ban on all DUVi exports to China plus servicing and spare-part controls; SCMP reported it on Oct 2–3, 2026.

- Stock by Q1 2026 (CTS estimate): ~270 NXT:1980i, ~22 NXT:2050i, ~6 NXT:2100i and ~45 older DUVi systems (~343 in total)
- Chinese fabs spent over $13B on NXT:1980i systems in 2024–2025 (about 90 in 2024 and 89 in 2025, per SCMP)
- China's share of ASML DUVi shipments rose from 26% (2022) to 60–70% (2024Q1–2026Q1), per CTS
- CTS estimates Huawei will make 62,000–160,000 B300-equivalents of AI chips in 2026, versus 6.89M B300-equivalents of US production
- Recommendations: ban all DUVi exports to China, restrict servicing at advanced-node Chinese fabs, prevent diversion of the stockpile to AI chips, control subcomponents; still allow dry DUV for legacy production

#### What happened
The Center for Technology & Statecraft (CTS) published a report arguing that immersion DUV (DUVi) lithography, not EUV, is now the main
bottleneck for China's AI chip production. It counts about 343 DUVi scanners in Chinese-owned fabs by early 2026, most of them ASML's
NXT:1980i, which Chinese fabs use with multi-patterning to make 7nm-class logic and HBM2/2e memory for Huawei's Ascend accelerators.
The South China Morning Post covered the report on Oct 2–3, 2026 (Techmeme, Oct 3), framing it as fuel for US calls for a complete ban.

#### Why it matters
It quantifies how much chipmaking capacity China secured before tighter controls, and gives a concrete policy target (a full DUVi ban
plus servicing limits) in the US export-control debate that shapes how much AI compute China can build.

Caveats: the counts are CTS estimates from trade and shipment data, not ASML disclosures; the exact publication day was not verified
(the report page says September 2026; SCMP says "late September"), so the date is month-precision.

Sources: [Center for Technology & Statecraft: DUV Immersion Lithography, the Foundation of the US AI Advantage (Brown, Khan)](https://techstatecraft.org/duv); [Tom's Hardware: China stockpiled 343 immersion DUV tools; 270 ASML scanners can produce 7nm processors (Oct 5)](https://www.tomshardware.com/tech-industry/semiconductors/china-stockpiled-343-immersion-duv-tools-for-advanced-chipmaking-report-claims-270-asml-scanners-can-produce-7nm-processors-without-sanctioned-euv-tools); [SCMP: China stockpiles ASML lithography tools, spurring US calls for complete export ban](https://www.scmp.com/tech/tech-war/article/3369557/china-stockpiles-asml-lithography-tools-spurring-us-calls-complete-export-ban)

### 2026-09-09: Google commits €13 billion to Finnish AI data centers, its largest single investment in Europe

***; Google; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-09-google-13bn-finland-ai-infrastructure/

On Sept 9, 2026 Google announced at least €13 billion (~$15B) of investment in Finnish digital infrastructure over 2027–2028: data centers in Hamina, Kajaani, Muhos and Vaala serving Gemini, Search, Maps and YouTube, plus a 22-year agreement supporting the life extension of the Loviisa nuclear plant, new onshore wind and a 94 MW battery. It is Google's largest single investment in Europe.

- Amount: at least €13 billion over two years (2027–2028); announced Sept 9, 2026 by Bikash Koley, VP Google Global Infrastructure
- Sites: Hamina (existing campus expansion), Kajaani, Muhos and Vaala
- Energy: 22-year agreement supporting the Loviisa nuclear plant's life extension; new onshore wind capacity; 94 MW battery system for cold, windless periods
- Google's estimates: €3.6B a year added to Finland's GDP and 37,000+ jobs supported during construction; €31M for local communities over four years; AI upskilling for 4,400+ workers
- Oct 2026: Finland ordered Google to halt some data-center work (see 2026-10-06-finland-orders-google-halt-data-center-work)

#### What happened
Google expanded beyond its long-running Hamina site to three new Finnish locations. Finland offers a cool climate, clean power and grid
capacity. The nuclear life-extension agreement is a long-term power deal tied to Fortum's Loviisa plant (Yle reports up to half its output).

#### Why it matters
It is one of the largest AI-infrastructure commitments in Europe in 2026 and part of the Nordic data-center boom (Bloomberg counted over
€67B of planned or active Finnish data-center investment by October). Weeks later, Finnish authorities ordered Google to halt some work.

Sources: [Google blog: Google deepens its commitment to Finland with a €13 billion investment in AI infrastructure](https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/google-ai-commitment-to-finland/); [Google Cloud Press Corner: two-year €13 billion investment in AI infrastructure](https://www.googlecloudpresscorner.com/2026-09-09-Google-Deepens-Commitment-to-Finland-with-Two-Year-EUR13-Billion-investment-in-AI-Infrastructure); [Yle: Google's Finland investment](https://yle.fi/a/74-20245301); [Evertiq: Google to invest €13 billion in Finnish AI infrastructure](https://evertiq.com/news/2026-09-09-google-to-invest-13-billion-in-finnish-ai-infrastructure); [Bloomberg: Finland's data center tally tops €67 billion on AI boom (Oct 8)](https://www.bloomberg.com/news/articles/2026-10-08/finland-s-data-center-tally-tops-67-billion-on-ai-boom)

### 2026-09-15: MediaTek launches Dimensity 9600 Pro, its first 2 nm phone chip, with a dual NPU for on-device models of up to 30B parameters

**; MediaTek; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-15-mediatek-dimensity-9600-pro/

On Sept 15, 2026 MediaTek announced the Dimensity 9600 Pro, its first smartphone chip on a 2 nm process. It has two NPUs: a low-power one for always-on AI and the NPU 1090, which MediaTek says supports on-device models of up to 30B parameters with 51% faster LLM prefill than the previous generation. Qualcomm announced a comparable 2 nm chip a week later.

- Announced Sept 15, 2026 in Hsinchu; 2 nm process (TSMC, per Investing.com); companion chip Dimensity 9600M
- CPU: 2+3+3 all-big-core design, two C2-Ultra cores up to 4.55 GHz; up to 17% higher single-core and 15% higher multi-core performance, 61% lower multi-core power than the previous generation (MediaTek figures)
- NPU 1090: 51% higher LLM prefill performance, 55% more tokens generated per watt, models up to 30B parameters (MediaTek figures, vs Dimensity 9500)
- Second-generation Super Efficient NPU: 40% lower power for always-on AI; an 'Agentic AI Engine' for proactive, personalised assistance
- First phones 'expected to launch this quarter'; press reports name the vivo X500 Pro and OPPO Find X10 Pro Max
- ChannelNews (Oct 12) says the chip supports Google's Gemini Nano; MediaTek's press release does not mention it

#### What happened
MediaTek's 2026 flagship phone chip is built around running large language models on the phone. All performance numbers are MediaTek's
own, compared with its Dimensity 9500.

#### Why it matters
With Qualcomm's Snapdragon 8 Elite Extreme Gen 6, it sets 30B-parameter local models as the 2026 flagship-phone target, the hardware
base for on-device agents.

Sources: [MediaTek: Dimensity 9600 Pro sets new standard for flagship smartphone chips](https://www.mediatek.com/press-room/mediatek-dimensity-9600-pro-sets-new-standard-for-flagship-smartphone-chips); [ChannelNews: MediaTek brings Gemini Nano to smartphones with new Dimensity 9600 Pro chipset](https://www.channelnews.com.au/mediatek-brings-gemini-nano-to-smartphones-with-new-dimensity-9600-pro-chipset/); [Let's Data Science: MediaTek launches Dimensity 9600 Pro on TSMC 2nm](https://letsdatascience.com/news/mediatek-launches-dimensity-9600-pro-on-tsmc-2nm-de42a972)

### 2026-09-17: Crusoe raises $3.9B Series F at a $30.9B valuation; a week later it drops its $1.25B Boom Supersonic turbine order

***; Crusoe, Boom Supersonic; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-17-crusoe-3-9b-series-f/

On Sept 17, 2026 Crusoe, the Denver AI data-center builder behind OpenAI's Abilene, Texas site, raised a $3.9B Series F at a $30.9B post-money valuation, about 3x its October 2025 price. It also showed "Spark" modular AI factories that can be delivered by truck. On Sept 25 Crusoe cancelled its $1.25B launch order for 29 of Boom Supersonic's 42 MW "Superpower" gas turbines. Boom had planned to use that deal to fund its Overture airliner.

- Series F: $3.9B at $30.9B post-money; co-leads Atreides Management, Mubadala Capital, Valor Equity Partners; with Founders Fund, GIC, Nvidia, Qatar Investment Authority, Radical Ventures, TPG (TechCrunch)
- Previous round: $1.38B at $10B (Oct 2025). New board members: Thomas Seifert (Cloudflare CFO), Bill Stein, JB Straubel (TechCrunch)
- Customers named: Meta, Microsoft, Oracle, OpenAI (Abilene); a reported ~$13B five-year deal with Jane Street; exploring an IPO with Goldman Sachs and Morgan Stanley (TechCrunch)
- Sept 25: the Boom deal is off: 29 Superpower turbines (42 MW each, ~1.2 GW in total), $1.25B, first deliveries planned for 2027. Boom CEO Blake Scholl announced it on X (TechCrunch)
- Crusoe spokesperson Andrew Schmitt: its energy plans 'haven't changed', it still plans to use turbines, 'just not Boom's', alongside wind, solar, batteries and the grid (TechCrunch)
- Scholl: Boom will deliver about 250 MW of Superpowers next year to other sites and targets 1 GW in 2028 (TechCrunch)

#### What happened
Crusoe started out mining crypto with stranded gas and now builds and runs AI campuses. It closed a $3.9B Series F led by Atreides,
Mubadala Capital and Valor, at a valuation about three times its 2025 round. Eight days later, Boom Supersonic CEO Blake Scholl said Crusoe
had cancelled its order as launch customer for Boom's aeroderivative "Superpower" turbines. Crusoe said it would keep using turbines from
other suppliers.

#### Why it matters
Power is the binding constraint on the AI build-out. On-site gas turbines became a stopgap for data centers waiting for grid connections,
and companies from other industries (here, a supersonic-jet startup) entered the market to sell them. Losing its launch customer weakens one
of those new suppliers.

Sources: [TechCrunch: Crusoe raises $3.9B to build massive data centers and small modular 'AI factories'](https://techcrunch.com/2026/09/17/crusoe-raises-3-9b-to-build-massive-data-centers-and-small-modular-ai-factories/); [TechCrunch: Crusoe abandons $1.25B plan to use Boom turbines at AI data centers](https://techcrunch.com/2026/09/25/crusoe-abandons-1-25b-plan-to-use-boom-turbines-at-ai-data-centers/); [TechCrunch: Crusoe reportedly raises $3B at a $30B valuation (Sept 3)](https://techcrunch.com/2026/09/03/crusoe-reportedly-raises-3b-at-a-30b-valuation/)

### 2026-09-18: Huawei sets Ascend 950 cluster cloud launch (China Sept 30, global Nov 30) and Ascend 960 roadmap

***; Huawei; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-18-huawei-ascend-950-cluster-cloud/

At Huawei Connect 2026 (2026-09-18) Huawei Cloud said its Ascend 950 AI cluster cloud service launches commercially in China on 2026-09-30 and globally on 2026-11-30 — 1,024-card clusters delivering 1 EFLOPS FP8 / 2 EFLOPS FP4 with 256TB unified memory — and set Ascend 960DT for Q1 2027 and 960PR for Q3 2027.

- Ascend 950 cluster: 1,024 cards; 1 EFLOPS FP8, 2 EFLOPS FP4; 256TB globally addressable memory; UnifiedBus interconnect
- Commercial launch: China 2026-09-30; global 2026-11-30
- Over 1,000 Ascend supernodes already deployed
- China launch (2026-09-30): reported as launched by an SMM flash (Oct 1); the first Ascend 950 liquid-cooled SuperPoD was installed at the Qingpu Yunhu computing centre in Shanghai on 2026-09-29 (Sohu, Oct 1). No Huawei Cloud press release, price or named customer found by 2026-10-09
- Roadmap: Ascend 960DT Q1 2027; Ascend 960PR Q3 2027
- Atlas 950 SuperPoD scales to 8,192 chips; Huawei claims 6.7x the compute of Nvidia's Vera Rubin NVL144 (vendor claim)

#### What happened
Huawei Cloud CEO Zhou Yuefeng announced dates at Huawei Connect 2026. DeepSeek V4 was validated on Ascend at launch, and DeepSeek said V4-Pro prices could fall as Ascend 950 scales.

**Outcome of the China launch (checked 2026-10-09).** Nothing says the Sept 30 launch slipped. An SMM news flash on Oct 1 said the cluster
service was "officially launched for commercial use on September 30", and Sohu reported on Oct 1 that the world's first Ascend 950
liquid-cooled SuperPoD (1,024 chips) was installed at the Qingpu Yunhu intelligent computing centre in Shanghai on Sept 29. Searches in
English and Chinese after Golden Week found no Huawei Cloud press release, console listing, price or named customer, so the launch is treated
as happened on the stated date but without first-hand confirmation.

#### Why it matters
Ascend 950 is China's main answer to US export controls; selling it as a global cloud service extends Huawei's AI compute beyond China.

Sources: [TechNode: Huawei sets commercial launch dates for Ascend 950 AI cluster cloud](https://technode.com/2026/09/18/huawei-sets-commercial-launch-dates-for-ascend-950-ai-cluster-cloud-service/); [Huawei Central: Ascend 950 AI cluster to debut globally on November 30](https://www.huaweicentral.com/huawei-ascend-950-ai-cluster-to-debut-globally-on-november-30/); [SMM flash (Oct 1): Ascend 950 AI computing cluster officially launched for commercial use on September 30](https://news.metal.com/newscontent/104143081-smm-flash-huawei-ascend-950-ai-computing-cluster-officially-launched-for-commercial-use-on-september-30); [Sohu (Oct 1): first Ascend 950 liquid-cooled SuperPoD lands in Shanghai Qingpu (Chinese)](https://www.sohu.com/a/1082965333_121332532); [DCD: Huawei announces annual Ascend cadence and supernode](https://www.datacenterdynamics.com/en/news/huawei-announces-annual-release-cadence-for-three-new-ascend-ai-chips-unveils-supernode-offering-company-says-will-outperform-nvidias-nvl144/)

### 2026-09-21: Meta announces Petal, the first 1-petabit-per-second transatlantic subsea cable (US–France, 2029)

**; Meta; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-21-meta-petal-petabit-subsea-cable/

On Sept 21, 2026 Meta announced Petal, a roughly 7,000 km subsea cable between the US and France's Atlantic coast designed for 1 petabit per second (1,000 Tbps), about twice today's top transoceanic systems. It will be the first to use multi-core fiber at scale, with NEC as system supplier, Sumitomo Electric for fiber and Orange for the French landing. It is expected in service in 2029.

- ~7,000 km (4,300 mi), US to France; design capacity 1 Pbps (1,000 Tbps)
- First subsea system to deploy multi-core fiber at scale
- NEC (system), Sumitomo Electric (fiber), Orange (French landing)
- Expected in service 2029

#### What happened
Meta added a petabit-class Atlantic link to its private subsea network, which already includes 2Africa.

#### Why it matters
AI training and inference across distant data centers needs far more intercontinental bandwidth. Petal is an example of AI
infrastructure spending reaching undersea networking.

Sources: [Meta: Announcing Petal, a first-of-its-kind transoceanic subsea cable](https://about.fb.com/news/2026/09/announcing-petal-meta-petabit-transoceanic-cable/); [DCD: Meta plans Petal, the first transoceanic subsea cable at petabit capacity](https://www.datacenterdynamics.com/en/news/meta-plans-petal-the-first-transoceanic-subsea-cable-at-petabit-capacity/)

### 2026-09-22: Qualcomm unveils Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6 (2 nm) for on-device AI agents; Extreme runs a 30B MoE locally

**; Qualcomm; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-22-snapdragon-8-elite-gen-6/

At Snapdragon Summit on Sept 22, 2026 Qualcomm announced two 2 nm flagship phone chips built around on-device agentic AI. The Snapdragon 8 Elite Extreme Gen 6 can run a 30B-parameter mixture-of-experts model locally. Both chips have sensing hubs that run always-on models of up to 200M parameters for agent personalization.

- Two platforms: Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6, both on a 2 nm process (Qualcomm via 9to5Google, Unite.AI)
- 8 Elite Extreme Gen 6: runs a 30B-parameter MoE model on-device (Unite.AI)
- Sensing hubs run small AI models up to 200M parameters for always-on personalization
- 8 Elite Gen 6 vs Gen 5: +10% CPU, +37% power efficiency, +14% NPU performance, +20% AI performance per watt (9to5Google)
- OEMs: HONOR, iQOO, Motorola, OnePlus, OPPO, REDMI, RedMagic, vivo, Xiaomi

#### What happened
Qualcomm's annual flagship refresh, in a standard and an "Extreme" tier, both pitched around local AI agents rather than chatbot features.

#### Why it matters
A 30B MoE running locally on upcoming Android flagships moves usable agent models onto phones, with no cloud round-trip or data leaving the
device.

Sources: [9to5Google: Qualcomm announces the Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6](https://9to5google.com/2026/09/22/snapdragon-8-elite-gen-6/); [Unite.AI: Qualcomm unveils two flagship Snapdragon chips built for agentic AI](https://www.unite.ai/qualcomm-unveils-two-flagship-snapdragon-chips-built-for-agentic-ai/); [The AI Insider: Snapdragon 8 Elite Gen 6 chips built for on-device AI agents](https://theaiinsider.tech/2026/09/23/qualcomm-unveils-snapdragon-8-elite-gen-6-chips-built-for-on-device-ai-agents/)

### 2026-09-23: Brookings paper projects $10.3 trillion US AI buildout for 2025–2032 and warns financing is moving off balance sheet

***; Brookings Institution, Columbia Business School; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-23-brookings-financing-the-ai-buildout/

In a Brookings Papers on Economic Activity paper (Sept 23, 2026; presented at the Sept 24–25 conference), Columbia's Stijn Van Nieuwerburgh projected US AI investment in data centers, power, networking and chips at $10.3 trillion over 2025–2032, about 3.63% of GDP a year. Relative to the economy, that is larger than the canal, railroad, electrification, highway and telecom booms. He warned that risk is migrating into opaque off-balance-sheet vehicles.

- Projection: $10.3T of US AI investment 2025–2032 (data-center buildings, power systems, networking, specialized chips and equipment), averaging 3.63% of GDP per year
- Larger relative to GDP than the major US canal, railroad, electrification, highway and telecommunications investment booms
- Financing shift: from transparent on-balance-sheet corporate funding to joint ventures, private credit, securitization, special-purpose vehicles, lease commitments and loan guarantees
- Risks named: uncertain AI demand, rapid technological change, timely access to power and hardware, and the credit quality of a limited number of data-center operators
- Assessment: not yet comparable to earlier credit booms, but off-balance-sheet structures could hide correlated exposures; main policy ask is better measurement and transparency
- Reported widely, including by the Wall Street Journal (Sept 24)

#### What happened
The fall 2026 Brookings Papers on Economic Activity included "Financing the AI Buildout" by Stijn Van Nieuwerburgh (Columbia Business
School). The paper sizes the US AI infrastructure boom against past infrastructure booms and traces how it is being financed.

#### Why it matters
It gives a mainstream-economics benchmark for the scale of the AI capex wave. It also warns specifically that the boom's risks are becoming
harder to see as financing moves into joint ventures, private credit and special-purpose vehicles.

Sources: [Brookings: Financing the AI buildout (Sept 23, 2026)](https://www.brookings.edu/articles/financing-the-ai-buildout/); [BPEA conference draft: Financing the AI Buildout (PDF)](https://www.brookings.edu/wp-content/uploads/2026/09/4c_Van-Nieuwerburgh.pdf); [Brookings: BPEA Fall 2026 Conference](https://www.brookings.edu/events/bpea-fall-2026-conference/); [IndexBox: Brookings study, AI buildout to exceed railroads and highways as share of GDP](https://www.indexbox.io/blog/brookings-study-ai-infrastructure-buildout-to-consume-36-of-gdp-topping-railroads-and-highways/)

### 2026-09-24: Google sets the first Project Suncatcher launch: a 4-TPU orbital prototype on SpaceX Transporter-18 (Oct 1)

***; Google, Planet, SpaceX; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-24-google-suncatcher-first-launch-set/

On Sept 24, 2026 Google said the first Project Suncatcher prototype, a refrigerator-sized satellite with 4 Trillium TPUs built with Planet, will launch on SpaceX's Transporter-18 from Vandenberg on Oct 1, 2026. Two satellites to test inter-satellite laser links follow in 2027. Google says orbit gives up to 8x more solar energy than the ground.

- Prototype: 4 Trillium TPUs, built with Planet; launch Oct 1, 2026 on SpaceX Transporter-18 from Vandenberg
- Follow-up: two satellites to test laser links in 2027
- Google: up to 8x more solar power in orbit than on the ground
- Launched on schedule Oct 1, 2026 (11:32 a.m. PT); see 2026-10-01-google-suncatcher-prototype-launched
- Launch listed for Oct 1, 2026 at 18:18 UTC (launch trackers); outcome tracked in upcoming item 2026-10-01-suncatcher-transporter-18-launch

#### What happened
Google's plan for solar-powered AI compute in space moves from paper to hardware.

#### Why it matters
It is the first test of TPUs as an orbital compute platform, as ground datacenters hit power limits (see the Oracle New Mexico notice).

Sources: [Google: Project Suncatcher facts](https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/); [Quartz: Google Project Suncatcher TPU satellite](https://qz.com/google-project-suncatcher-tpu-satellite-spacex-orbit-092426); [Google video: latest moonshot to put machine learning in space](https://www.youtube.com/watch?v=o1JK79jszqo); [WAI Hub: Transporter 18 launch page](https://www.whataboutit.space/launches/transporter-18)

### 2026-09-24: Oracle sends a force-majeure notice on the 2.45 GW New Mexico Stargate campus after gas-pipeline delays

***; Oracle, OpenAI, Blue Owl; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-24-oracle-force-majeure-stargate-new-mexico/

On Sept 24, 2026 Oracle sent a force-majeure notice for the 2.45 GW Stargate campus in New Mexico ('Project Jupiter', developed by Blue Owl, targeted for 2028). The notice would let Oracle delay payments if the site misses 2028. The cause is energy: a gas pipeline about six months late, a denied pipeline permit and a pending fuel-cell air permit. Oracle says it is not exiting and remains on schedule.

- Site: 2.45 GW campus in New Mexico, developed by Blue Owl, targeted for 2028
- Energy Transfer pipeline ~6 months late (now 2027-02-01); a pipeline permit denied; fuel-cell air permit pending (state deadline Nov 23)
- Oracle: not exiting, 'remains on our planned schedule'
- Originally reported by Bloomberg
- Oracle on X (Sept 24): 'Project Jupiter remains on our planned schedule', citing a 'reimagined' power plan and 3,600+ construction workers; it did not deny the notice
- WSJ (Sept 25): the project faces power and permitting hurdles that expose Oracle's AI build-out risks

#### What happened
A flagship Stargate site hit power-supply problems, and Oracle protected itself contractually.

#### Why it matters
Power and permitting, not chips, are now the binding constraint on the largest AI datacenter projects.

Sources: [TechCrunch: Oracle sends force majeure notice on its New Mexico Stargate data center](https://techcrunch.com/2026/09/24/oracle-sends-force-majeure-notice-on-its-new-mexico-stargate-data-center/); [Bloomberg: Oracle cites force majeure to shield itself on controversial data center](https://www.bloomberg.com/news/articles/2026-09-24/oracle-cites-force-majeure-to-shield-itself-on-controversial-data-center); [WSJ: Cracks in Oracle's AI data center build-out appear in massive New Mexico project](https://www.wsj.com/finance/investing/cracks-in-oracles-ai-data-center-build-out-appear-in-massive-new-mexico-project-effb51c2); [Oracle on X: Project Jupiter remains on schedule](https://x.com/Oracle/status/2103105677874897334)

### 2026-09-25: SemiAnalysis: China has 24+ GW of data-center capacity (US: 56 GW by end-2026), ~50 GW more in the pipeline; ByteDance uses a fifth

***; SemiAnalysis; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-25-semianalysis-china-datacenter-24gw/

SemiAnalysis's China Datacenter Model (Sept 25, 2026; tracking 1,000+ facilities from 60+ players) puts China's delivered data-center capacity at over 24 GW, more than EMEA and the rest of Asia combined, against 56 GW in the US by end-2026. About 20 GW of dated pipeline and 30 GW of announced projects follow. ByteDance occupies about one-fifth of Chinese capacity. The FT reported the figures on Oct 8.

- China delivered capacity 24+ GW; pipeline ~20 GW dated plus ~30 GW announced (≈50 GW under construction or planned); US 56 GW at 2026 year-end
- ByteDance leases about 1/5 of China's delivered capacity; ByteDance/Alibaba/Tencent/Baidu capex reached ~$20B in Q2 2026, more than double a year earlier
- Inner Mongolia is the leading AI hub (power about half the price of Tier-1 cities, renewables, generous quotas) under the 'Eastern Data, Western Compute' policy
- Build speed: a standard 100 MW facility in ~12 months; Alibaba's modular '100-day datacenter' from shell to servers; Chinese permitting 3–6 months vs 12–13 months in the US

#### What happened
SemiAnalysis launched a facility-level model of Chinese data centers. The FT's Oct 8 story brought the numbers to a wider audience.

#### Why it matters
Compute capacity is the core of the US–China AI race. China has less than half of US capacity today but builds and permits faster, with
cheap inland power.

Sources: [SemiAnalysis: The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model](https://newsletter.semianalysis.com/p/the-chinese-ai-infrastructure-boom); [Financial Times (Oct 8): China's data center capacity (SemiAnalysis figures)](https://www.ft.com/content/e1dd8bff-b06d-4a40-bbb7-c0a6a36f1c8e); [SCMP: ByteDance occupies one-fifth of China's data centre capacity: SemiAnalysis](https://www.scmp.com/tech/big-tech/article/3369056/bytedance-grabs-one-fifth-chinas-data-centre-capacity-ai-drives-infrastructure-boom)

### 2026-09-27: The Information: China asks ByteDance, Alibaba and others for Nvidia RTX Pro 5500 purchase plans and signals approval

***; MIIT, Nvidia, ByteDance, Alibaba; Chips & compute; confidence medium; https://postcutoff.com/e/2026-09-27-china-nvidia-rtx-pro-5500-purchase-plans/

On Sept 27, 2026 The Information reported that China's Ministry of Industry and Information Technology asked ByteDance, Alibaba and other companies to submit plans to buy Nvidia's new RTX Pro 5500, a Blackwell workstation GPU with 84GB of memory, and told some of them that approval was expected. Beijing had until then discouraged purchases of Nvidia chips to support domestic chipmakers. Reuters could not verify the report.

- Source: The Information (Sept 27, 2026), relayed by Reuters, which said it could not independently verify it
- Ministry: MIIT asked ByteDance, Alibaba and others to submit RTX Pro 5500 purchase plans; some were told approval is anticipated
- RTX Pro 5500: Blackwell-based workstation card with 84GB memory, unveiled by Nvidia in September 2026; some executives expect it to fall outside US export restrictions
- Context: the US cleared H200 sales to Chinese firms earlier in 2026, but ByteDance and Tencent were told to keep first shipments in Hong Kong rather than the mainland (TNW)
- No public comment from MIIT, Nvidia, ByteDance or Alibaba at the time of reporting
- Orders (Reuters exclusive, Oct 20, 2026, two sources; read via The Edge Malaysia): ByteDance, Alibaba and Tencent ordered more than 3 million RTX PRO 5500 cards, about $11,400 each, combined sticker value above $35B; orders from all Chinese customers exceed 6 million. Nvidia told customers it cannot guarantee fulfilment, may supply partners only 5-10% of requested volume, can make no more than 3 million of the cards worldwide in 2027 (TSMC capacity shifting to Rubin); early samples in December, commercial deliveries from Q1 2027. Buyers mainly want them for inference; the card performs about like the discontinued RTX 6000D and is weaker than H200/B300. Tansun Technology disclosed an order of 20,000 cards on Oct 13. MIIT did not comment
- Follow-up (Financial Times, Oct 13, 2026; paywalled, read through the Techmeme summary): sources say Chinese tech companies bought more than 1M Nvidia gaming and workstation chips, mostly RTX 5090 and RTX Pro, for basic AI workloads amid an AI chip shortage

#### What happened
Rather than blocking Nvidia purchases outright, Beijing appears to be managing them: companies must report their plans and the ministry decides.
The chip is a workstation part, not a data-center accelerator, which is why it may sit outside US export limits.

#### Why it matters
Chinese labs remain compute-constrained. A Beijing-approved channel for a large-memory Blackwell card would ease inference capacity while keeping
state control over how much Nvidia hardware enters China. Single-source report; confidence medium.

Sources: [TNW: China may let Alibaba, ByteDance buy Nvidia's RTX Pro 5500](https://thenextweb.com/news/china-alibaba-bytedance-nvidia-rtx-pro-5500); [Reuters via Yahoo: China may allow ByteDance, Alibaba to buy Nvidia RTX PRO 5500 chips](https://www.yahoo.com/news/world/articles/china-may-allow-bytedance-alibaba-131252441.html); [Seeking Alpha: China may allow Alibaba, ByteDance to buy Nvidia's new RTX Pro 5500 chips](https://seekingalpha.com/news/4647403-china-may-allow-alibaba-bytedance-to-buy-nvidia-s-new-rtx-pro-5500-chips); [Benzinga: China reportedly green lights RTX PRO 5500](https://www.benzinga.com/markets/tech/26/09/62013860/nvidia-chips-for-alibaba-bytedance-china-reportedly-green-lights-rtx-pro-5500-amid-us-curbs); [Financial Times: Chinese tech companies bought 1M+ Nvidia gaming and workstation chips (Oct 13, 2026)](https://www.ft.com/content/9dbd99db-6b0c-43bd-b7be-849f129df098); [The Edge Malaysia (Reuters): ByteDance, Alibaba, Tencent place over US$35b worth of Nvidia chip orders (Oct 20)](https://theedgemalaysia.com/node/822533); [Reuters: ByteDance, Alibaba, Tencent place over $35 billion worth of Nvidia chip orders (Oct 20)](https://www.reuters.com/business/retail-consumer/bytedance-alibaba-tencent-place-over-35-billion-worth-nvidia-chip-orders-2026-10-20/)

### 2026-09-29: DeepSeek partners with Huawei, open-sourcing TileLang-based programming tools and libraries for Ascend chips as a CUDA alternative

***; DeepSeek, Huawei; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-29-deepseek-huawei-tilelang-ascend/

On Sept 29, 2026 DeepSeek said on its official WeChat account that it is working with Huawei on programming tools for Huawei's Ascend AI chips. It is open-sourcing Ascend infrastructure built around TileLang, a high-level open-source language for AI kernels, plus compute and communication libraries. The aim is an "independent, self-controlled" software stack that lowers Chinese developers' reliance on Nvidia's CUDA.

- Announced on DeepSeek's official WeChat account on Sept 29, 2026 (Reuters)
- Open-sourced: TileLang support for Ascend, plus Ascend-optimized compute and communication libraries (operators, matrix computation, inter-card communication, attention, data filtering, per secondary reports)
- DeepSeek says TileLang offers 'a simpler programming model' than CUDA while reaching the hardware's full performance
- DeepSeek: 'To build a new generation of independent, self-controlled GPU software ecosystems, the first priority is establishing a high-level language that is universal, easy to program' and still reaches full performance
- The two companies also worked on a supernode of 128 Ascend 950 processors, optimizing how computation and chip-to-chip communication are distributed (Reuters via Techzine)
- Background: Huawei's own CANN stack had been criticized as hard and unstable to use; TileLang is an existing open-source tile-based kernel language that DeepSeek had used for GPU kernels before
- Released repos: deepseek-ai/DeepGEMM-Ascend (created Sept 29; ~583 stars by Oct 16), a port of DeepGEMM that is fully API-compatible and supports BF16, FP8 and FP4 GEMM, MQA logits and MegaMoE on Ascend NPUs; and deepseek-ai/DeepEP-Ascend (created Sept 30; ~257 stars), described as 'a high-performance communication library for machine learning training and inference on Huawei Ascend NPUs' (READMEs and repo descriptions)
- The Information (Oct 19, Qianer Liu and Juro Osawa; paywalled, Techmeme summary): DeepSeek and Huawei are collaborating on future AI chip designs, and DeepSeek plans to start using the Ascend 950DT for pre-training as soon as the end of 2026. Earlier reports summarized in the press (Bloomberg) had said DeepSeek meant to use the 950DT for running models, not training, so the training role is still contested

#### What happened
DeepSeek announced the partnership and the open-source release on WeChat. It covers a high-level programming language (TileLang) with an
Ascend backend and libraries for compute and inter-chip communication. These mirror the toolkit DeepSeek had built for Nvidia GPUs.

#### Why it matters
Nvidia's lock on AI chips rests largely on CUDA. China's leading open-model lab is now publishing a developer stack for Huawei's chips, which
tackles the software barrier as well as the hardware one. Its V4 models already run on Ascend.

Unverified: the exact names of the released repositories were not in the reports read for this run.

Sources: [Reuters via Investing.com: DeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidia](https://www.investing.com/news/stock-market-news/deepseek-partners-with-huawei-to-develop-chip-programming-tools-reducing-reliance-on-nvidia-4924051); [US News (Reuters): DeepSeek partners with Huawei to develop chip programming tools](https://money.usnews.com/investing/news/articles/2026-09-29/deepseek-partners-with-huawei-to-develop-chip-programming-tools-reducing-reliance-on-nvidia); [SDxCentral: DeepSeek teams up with Huawei to beef up Ascend chip programming tools](https://www.sdxcentral.com/news/deepseek-teams-up-with-huawei-to-beef-up-ascend-chip-programming-tools/); [Techzine: DeepSeek brings AI software to Huawei's Ascend chips](https://www.techzine.eu/news/devops/144657/deepseek-brings-ai-software-to-huaweis-ascend-chips/); [GitHub: deepseek-ai/DeepGEMM-Ascend](https://github.com/deepseek-ai/DeepGEMM-Ascend); [GitHub: deepseek-ai/DeepEP-Ascend](https://github.com/deepseek-ai/DeepEP-Ascend); [TileLang on GitHub](https://github.com/tile-ai/tilelang); [Reuters: DeepSeek partners with Huawei to develop chip programming tools](https://www.reuters.com/world/asia-pacific/deepseek-partners-with-huawei-develop-chip-programming-tools-reducing-reliance-2026-09-30/); [The Information: How DeepSeek is shaping Huawei's AI chips in a growing challenge to Nvidia (paywalled)](https://www.theinformation.com/articles/deepseek-shaping-huaweis-ai-chips-growing-challenge-nvidia)

### 2026-09-30: Micron posts record $54.2B quarter on AI memory demand and guides to $61.5B

***; Micron; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-30-micron-record-fq4-2026/

On Sept 30, 2026 Micron reported fiscal Q4 2026 revenue of $54.23B, up from $11.32B a year earlier, with GAAP net income of $37.70B. Fiscal 2026 revenue was $133.19B against $37.38B the year before. It guided fiscal Q1 2027 revenue to $61.5B ± $1.5B, above analyst estimates. The AI build-out has turned memory into one of the most profitable parts of the chip industry.

- Mehrotra (Oct 1, The Register): 'In calendar 2027 as well as 2028, we see demand exceeding supply'; 'We do not have line of sight to when supply and demand will return to balance'; customers will pay 'much higher prices' than in 2026; HBM demand is growing faster than server DRAM demand
- FQ4 2026 revenue $54.23B (vs $41.46B prior quarter, $11.32B a year earlier); GAAP net income $37.70B ($32.87/share); non-GAAP $33.42/share
- FY2026 revenue $133.19B (vs $37.38B in FY2025); GAAP net income $84.97B
- FQ1 2027 guidance: revenue $61.5B ± $1.5B; analysts had expected about $57B (CNBC)
- Micron's CEO used the White House's new term: 'AI is becoming Super Intelligence (SI), and memory enhances this intelligence...'
- Investor slides (Investing.com): AI demand keeps supply tight; most of calendar 2027 HBM supply reportedly already contracted at higher prices
- Micron also said humanoid robots could need 200GB+ of DRAM and terabytes of NAND each (Benzinga)

#### What happened
Micron's annual revenue more than tripled as AI data centers absorbed HBM, high-capacity server DRAM and SSDs. Quarterly net
income was about 70% of revenue.

#### Why it matters
Memory has become a bottleneck and a profit center of the AI build-out. Micron's guidance is a leading indicator that AI
infrastructure spending was still accelerating into late 2026. The CEO adopting "Super Intelligence (SI)" days after the executive
order shows how fast the new term spread.

Unverified: the HBM contracting detail comes from a secondary summary of the slides, not the press release.

Sources: [The Register: RAM supply set to worsen, says Micron](https://www.theregister.com/systems/2026/10/01/ram-supply-set-to-worsen-says-micron-as-ceo-celebrates-much-higher-prices/5300346); [Micron FQ4 2026 press release (SEC 8-K Ex. 99.1)](https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm); [CNBC: Micron (MU) Q4 earnings report 2026](https://www.cnbc.com/2026/09/30/micron-mu-q4-earnings-report-2026.html); [Investing.com: Micron FQ4 2026 slides: record revenue, AI demand drives supply tightness](https://www.investing.com/news/company-news/micron-fq4-2026-slides-record-revenue-ai-demand-drives-supply-tightness-93CH-4926074); [Benzinga: Micron says humanoid robots could drive future memory demand](https://www.benzinga.com/markets/earnings/26/09/62096820/micron-humanoid-robots-memory-driver-server-level-chips)

### 2026-09-30: Tencent leases ~100,000 advanced AI chips from Oracle in Southeast Asia in a ~$7B five-year deal (FT)

***; Tencent, Oracle; Chips & compute; confidence medium; https://postcutoff.com/e/2026-09-30-tencent-oracle-7b-overseas-chip-lease/

The Financial Times reported on Sept 30, 2026 that Tencent signed a roughly $7B, five-year contract with Oracle earlier in 2026 for access to about 100,000 advanced AI chips hosted in Oracle data centers in Southeast Asia. The chips never enter China, which US export rules allow; it is reported as Tencent's largest overseas compute lease and shows remote leasing outpacing chip export controls.

- Value ~$7B over 5 years; ~100,000 advanced AI chips (chip generation not disclosed); roughly 30% paid upfront (FT via Startup Fortune, AI Weekly)
- Capacity sits in Oracle data centers in Southeast Asia; Chinese firms may not buy top Nvidia chips outright but may lease compute overseas under US rules
- Signed 'this year'; the upfront payment reportedly weighed on Tencent's Q2 free cash flow
- Reuters said it could not immediately verify the figures; neither company disclosed chip models or host countries
- Oracle shares rose about 2.5% in overnight trading after the report (Stocktwits)
- Context: ByteDance is already one of Oracle's largest GPU customers in Asia-Pacific (Startup Fortune)
- Financing (Bloomberg, Oct 12, sources; headline and Techmeme summary only): Tencent is planning a ~$5B bond offering, including $1.5B to $2B in dollar tranches, and told investors it is mainly for AI capital expenditure, with the bulk going toward chip purchases. Reuters the same day cited a source for a $4.5B sale (headline only)

#### What happened
On Sept 30, 2026 the Financial Times reported that Tencent had signed a five-year deal worth about $7B with Oracle to lease capacity
of roughly 100,000 advanced AI chips in Oracle's Southeast Asian data centers. About 30% was reportedly paid upfront. Tencent cannot
buy the most advanced Nvidia accelerators for use in China, but US rules do not bar Chinese companies from renting compute abroad.
Reuters, following the FT, said it could not independently verify the figures.

#### Why it matters
It is one of the largest known overseas compute leases by a Chinese AI company and a concrete example of the "remote access" gap in
US chip export controls, which govern physical shipments rather than cloud rental.

Sources: [Bloomberg: Tencent Tells Investors Its $5 Billion Bond Mainly for AI Capex (Oct 12; paywalled, not read)](https://www.bloomberg.com/news/articles/2026-10-12/tencent-tells-investors-its-5-billion-bond-mainly-for-ai-capex); [Financial Times: Tencent signs $7bn deal with Oracle for AI chips (paywalled)](https://www.ft.com/content/8799b33d-f07c-4a03-82f0-bf5d3d1d29e9); [Startup Fortune: Tencent leases 100,000 AI chips from Oracle in a $7 billion five year deal](https://startupfortune.com/tencent-leases-100000-ai-chips-from-oracle-in-a-7-billion-five-year-deal/); [AI Weekly: Tencent leases 100,000 AI chips from Oracle in $7B deal](https://aiweekly.co/alerts/tencent-leases-100000-ai-chips-from-oracle-in-7b-deal); [Stocktwits: Why is ORCL stock jumping in overnight trading?](https://stocktwits.com/news-articles/markets/equity/why-is-orcl-stock-jumping-over-2-5-in-overnight-trading/cZDirDpRBLH); [Rohan Paul on X summarizing the FT report](https://x.com/rohanpaul_ai/status/2105504789526409498)

### 2026-09-30: Faulty redaction exposes Google's Nebraska data-center power and water figures and expected tax refunds, which Google had claimed as trade secrets

**; Google; Chips & compute; confidence high; https://postcutoff.com/e/2026-09-30-google-lincoln-data-center-redaction-water-power/

On Sept 30, 2026 Lincoln, Nebraska station 10/11 News (KOLN) reported that the redactions in Google's annual report to Nebraska's Department of Water, Energy and Environment could be undone by copying and pasting the blacked-out text. The report showed that Google's Lincoln data center (Agate LLC) drew 52.65 MW at peak and used 13.299 million gallons of water, figures Google had withheld as trade secrets, and that three Google data-center entities expected about $117M in Nebraska tax refunds for 2025. The story reached 504 points on Hacker News on Oct 4.

- Lincoln site (Agate LLC): 52.65 MW at peak electrical demand; 13.299 megagallons of water across cooling towers, evaporative systems and site operations (about 20 Olympic pools), per the annual report to the state
- Google had marked these figures as trade secrets under Nebraska statutes (Neb. Rev. Stat. §§ 81-1527, 84-712.05; NAC Title 115 ch. 2)
- Expected refunds for 2025 taxes: Agate LLC (Lincoln) $55,822,472; Fireball LLC (Papillion) $39,171,573.39; Westwood Solutions LLC (Omaha) $22,558,881
- How it surfaced: reporter Madison Pitsch highlighted the redaction boxes, copied the text and pasted it into another document
- Reach: Hacker News front page, 504 points (Oct 4, 2026); no Google or city response reported in the article

#### What happened
A local TV newsroom obtained Google's annual compliance report for its Lincoln, Nebraska data center (operated as Agate LLC). The
power and water figures had been blacked out as trade secrets, but the redaction was only cosmetic: copying the hidden text revealed
peak demand of 52.65 MW, 13.299 million gallons of water, and the tax refunds Google's three Nebraska data-center companies expected
under the state's incentive programme.

#### Why it matters
AI data centers' water and power use is a growing local political issue, and operators often keep site-level figures confidential.
This leak gives rare site-level numbers for a Google facility, and its HN reach shows strong public interest in data-center transparency.

Sources: [10/11 News (KOLN): More questions than answers about Lincoln's Google data center water, electricity usage (Sept 30, 2026)](https://www.1011now.com/2026/09/30/more-questions-than-answers-about-lincolns-google-data-center-water-electricity-usage/); [Hacker News discussion (504 points)](https://news.ycombinator.com/item?id=49957068)

### 2026-10-01: Anthropic prospectus: Broadcom to lend up to $42B in convertible notes toward a $125.2B five-year TPU lease

***; Anthropic, Broadcom, Google; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-01-broadcom-42b-convertible-loan-anthropic/

Reuters reported on Oct 1, 2026, from Anthropic's IPO prospectus, that Broadcom agreed to lend Anthropic up to $42B through convertible notes. The money helps finance a five-year, $125.2B lease of TPU computing capacity starting in 2027. Broadcom thus becomes Anthropic's chip supplier, its lessor and its lender at once. The filing flags "potential conflicts of interest".

- Up to $42B from Broadcom in convertible notes, which can later convert to equity; about one third of the $125.2B five-year TPU lease commitment
- Capacity starts arriving in 2027; Anthropic is expected to become Broadcom's largest compute customer that year
- Anthropic does not expect any notes to be sold before its IPO completes; payment or performance defaults could accelerate lease obligations and restrict access to the facility (Finimize summary of the filing)
- Filing language (per reports): 'potential conflicts of interest'; Broadcom's 'decisions around pricing and hardware could affect' Anthropic's ability to procure enough computing infrastructure
- Press framed it as the latest 'circular' AI financing deal (supplier funds customer who buys its product), alongside similar arrangements with Amazon, Microsoft, Nvidia and AMD
- Oct 2 (Bloomberg): Broadcom's bank syndicate began gathering $60B of AI-chip financing meant to benefit Anthropic and others. Bank of America, Citigroup and Morgan Stanley are to send syndication letters for a $42B Class A senior-secured tranche, and Blackstone leads an $18B Class B junior tranche, committing $9B from its funds and syndicating the rest
- Earlier prospectus reporting (Reuters, Sept 28–29): Broadcom lease arrangements are non-cancelable by either party except on default; total infrastructure obligations ~$518B
- MarketBeat (Oct 12), citing the prospectus seen by Reuters: Anthropic has about $161B of commitments tied to Broadcom; of the $60B debt package, the $42B senior secured loans are partly guaranteed by Broadcom and the $18B junior loans are not; Anthropic is 'reportedly targeting a public debut as soon as mid-November, potentially before Thanksgiving', later than earlier expectations of October. The $161B figure and the mid-November timing were not checked against Reuters directly
- Oct 16 (Financial Times, paywalled; via Reuters headline and Investing.com summary): in the ~$60B package, called Project North Star and tied to about 1.5 GW of Google TPUs, Anthropic offered Blackstone a 10% coupon on the $9B it is expected to take of the $18B junior tranche backed by the five-year lease payments; total returns exceed 10% after issuance discounts and fees. Similarly rated borrowers pay about 7% in public bond markets; the earlier Project Big Sky raised about $4.4B at 8.5% four months before. The other $9B of junior debt is to be marketed in the coming months, possibly at lower yields; the debt could become tradable in the 144A market once drawn in 2027. Blackstone declined to comment
- X reaction to the FT yield report (fxtwitter, Oct 17): JustDario: 'Why Anthropic claims to be worth 2 trillion $ and then behave like a company that is on its way to be worth close to zero?' ~201k views; Ed Zitron, quoting it: 'Blackstone losing every dollar on this deal' … 'billions in debt at *10%* - what an insane deal!' ~148k

#### What happened
A further detail from the leaked Anthropic S-1 prospectus: Broadcom, which designs Google's TPUs, will also lease TPU capacity to
Anthropic and lend it up to $42B toward the cost. The loan is convertible, so Broadcom can end up with Anthropic shares.

#### Why it matters
It shows how frontier-lab compute is financed in 2026. Chip and cloud suppliers fund the labs that buy from them. That inflates
commitments (Anthropic's ~$518B) and ties the suppliers' fortunes to the labs' IPOs.

Sources: [MarketBeat: Broadcom's Anthropic Bet Gets Bigger as $60 Billion Financing Raises the Stakes (Oct 12, 2026)](https://www.marketbeat.com/articles/broadcoms-anthropic-bet-gets-bigger-as-60-billion-financing-raises-the-stakes/); [Investing.com: Anthropic offers Blackstone over 10% yield in $60B AI chip financing (FT report, Oct 16)](https://www.investing.com/news/company-news/anthropic-offers-blackstone-over-10-yield-in-60b-ai-chip-financing--ft-reports-4955122); [Financial Times: Anthropic offers Blackstone double-digit yield in latest chip financing (paywalled, not read)](https://www.ft.com/content/cb4034e1-475f-40b9-a5f9-256e91aacb89); [Reuters: Broadcom to lend Anthropic up to $42 billion to lease its chips, filing says](https://www.reuters.com/business/broadcom-lend-anthropic-up-42-billion-lease-its-chips-filing-says-2026-10-01/); [Yahoo Finance: Broadcom to lend Anthropic up to $42 billion, in latest circular investing deal](https://finance.yahoo.com/technology/article/broadcom-to-lend-anthropic-up-to-42-billion-in-deal-to-lease-chips-121617505.html); [Finimize: Anthropic's IPO filing puts Broadcom at the center](https://finimize.com/content/anthropics-ipo-filing-puts-broadcom-at-the-center); [Bloomberg: Broadcom starts amassing $60 billion to fund chips for Anthropic](https://www.bloomberg.com/news/articles/2026-10-02/broadcom-starts-amassing-60-billion-to-fund-chips-for-anthropic); [Dataconomy: Broadcom banks seek $60B to finance AI chips](https://dataconomy.com/2026/10/02/broadcom-anthropic-60-billion-usd-ai-chip-debt/); [Ed Zitron on X: on the 10% Blackstone yield (~148k views)](https://x.com/edzitron/status/2111262199872389375)

### 2026-10-01: Google's first Project Suncatcher satellite (4 TPUs running Gemma) launches on SpaceX Transporter-18

***; Google, Planet, SpaceX; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-01-google-suncatcher-prototype-launched/

On Oct 1, 2026 a SpaceX Falcon 9 (Transporter-18, 130 payloads) lifted off from Vandenberg at 11:32 a.m. PT carrying the first Project Suncatcher prototype, a refrigerator-sized Planet-built satellite with 4 Google TPUs that will run a version of Gemma in sun-synchronous orbit. It is Google's first in-orbit test of AI compute; per NPR, Gemma runs 15 minutes at a time because of heat limits.

- Launch: Oct 1, 2026, 11:32 a.m. PT, Falcon 9 Transporter-18 from Vandenberg; 130 payloads; booster's 25th flight, landed at Vandenberg
- Payload deployment scheduled to start 54.5 minutes after launch
- Satellite: built with Planet, 4 TPUs, sun-synchronous orbit; runs a version of Gemma for 15 minutes at a time (heat management)
- Planned mission length: about one year; two more satellites to test laser links in 2027
- Suncatcher lead Travis Beals (NPR): 'I don't see this being cheaper in the next five years.'
- No public confirmation yet (as of Oct 2) of first contact or in-orbit TPU results

#### What happened
Transporter-18 launched on schedule on Oct 1, 2026 and carried Google's Project Suncatcher pathfinder, built with Planet, to low Earth orbit.
The satellite has four TPUs and will run a version of Google's Gemma model in sessions of about 15 minutes, a limit set by heat dissipation.
Its sun-synchronous orbit keeps the solar panels almost always lit. The mission is planned to run for about a year. Two satellites to test
inter-satellite laser links follow in 2027.

#### Why it matters
This is the first time Google has put its own AI accelerators in orbit to test radiation tolerance and thermal behaviour, the first hardware step
toward solar-powered orbital data centers. The project lead says it is unlikely to be cheaper than ground compute within five years.

Sources: [NPR (via KLCC): Google launches Project Suncatcher, a step towards AI data centers in space](https://www.klcc.org/npr-news/2026-10-01/google-launches-project-suncatcher-a-step-towards-ai-data-centers-in-space); [CNBC: SpaceX launches Google AI chips into orbit](https://www.cnbc.com/2026/10/01/spacex-to-launch-google-ai-chips-to-orbit-with-planet-labs-satellites.html); [Starlust: Transporter-18 sends Google's AI test satellite to orbit with 129 other payloads](https://starlust.org/watch-space-xs-transporter-18-sends-googles-ai-test-satellite-to-orbit-with-129-other-payloads/); [Google: Project Suncatcher facts](https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/)

### 2026-10-02: SpaceXAI turns into an AI cloud provider: talks with Microsoft on leasing capacity, on top of Anthropic and Google contracts

***; SpaceXAI, SpaceX, Microsoft, Anthropic, Google; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-02-spacexai-ai-cloud-microsoft-talks/

The Information reported on Oct 2, 2026 that SpaceX's AI unit (SpaceXAI, ex-xAI) held talks over the summer with Microsoft about leasing it computing capacity, and that it has "lined up billions of dollars a month in commitments from compute-hungry AI labs". Earlier reporting (Fortune, July) put Anthropic at $1.25B a month for ~325,000 Nvidia GPUs and Google at $920M a month for ~110,000 GPUs, both cancellable on 90 days' notice. Musk had initially resisted renting out compute.

- The Information (Oct 2): talks with Microsoft over the summer about leasing capacity; no deal reported
- Fortune (July 19): Anthropic contract ~$1.25B/month for ~325,000 Nvidia GPUs; Google ~$920M/month for ~110,000 GPUs; both with 90-day cancellation; together ~ $26B a year
- Fortune: SpaceX's AI segment had 2025 revenue of $3.2B and an operating loss of ~$6.4B
- Anthropic's Colossus 1 deal was first announced at Code with Claude (May 2026)

#### What happened
SpaceXAI, which still trains Grok, now earns much of its AI revenue by renting Colossus capacity to rival labs and hyperscalers.

#### Why it matters
A frontier lab acting as a neocloud for its competitors (Anthropic, Google, possibly Microsoft) shows how scarce compute has become, and
blurs the line between AI labs and infrastructure providers.

Sources: [The Information: How SpaceX's AI unit turned itself into an AI cloud firm](https://www.theinformation.com/articles/spacexs-ai-unit-turned-ai-cloud-firm); [The Information on X](https://x.com/theinformation/status/2105693008985284875); [Fortune: SpaceX is quietly building an AI compute business](https://fortune.com/2026/07/19/spacex-ai-compute-renting-business-google-anthropic-pentagon-deals-revenue-valuation-elon-musk-colossus-data-centers/)

### 2026-10-02: Amazon explores moving ~$8B of installed Nvidia Grace Blackwell chips into a sale-leaseback vehicle; raises GPU Capacity Block prices 15%

**; Amazon, Nvidia; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-02-amazon-nvidia-chip-sale-leaseback/

The Financial Times reported (Oct 2, 2026) that Amazon has approached investors about transferring about $8B of Nvidia Grace Blackwell systems, already running in more than a dozen US data centers in five states, into a special-purpose vehicle funded by debt, and leasing them back. Separately AWS raised prices of EC2 Capacity Blocks for ML (reserved Nvidia GPU time) by 15%. Nothing on the leaseback is agreed.

- ~$8B of Grace Blackwell hardware across 12+ US sites in five states (incl. Nevada and Virginia), per FT as reported by Reuters/Yahoo and Digitimes
- Structure: SPV funded by debt sold to outside investors, with an equity slice of up to 10% (reports differ on who holds it); investors expect an investment-grade rating tied to Amazon's AA credit
- Purpose: a more asset-light balance sheet; Amazon's 2026 capex budget is ~$200B (up from $131B in 2025)
- AWS raised EC2 Capacity Blocks for ML prices by 15%; e.g. p5.48xlarge (8× H100) at $41.528/hour in major US regions (Yahoo Finance)

#### What happened
Amazon is exploring off-balance-sheet financing for GPUs it already operates, while charging more for scarce reserved GPU capacity.

#### Why it matters
Hyperscalers are now using structured finance (SPVs, leasebacks, syndicated chip loans) to fund AI capex. Price rises for GPU capacity
signal that compute is still scarce in late 2026.

Sources: [Financial Times: Amazon seeks to offload $8bn of Nvidia chips to investors](https://www.ft.com/content/97d8d346-519e-48fb-8df8-66cf5f12ef62); [Yahoo Finance: Amazon hiking AI chip rental prices, exploring Nvidia leaseback](https://finance.yahoo.com/technology/ai/articles/amazon-hiking-ai-chip-rental-122155988.html); [Digitimes: Amazon weighs US$8B chip sale-leaseback to support credit rating](https://www.digitimes.com/news/a20261002PD235/amazon-chips-vehicle-nvidia-technology.html)

### 2026-10-02: Nvidia announces a 64GB DGX Spark for $4,999 as memory prices push the 128GB model to $6,950

**; Nvidia; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-02-nvidia-dgx-spark-64gb/

On Oct 2, 2026 Nvidia announced a DGX Spark with 64GB of unified memory, half the original, from $4,999 starting Oct 23. That is $1,000 more than the 128GB model cost at launch. Reports say the 128GB version now costs $6,950. The new model is a sign of how the memory shortage is raising prices for local-AI hardware.

- DGX Spark 64GB: GB10 Grace Blackwell superchip, DGX OS and the Nvidia AI stack; from $4,999; available from Oct 23 via Acer, ASUS, Dell, Gigabyte, HP and MSI (Nvidia blog)
- Nvidia: one unit runs models up to 100B parameters; two units clustered with NVIDIA Sync Cluster Assistant pool 128GB for models up to 200B; in Nvidia's Qwen 3.8 27B test two units gave up to 1.7x the performance of one
- Original 128GB DGX Spark launched at $3,999; reported current price $6,950 (KuCoin/Futu reports; PCMag: the 64GB model costs $1,000 more than the 128GB model did at launch)
- Reaction: analyst @firstadopter (~11k views) said it costs 'roughly the same street price as the previous 128GB version' with half the memory

#### What happened
Nvidia added a cheaper, lower-memory DGX Spark model, sold through its PC partners, for running local agents and fine-tuning on the
desk. Because memory prices have risen, the cut-down model still costs more than the full model did at launch.

#### Why it matters
The DRAM shortage caused by the AI build-out now shows up directly in the price of developer hardware for running models locally.

Sources: [NVIDIA blog: DGX Spark 64GB gives developers more ways to build and scale local AI](https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/); [PCMag: Facing memory crunch, Nvidia creates 64GB DGX Spark; it still costs $4,999](https://www.pcmag.com/news/facing-memory-crunch-nvidia-creates-64gb-rtx-spark-it-still-costs-4999); [tbreak: Nvidia's 64GB DGX Spark starts at $4,999 on October 23](https://tbreak.com/nvidia-dgx-spark-64gb-price/); [Futu News: 64GB DGX Spark at $4,999; 128GB version rises to $6,950](https://news.futunn.com/en/post/1000536008/nvidia-has-launched-a-64gb-version-of-the-dgx-spark)

### 2026-10-02: Oracle agrees to buy 10–20% of Wisconsin's Point Beach nuclear output for Project Lighthouse, its data-center campus with OpenAI

**; Oracle, We Energies, OpenAI; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-02-oracle-point-beach-nuclear-project-lighthouse/

On Oct 2, 2026 We Energies announced an agreement letting Oracle "subscribe" to 10–20% of the electricity from the Point Beach Nuclear Plant (reported as about 125–250 MW) for Project Lighthouse, the Port Washington, Wisconsin data-center campus Oracle is co-developing with OpenAI. The utility says it would save other customers about $300 million in fuel costs; the deal needs approval from the Wisconsin Public Service Commission.

- Announced Oct 2, 2026 by We Energies; Oracle subscribes to 10–20% of Point Beach output (Civic Media); The Daily Reporter reports 125–250 MW
- Expected customer savings: about $300 million in fuel costs (We Energies), between 2027 and 2033 per press reports
- Point Beach (Manitowoc County) is Wisconsin's only operating nuclear plant, owned by NextEra since 2007; We Energies buys most of its output under an agreement expiring in 2033 and has committed to ~86% of output through 2053
- We Energies: 'Oracle will fully fund Project Lighthouse's energy costs to protect ratepayers' electricity bills and support grid reliability'
- Project Lighthouse: Port Washington (Ozaukee County) campus co-developed with OpenAI; We Energies cites $11B+ economic impact, 4,000+ construction jobs over three years and 1,000 operations jobs; press reports a $15B campus
- Requires Wisconsin PSC approval; the same day We Energies filed for a new gas plant at the Fox Energy Center

#### What happened
Instead of a new reactor, Oracle takes a share of an existing plant's increasingly expensive output, which We Energies is contractually
obliged to buy from NextEra. The utility frames this as a ratepayer benefit: the data center absorbs costs other customers would
otherwise pay.

#### Why it matters
Data-center power costs became a political issue in 2026. This deal is a template for labs and cloud providers paying for grid power
in ways that can be presented as lowering household bills, and it ties an OpenAI-linked campus to nuclear power.

Sources: [We Energies: agreement with Oracle expected to save customers approximately $300 million](https://news.we-energies.com/we-energies-agreement-with-oracle-expected-to-save-customers-approximately-300-million/); [Civic Media: We Energies strikes deal to power Oracle data center with nuclear energy](https://civicmedia.us/news/2026/10/02/we-energies-strikes-deal-to-power-oracle-data-center-with-nuclear-energy); [The Daily Reporter: Oracle plans to buy nuclear power for its data center in Wisconsin](https://dailyreporter.com/2026/10/02/oracle-offers-to-buy-nuclear-power-in-wisconsin/); [BizTimes: User of Port Washington data center commits to absorb $300 million in rising costs for Point Beach](https://biztimes.com/user-of-port-washington-data-center-commits-to-absorb-300-million-in-rising-costs-for-point-beach-nuclear-plant/)

### 2026-10-03: Musk confirms TSMC talks over Terafab, the Tesla–SpaceX AI chip fab in Texas, then says TSMC may at most sublease part of it

***; TSMC, SpaceX, Tesla, xAI, Intel; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-03-musk-confirms-tsmc-terafab-talks/

On Oct 3, 2026 Elon Musk confirmed on X that his Terafab chipmaking project was in talks with TSMC ("Just discussions, but something may come of it"). Tim Culpan's Culpium had reported that TSMC was exploring ways to work with Terafab, possibly by owning and operating the fab. On Oct 7 Musk ruled that out: "we will build and run the fab", and "Maybe TSMC subleases part of the Terafab if they want, but nothing more than that." The same day Intel CEO Lip-Bu Tan told Bloomberg that Intel will stay involved. Terafab is the Tesla–SpaceX project for a giant Texas fab making AI chips only for Musk's companies.

- Musk's reply on X (Oct 3, 2026): 'Just discussions, but something may come of it' (~690k views by Oct 4, fxtwitter)
- Oct 7, 2026 (06:33 UTC), Musk replying on X to a video post by Herbert Ong, which said the likeliest deal had TSMC owning and running the fab: 'No, we will build and run the fab. Let there be ZERO doubt about that. Maybe TSMC subleases part of the Terafab if they want, but nothing more than that.' (~340k views on Oct 7, fxtwitter). TSMC did not immediately respond to Bloomberg
- Intel CEO Lip-Bu Tan told Bloomberg on Oct 7 that Intel will stay involved in Terafab; he spoke ahead of an event in Tokyo marking Intel's 50 years in Japan. Intel has been Terafab's development partner since April (Investing.com and Digitimes, citing Bloomberg)
- Bloomberg's Oct 7 story also recalls its report from the previous week that TSMC is weighing an expansive new chipmaking campus in Texas
- Culpium (Tim Culpan, Oct 2) reported that TSMC was exploring a collaboration; Culpan wrote that 'the most likely scenario is for TSMC to try to own and operate the new factory'. Culpium also reports that the Intel collaboration on Terafab is 'not making progress' and that most Terafab hires so far are ex-Intel staff
- Reuters (Sept 30, citing two people familiar) reported that TSMC was weighing an additional investment in Texas, separate from its ~$265B Arizona build-out; no decision was made and TSMC declined to comment
- Background: Terafab was announced on March 21, 2026 in Austin; Intel joined in April with its 14A (1.4 nm) process chosen for the full-scale fab (Wikipedia, citing Electrek and Reuters)
- Site: the former Gibbons Creek power station in Grimes County, Texas; Tesla and SpaceX committed ~$16.8B to the first phase (Aug 2026); state filings list construction from Dec 1, 2026 to end-2028 (Austin American-Statesman)
- Stated long-term goal: 1 million wafer starts per month and 100–200 billion custom AI and memory chips a year, supplying only Tesla, SpaceX and xAI (Wikipedia; Tom's Hardware)
- No agreement, investment or timeline had been announced by TSMC as of Oct 7
- Culpium follow-up (Tim Culpan, Oct 11, 'Inside TSMC, Texas, and the Terafab Tango', paywalled; from the Techmeme headline and the visible opening only): TSMC executives in charge of overseas operations met Texas officials and began more serious discussions in the first half of 2026, after talking on and off for more than 10 years; the piece also covers the pace at which Intel is proceeding 'while TSMC stays in the hunt'

#### What happened
Tim Culpan's newsletter Culpium reported on Oct 2 that TSMC was exploring ways to work with Terafab, two days after Reuters reported
that TSMC was weighing a Texas investment beyond Arizona. Asked about the report on X, Musk answered: "Just discussions, but something
may come of it." Culpan's follow-up says the Intel partnership announced in April is not making progress.

On Oct 7 Musk narrowed this down. Herbert Ong had posted a video saying the likeliest deal had TSMC owning and running the fab. Musk
replied: "No, we will build and run the fab. Let there be ZERO doubt about that. Maybe TSMC subleases part of the Terafab if they want,
but nothing more than that." Bloomberg reported this as Musk ruling out a TSMC role in Terafab's operations; TSMC did not immediately
comment. The same day Intel CEO Lip-Bu Tan told Bloomberg that Intel, Terafab's development partner since April, will stay involved.

#### Why it matters
Terafab is Musk's attempt to make his own AI chips at a huge scale for Tesla (Optimus, Cybercab), SpaceX (planned space data centers)
and xAI. Bringing in TSMC, the leading foundry, to run the fab would have made the plan far more credible than building leading-edge
capacity from scratch. Musk's Oct 7 reply rules out the scenario Culpium called most likely. Finimize notes that Tesla and SpaceX
would then carry the execution risk of running the fab themselves, which makes Terafab harder to count on as near-term US chip supply.
TSMC may still expand in Texas on its own (Bloomberg). Nothing is agreed yet.

Sources: [Culpium: Inside TSMC, Texas, and the Terafab Tango (Oct 11, paywalled)](https://www.culpium.com/p/inside-tsmc-texas-and-the-terafab); [Elon Musk on X: 'Just discussions, but something may come of it'](https://x.com/elonmusk/status/2106290920605790717); [Elon Musk on X (Oct 7): 'No, we will build and run the fab … Maybe TSMC subleases part of the Terafab if they want, but nothing more than that'](https://x.com/elonmusk/status/2107721071571128709); [Herbert Ong on X (Oct 6): the video post Musk replied to](https://x.com/herbertong/status/2107482007596892239); [Bloomberg: Elon Musk Rules Out Potential TSMC Role in Terafab Operations (Oct 7)](https://www.bloomberg.com/news/articles/2026-10-07/elon-musk-rules-out-potential-tsmc-role-in-terafab-operations); [Yahoo Finance UK (Bloomberg syndication, Debby Wu): Elon Musk Rules Out Potential TSMC Role in Terafab Operations](https://uk.finance.yahoo.com/news/elon-musk-rules-potential-tsmc-084001178.html); [Digitimes: Musk limits TSMC to possible Terafab sublease; Intel says it will stay involved (Oct 7)](https://www.digitimes.com/news/a20261007VL217/intel-spacex-tesla-tsmc-elon-musk.html); [Investing.com: Intel to continue working on Musk's Terafab despite TSMC talks – Bloomberg (Oct 7)](https://uk.investing.com/news/stock-market-news/intel-to-continue-working-on-musks-terafab-despite-tsmc-talks--bloomberg-4898289); [Finimize: Musk Says Tesla And SpaceX Will Run Terafab (Oct 7)](https://finimize.com/content/musk-says-tesla-and-spacex-will-run-terafab); [Culpium: Elon Musk confirms TSMC-Terafab talks](https://www.culpium.com/p/elon-musk-confirms-tsmc-terafab-talks); [Culpium: Exclusive — TSMC exploring plans to work with Terafab](https://www.culpium.com/p/exclusive-tsmc-exploring-plans-to); [Tom's Hardware: Elon Musk confirms discussions with TSMC about Terafab](https://www.tomshardware.com/tech-industry/semiconductors/elon-musk-confirms-discussions-with-tsmc-about-terafab-chipmaking-collaboration-intel-is-the-only-other-named-partner-terafab-to-exclusively-supply-tesla-spacex-and-xai); [Reuters (via WTVB): TSMC evaluates potential Texas investment, sources say](https://wtvbam.com/2026/09/30/tsmc-evaluates-potential-texas-investment-sources-say/); [Tesla North: Elon Musk confirms TSMC talks for $16.8B Texas Terafab project](https://teslanorth.com/2026/10/03/elon-musk-confirms-tsmc-terafab-talks/); [Austin American-Statesman: Elon Musk, SpaceX confirm $16.8B Terafab facility in Texas](https://www.statesman.com/business/article/elon-musk-terafab-spacex-grimes-county-22375537.php); [Wikipedia: Terafab](https://en.wikipedia.org/wiki/Terafab)

### 2026-10-05: Etched fields funding offers at a $40B–$50B valuation, about double its $21B round in August

***; Etched; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-05-etched-funding-talks-40b-50b/

On Oct 5, 2026 TechCrunch reported, citing sources, that AI inference-hardware startup Etched is in early talks to raise new funding after receiving offers that value it at $40B to $50B. That would roughly double the $21B valuation of its $700M round led by Jane Street in August, which itself doubled a $10.3B valuation from July. Investors are betting on inference hardware as an alternative to Nvidia.

- Oct 5 (TechCrunch, sources): Etched has received offers from top-tier investors at a $40B–$50B valuation; talks are early and no deal is reported
- Aug 18, 2026: $700M at a $21B valuation, led by Jane Street after it tested, bought and installed Etched's first shipped AI cluster; Kleiner Perkins, Sequoia, a16z, Peter Thiel, Tiger Global, Bain Capital Ventures, Blackstone and others also joined (TechCrunch)
- Earlier valuations: $5B in December 2025; $300M Series C at $10.3B in July 2026 (TechCrunch)
- Etched sells 'frontier inference clusters': it designed a low-voltage prefill chip and a cluster-scale shared-memory system to speed inference; more than $1B in customer orders booked (TechCrunch)
- Company: founded by Gavin Uberti, Chris Zhu and Robert Wachen; about 400 employees, roughly 15% from Nvidia; a 10 MW data center in Silicon Valley (TechCrunch, Oct 5)

#### What happened
TechCrunch reported on Oct 5 that Etched, which builds inference-only AI hardware, has offers at a $40B–$50B valuation and is in early
funding talks. The company has risen fast: $5B in December 2025, $10.3B in a July 2026 Series C, and $21B in an August round of $700M led
by Jane Street. Jane Street led that round after testing and buying Etched's first shipped cluster, so it is both investor and customer.
Etched says it has more than $1B in orders.

#### Why it matters
It is one of the steepest valuation climbs among Nvidia challengers. Investors are treating inference, not training, as the next big
compute market as agentic workloads grow.

Unverified: the Oct 5 figures come from unnamed sources; no term sheet, lead investor or amount has been reported. This dataset had no
earlier Etched entry, so the August round is recorded here as background.

Sources: [TechCrunch (Oct 5): Etched fields funding offers at $40B valuation, sources say](https://techcrunch.com/2026/10/05/etched-fields-funding-offers-at-40b-valuation-sources-say/); [TechCrunch (Aug 18): Etched's valuation doubles to $21B in a month](https://techcrunch.com/2026/08/18/etcheds-valuation-doubles-to-21b-in-a-month/); [Quartz: Etched raises $700 million at $21 billion valuation](https://qz.com/etched-ai-chip-startup-700-million-21-billion-valuation-081926)

### 2026-10-05: Anthropic hires Google's head of data center global design and construction, John M. Wilson, to lead EMEA capacity delivery

**; Anthropic, Google; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-05-anthropic-hires-google-dc-design-head-john-wilson/

Data Center Dynamics reported on Oct 5, 2026 that John M. Wilson, Google's head of data center global design and construction, has joined Anthropic as a member of technical staff leading EMEA capacity delivery after six years at Google. He is the latest of a series of Google data-center veterans hired as Anthropic builds out its own infrastructure team.

- Role at Anthropic: member of technical staff, leading EMEA (Europe, Middle East, Africa) capacity delivery (DCD)
- Previously ~6 years at Google as head of data center global design and construction; before that European head of design and construction at Amazon (DCD, per search summaries)
- Earlier Google-to-Anthropic data-center hires reported by DCD and others: Sana Ouji (energy lead, April 2026), Mike Brinker (capacity delivery, May 2026), Winnie Leung (head of data center infrastructure), Brett Rogers, Liwen Mao, Adam Johnson, Peter Sarossy, Zach Miller, Soheil Farshchian

#### What happened
The DCD article could not be read directly (bot-blocked). The facts above come from search-result extracts of it, so wording and
details beyond the name and role are unverified.

#### Why it matters
Anthropic has roughly $518B of compute commitments and is moving from renting capacity to designing and delivering its own sites. Hiring
the person who ran Google's global data-center design and construction, to lead its EMEA region, suggests Anthropic is planning
large self-directed builds in Europe and the Middle East.

Sources: [DCD: Google's head of data center global design and construction joins Anthropic](https://www.datacenterdynamics.com/en/news/googles-head-of-data-center-global-design-and-construction-joins-anthropic/); [DCD: Mike Brinker becomes latest Googler to join Anthropic's data center team](https://www.datacenterdynamics.com/en/news/mike-brinker-becomes-latest-googler-to-join-anthropics-data-center-team/); [Construction Digital: Why has Anthropic hired Google leader in data centre push? (Ouji, ex-Google hires list)](https://constructiondigital.com/news/anthropic-hired-google-leader-data-centre-push)

### 2026-10-05: Singapore's DayOne Data Centers files publicly for a US IPO of up to $5B; H1 revenue tripled to $512M

**; DayOne Data Centers; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-05-dayone-data-centers-us-ipo-filing/

On Oct 5, 2026 Singapore-based DayOne Data Centers made its US IPO filing public (Reuters). It reported first-half 2026 revenue more than tripling year on year to $512M, while its net loss widened to $77.2M. The WSJ reported on Oct 6 that the SoftBank-backed company seeks to raise up to $5B by listing American depositary shares before the end of 2026.

- H1 2026 revenue $512M (more than 3x year on year); net loss widened to $77.2M (Reuters, via Techmeme)
- Seeks up to $5B; ADS listing planned by end of 2026 (WSJ). Earlier Bloomberg reports: confidential filing, ~$20B valuation target, listing in November
- Develops and operates data centers for cloud and AI customers in Asia
- SEC filing: DayOne Data Centers Ltd, Form DRS (confidential draft, later made public)
- Oct 13, 2026 (Reuters, read through the Techmeme summary): DayOne raised about $3.67B from seven banks, including an offshore term loan facility of about $2.4B, to fund its Malaysia expansion

#### What happened
The Reuters and WSJ articles are paywalled; this entry uses their Techmeme summaries and the EDGAR filing listing. DayOne joins a line of
AI-infrastructure companies (Nscale, Lambda) heading to public markets in late 2026.

#### Why it matters
Asian data-center builders are tapping US public markets to fund AI capacity. The fast-growing, loss-making numbers show how capital-heavy
the build-out is.

Sources: [Reuters: DayOne Data Centers files for US IPO](https://www.reuters.com/legal/transactional/dayone-data-centers-files-us-ipo-2026-10-05/); [WSJ: SoftBank-backed DayOne Data Centers plans to raise up to $5.0 bln in U.S. IPO](https://www.wsj.com/tech/softbank-backed-dayone-data-centers-plans-to-raise-up-to-5-0-bln-in-u-s-ipo-c568ead1); [SEC EDGAR: DayOne Data Centers Ltd draft registration statement](https://www.sec.gov/Archives/edgar/data/0002118192/000119312526341952/filename1.htm); [Bloomberg Government: DayOne is said to confidentially file for $5 billion US IPO](https://news.bgov.com/capital-markets/dayone-is-said-to-confidentially-file-for-5-billion-ipo-in-us); [Reuters: Singapore data centre developer DayOne secures $3.7 billion financing for Malaysia (Oct 13, 2026)](https://www.reuters.com/world/asia-pacific/singapore-data-centre-developer-dayone-secures-37-billion-financing-malaysia-2026-10-13/)

### 2026-10-05: Ghost launches Core, a $3,499 local 'personal AI computer' for always-on agents, with an $11M a16z-led seed

**; Ghost; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-05-ghost-core-personal-ai-computer/

On Oct 5, 2026 Ghost, founded by 19-year-old Zain Javaid, emerged from stealth with Core, a $3,499 box with an Nvidia RTX Pro 4000 SFF Blackwell GPU that runs open-weight models (Qwen 3.8, Gemma 4) and agents locally on encrypted personal data, plus an $11M seed led by Andreessen Horowitz. The launch post drew about 2M views on X.

- Price $3,499; pre-orders from Oct 5, shipping planned for the last week of October 2026
- GPU: Nvidia RTX Pro 4000 SFF Blackwell (24 GB per a buyer's post); pre-installed Qwen-3.8-Next, Qwen-3.8-27B and Gemma-4-31B; other Hugging Face or custom models can be loaded
- Pitch: a 'brain in a box' that centralises the user's data, keeps it encrypted locally with no cloud dependency, and runs agents continuously
- Funding: $11M seed led by a16z, with Abstract, Audacious Ventures, SV Angel and Nova
- Team: Zain Javaid (CEO, 19), Nicholas Chua, Yifei Chen, Gautam Sharda
- Launch post by Javaid: ~2.0M views and ~8.6k likes by Oct 6

#### What happened
Ghost is selling dedicated local hardware for personal agents, a bet that people will want an always-on assistant over private data without
sending it to a cloud lab. It ships with Chinese and Google open-weight models.

#### Why it matters
A small launch, but a widely seen one, and a marker of the local-agent hardware niche growing alongside cloud agents such as
OpenAI's Dots and Claude Cowork's move to the cloud. Earlier runs skipped it as minor; it was added for its reach.

Sources: [TechCrunch: At 19, Ghost founder raises $11 million to build a $3,499 computer for your personal AI](https://techcrunch.com/2026/10/05/at-19-ghost-founder-raises-11-million-to-build-a-3499-computer-for-your-personal-ai/); [Zain Javaid on X: Introducing Core (~2M views)](https://x.com/zainmfj/status/2107146556176798081)

### 2026-10-06: Finland orders Google's Tuike Finland to halt land clearing at its Muhos and Kajaani data-center sites until environmental assessments are done

***; Google, Tuike Finland, Permit and Supervision Agency (Finland); Chips & compute; confidence high; https://postcutoff.com/e/2026-10-06-finland-orders-google-halt-data-center-work/

On Oct 6, 2026 Finland's Permit and Supervision Agency (Lupa- ja valvontavirasto, LVV) told Google's subsidiary Tuike Finland Oy to stop all work that significantly alters the environment at its planned data-center sites in Muhos and Kajaani. This covers tree removal, topsoil stripping, excavation, quarrying, ditching, and building site roads and storage areas. The work must stop immediately, and no later than Oct 23, until the projects' environmental impact assessments (EIAs) are complete. About 330 ha at Muhos and just under 200 ha at Kajaani had already been cleared or prepared. Google said it had fallen short of its own standards and would follow the agency's guidance. The order came a month after Google announced a €13bn (about $15bn) Finnish data-center plan, which it called its largest single investment in Europe.

- Instrument: a call to suspend ('kehotus') with a hearing, under section 32(2) of Finland's EIA Act (YVA-laki), one per site: LVV-U/37128/2026 (Muhos) and LVV-U/22159/2026 (Kajaani-Otanmäki) (LVV notices)
- Covered: tree removal, topsoil stripping, excavation, quarrying, crushing, moving earth masses, ditching, and building site roads and storage areas. Not covered: planning, surveying, soil investigations and other minor, easily reversible work (LVV)
- Already done, per LVV: tree removal, topsoil stripping, site roads and storage areas, changes to ditches and culverts, and stormwater arrangements
- Area cleared or prepared, per Tuike's own Sept 25 report to LVV: about 330 ha of the ~571 ha Muhos project area and just under 200 ha of the ~405 ha Kajaani-Otanmäki area. Google confirmed both figures to Yle
- Deadlines: written reply by Oct 14, 2026, stating in particular whether Tuike will comply; work to stop by Oct 23, 2026. The suspension lasts until the EIA procedures are complete; the EIA reports are expected at the end of 2026
- If ignored: LVV may issue a binding order under special legislation or section 35 of the EIA Act and back it with a conditional fine under the Act on Conditional Fines (1113/1990)
- Google to CNBC: 'We understand the concerns and have fallen short of our own high standards in this instance'; it will follow LVV's guidance, says it 'acted in good faith under the Forestry Act', and will plant trees across 130 ha at Muhos
- Plans per the EIA programmes (cited in LVV's notices): Muhos would have nine data-center buildings, over 300 MW of backup power and cooling water from the Oulujoki river; Kajaani-Otanmäki would have eight buildings, a 400 kV line and over 300 MW of backup power
- Trigger: on Sept 17 the Finnish Association for Nature Conservation alleged suspected illegal logging and asked LVV to halt the Muhos works; its chair Hanna Halmeenpää called it 'gross incompetence' by Tuike Finland Oy or 'gross disregard for Finnish legislation, or at worst both' (STT via Kaleva; Futurism). LVV requested reports on Sept 18
- LVV unit head Sari Myllyoja told Yle the call was 'very exceptional' and that it could take several months before Google can restart the halted work
- Context (Bloomberg analysis, Oct 8): planned and active data-center investments in Finland now top €67B, drawn by the cool climate, renewable power and grid connections

#### What happened
On Oct 6, 2026 Finland's Permit and Supervision Agency (Lupa- ja valvontavirasto, LVV) published a call to suspend work, addressed
to Tuike Finland Oy, the Google subsidiary developing both projects. Since Jan 1, 2026 LVV has been the coordinating authority for
environmental impact assessments (EIAs, "YVA" in Finnish). Tuike must stop "immediately and no later than 23.10.2026" all preparatory
work that significantly alters the environment at its planned data centers in Muhos (North Ostrobothnia) and Kajaani-Otanmäki (Kainuu),
until both projects' EIA procedures are complete. Both procedures have been running since summer 2025. Tommi Muilu, head of LVV's
environment department, said that large-scale preparatory work had already been done, including tree removal, topsoil stripping, site
roads and storage areas, changes to ditches and culverts, and stormwater arrangements. In his words, these measures "will change the
environment and cause impacts, the identification and assessment of which are among the key objectives of the EIA procedure" (CNBC's
translation).

The notices rely on reports that LVV requested on Sept 18 from Tuike, the municipality of Muhos and the city of Kajaani. In its Sept 25
reply, Tuike said it had removed topsoil, prepared site roads and temporary storage areas, improved ditches and culverts for
stormwater, and built (at Kajaani, was building) stormwater retention ponds. The felled and prepared area was about 330 ha at Muhos
and just under 200 ha at Kajaani-Otanmäki.
Tuike argued that no permit was needed: the work had started under section 109 of the Building Act, forest-use notifications were filed
under the Forest Act, and the most valuable nature zones were being protected. Tuike must reply in writing by Oct 14 and state whether
it will comply. If it does not comply, LVV may issue a binding order backed by a conditional fine, although its stated first aim is
voluntary correction. Sari Myllyoja, head of LVV's communities and infrastructure unit, told Yle the following:
- The call is "very exceptional".
- It was based on the parties' reports and on drone images published in the media.
- LVV had not yet visited the sites.
- It could take several months before Google can resume the halted work.

Google answered the same day. "We understand the concerns and have fallen short of our own high standards in this instance," a
spokesperson told CNBC, adding that Google would follow LVV's guidance. Sondre Ronander, Google's communications head for Finland and
Norway, told Yle that the company acted in good faith under the Forest Act, carried out nature surveys and protected valuable sites. He
said Google will plant trees across 130 ha at Muhos.

Background:
- In November 2024 Google bought about 1,400 ha of state land in Muhos and Kajaani from Metsähallitus, for an estimated €27M (Yle).
- On Sept 9, 2026 it announced a further €13bn investment in Finland over two years (Yle). The money covers data centers in Kajaani,
  Muhos and Vaala, an expansion at Hamina, and a power deal for up to half the output of the Loviisa nuclear plant. Google called it
  its largest single investment in Europe.
- On Sept 17 the Finnish Association for Nature Conservation told the STT news agency that it suspected illegal logging at both sites
  and asked LVV to halt the Muhos works. STT noted that under Finnish law, permanent land-use-change fellings of more than 200 ha
  need an EIA before they are carried out. Google denied breaking the law at the time, saying the trees were felled in line with the
  Forest Act (STT via Kaleva).

Local leaders accepted the decision (Yle). Kajaani city manager Terho Savolainen said it delays the project but does not change its
basis. Muhos municipal manager Kimmo Hinno said he did not think Google broke the law knowingly, and called a full Google withdrawal "quite
a nightmare". The zoning plans for both sites have been approved, but both are under appeal at the administrative court.

#### Why it matters
Finland has become a key European data-center location, and CNBC notes it has been dubbed the "Texas of Europe" for its available
land and power. This case shows that environmental procedure, not only chips and electricity, can slow hyperscale AI capacity. In particular,
it raises the question of whether large-scale "preparatory" forest clearing may run ahead of an unfinished EIA. An LVV unit head called the
move very exceptional. The next steps are Tuike's Oct 14 reply, the Oct 23 stop deadline, and whether LVV escalates to a binding order
with a fine.

Sources: [The Guardian (6 Oct): Google data centres in Finland, temporary halt over environmental impact assessment](https://www.theguardian.com/technology/2026/oct/06/google-datacentres-finland-temporary-halt-environmental-impact-assessment); [Hacker News discussion (277 points, 11 Oct; submitted as 'Google is illegaly cutting forest to build data centers')](https://news.ycombinator.com/item?id=50048638); [LVV press release (Finnish): Lupa- ja valvontavirasto kehottaa Tuike Finland Oy:tä keskeyttämään Muhoksen ja Kajaanin datakeskushankkeiden valmistelutoimet (Oct 6, 2026)](https://lvv.fi/-/lupa-ja-valvontavirasto-kehottaa-tuike-finland-oy-ta-keskeyttamaan-muhoksen-ja-kajaanin-datakeskushankkeiden-valmistelutoimet); [LVV notice and hearing, Muhos data center (PDF, LVV-U/37128/2026)](https://lvv.fi/documents/242146171/254731747/Kehotus%20ja%20kuuleminen%20Tuike%20Finland%20Oy,%20Muhoksen%20palvelinkeskus.pdf/754fefca-df3b-c546-c36f-7d78c7c66c17?t=1791285171979); [LVV notice and hearing, Kajaani-Otanmäki data center (PDF, LVV-U/22159/2026)](https://lvv.fi/documents/242146171/254731747/Kehotus%20ja%20kuuleminen%20Tuike%20Finland%20Oy,%20Kajaanin-Otanm%C3%A4en%20datakeskushanke.pdf/ada68b44-8901-3cc3-6434-b2208bbbb933?t=1791285254141); [Yle: Google: 'Emme yltäneet tavoitteisiin' – viranomainen laittoi datakeskustyömaat jäihin Muhoksella ja Kajaanissa (Oct 6)](https://yle.fi/a/74-20250109); [Yle: Näin Muhoksen ja Kajaanin kuntajohtajat kommentoivat Googlen työmaiden keskeytystä (Oct 6)](https://yle.fi/a/74-20250150); [CNBC: Google ordered to halt work on two data centers in 'Texas of Europe' (Oct 7)](https://www.cnbc.com/2026/10/07/google-finland-data-center-halt.html); [Xinhua: Finnish authority halts work at Google data center sites over environmental assessment (Oct 7)](https://english.news.cn/europe/20261007/c90bb9bed188430da1c130769453da35/c.html); [AFP via aNews: Finland orders halt to work on Google's data centre sites (Oct 6)](https://www.anews.com.tr/world/2026/10/06/finland-orders-halt-to-work-on-googles-data-centre-sites); [Helsinki Times: Google ordered to halt data centre works in Finland over environmental concerns (Oct 6)](https://www.helsinkitimes.fi/business/29355-google-ordered-to-halt-data-centre-works-in-finland-over-environmental-concerns.html); [PPC Land: Google faces Finnish order to stop work at 2 data centre sites (Oct 6)](https://ppc.land/google-faces-finnish-order-to-stop-work-at-2-data-centre-sites/); [Kaleva (STT): Luonnonsuojeluliitto: Googlen datakeskustyömailla tehty epäiltyjä laittomia hakkuita Muhoksella ja Kajaanissa (Sept 17)](https://www.kaleva.fi/luonnonsuojeluliitto-googlen-datakeskustyomailla-t/13671648); [Futurism: Wolves Descend on Finnish Villages After Google Hacked Down 1,300 Acres of Habitat (Oct 23, citing The Guardian)](https://futurism.com/future-society/wolves-finland-google-ai-data-centers); [Futurism: Google Accused of Illegally Killing a Massive Forest for Its New Data Center Project (Oct 2)](https://futurism.com/artificial-intelligence/google-accused-illegally-killing-forest-data-center); [Yle News: Google announces €13bn additional investment in Finland (Sept 9, 2026)](https://yle.fi/a/74-20245301); [Yle News: Google plans 'historic' data centre expansion in Finland (Nov 28, 2024; ~1,400 ha land purchase)](https://yle.fi/a/74-20128014); [Bloomberg: Finland's data center tally tops 67 billion (euros, per Techmeme) on AI boom (Oct 8, 2026)](https://www.bloomberg.com/news/articles/2026-10-08/finland-s-data-center-tally-tops-67-billion-on-ai-boom)

### 2026-10-06: Google signs ~3.6 GW power deal with Constellation Energy, about a quarter of it nuclear

***; Google, Constellation Energy; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-06-google-constellation-3-6gw-power-deal/

Reuters reported on Oct 6, 2026 that Google agreed to buy about 3,590 MW of electricity from Constellation Energy, with nuclear power making up about 25%, to supply its AI data centers. It is one of the largest single corporate power deals.

- Contracted capacity: ~3,590 MW (~3.6 GW); nuclear ~25% (Reuters)

#### What happened
Google signed a power deal with Constellation Energy covering about 3.6 GW, roughly a quarter of it nuclear (Reuters; details come from the headline summary, and terms and duration are not yet verified).

Sources: [Reuters: Google enters massive 3.6 GW power deal with Constellation Energy](https://www.reuters.com/business/energy/google-enters-massive-36-gw-power-deal-with-constellation-energy-2026-10-06/)

### 2026-10-06: SpaceX seeks $40B in Apollo-led debt to buy Nvidia chips for its AI data centers (FT)

***; SpaceX, SpaceXAI, Apollo Global Management, Nvidia; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-06-spacex-40b-debt-nvidia-chips-apollo/

On Oct 6, 2026 the Financial Times reported that SpaceX is seeking to raise about $40B, roughly $10B in bank loans and $30B in investment-grade debt, led by Apollo, to pay for Nvidia AI chips for its Colossus data centers that run Grok. Talks are early and the deal is expected to close in 2027.

- Size: ~$40B = ~$10B bank loans + ~$30B investment-grade debt (FT, via Reuters/Bloomberg)
- Apollo Global Management expected to lead and place the debt; bond manager Pimco among a small group of lenders in talks
- Use: Nvidia chips for SpaceX's AI computing (Colossus data centers running Grok / SpaceXAI)
- Expected to close in 2027; talks at an early stage and could end without a deal
- Market reaction: SpaceX shares fell ~1% after hours; Nvidia rose ~0.5%
- Context cited by coverage: Musk said in September that Colossus 2 could more than double its Nvidia chip count by December; Morgan Stanley estimates AI infrastructure needs $1.5T of external financing by 2028

#### What happened
The FT story is paywalled; this entry is written from summaries by Reuters, Bloomberg and investingLive. SpaceX, which absorbed xAI and
now runs it as SpaceXAI, wants to fund a large Nvidia chip order mostly with investment-grade bonds rather than equity, with Apollo
arranging it.

#### Why it matters
It is one of the largest debt packages tied to a single AI chip purchase, and it shows frontier compute moving from equity to debt
financing, a trend the Bank of England and others have flagged.

Sources: [Financial Times: SpaceX seeks $40bn to buy Nvidia chips (paywalled)](https://www.ft.com/content/d3f5928d-f38c-4666-8f7a-8737f9c45f51); [Bloomberg: SpaceX Seeks $40 Billion Debt Deal to Buy Nvidia Chips for AI Expansion](https://www.bloomberg.com/news/articles/2026-10-06/spacex-seeking-to-raise-40-billion-to-buy-nvidia-chips-ft-says); [investingLive: SpaceX seeks $40 billion in debt led by Apollo to fund Nvidia chip order, FT says](https://investinglive.com/stocks/spacex-seeks-40-billion-in-debt-led-by-apollo-to-fund-nvidia-chip-order-ft-says/)

### 2026-10-06: Planned US gas power capacity through 2030 rose 44% in eight months, to 60.4 GW, driven mainly by AI data centers (Environment America / Frontier Group)

**; Environment America Research & Policy Center, Frontier Group; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-06-environment-america-gas-plant-plans-up-44-percent-data-centers/

An analysis by Environment America Research & Policy Center and Frontier Group, reported by Electrek on Oct 6, 2026, found that planned grid-connected gas power capacity in the US through 2030 grew from 41.8 GW in December 2025 to 60.4 GW, a 44% rise in about eight months. The groups name AI data centers as the main driver. Planned wind additions fell 33.9% in the same period. Gas plants built on-site for data centers are not counted.

- Planned grid-connected gas additions through 2030: 60.4 GW, up from 41.8 GW in December 2025 (+44%); proposed grid-connected gas capacity has nearly tripled since early 2023 (Electrek, from the groups' analysis of filings with the US Energy Information Administration)
- Planned wind additions fell 33.9% since December 2025; solar still leads planned new capacity
- By state: Texas has 33.2% of the proposed new gas capacity (about 20 GW, almost 25% on top of its existing gas fleet), then Louisiana 6.6 GW, Pennsylvania 4.4 GW, Oklahoma 3 GW, Mississippi 2.6 GW; the top five states hold about 60%
- Not counted: off-grid gas plants that data-center operators build for their own sites, which could add tens of gigawatts
- Health modelling cited for 2028 in a high-growth data-center scenario: nearly 600,000 asthma symptom cases and about 1,300 premature deaths
- Andre Delattre (The Public Interest Network): 'Surging electricity demand to power data centers is putting our nation's transition to clean energy at risk'
- These are planned projects from EIA filings; not all will be built. We read press coverage (Electrek, Futurism), not the groups' own page, and found no release date for the analysis

#### What happened
Two US environmental research groups compared power-plant plans filed with the Energy Information Administration in December 2025 and in
mid-2026. Planned gas capacity rose by almost half, and planned wind fell by a third. They attribute most of the change to electricity
demand from data centers for AI.

#### Why it matters
It puts a number on how the AI build-out changes the US power mix: utilities are planning gas plants that run for decades to meet demand
from data centers. The figures come from advocacy groups and cover plans, not finished plants.

Sources: [Electrek (Michelle Lewis, Oct 6, 2026): US gas plant plans jump 44% in just 8 months – here's why](https://electrek.co/?p=487031); [Futurism (Oct 11): The AI data center boom is causing a gigantic spike in gas power plants](https://futurism.com/future-society/ai-data-center-boom-smog-spewing-gas-power)

### 2026-10-06: openTPU: an open-source AI accelerator 'developed by AI' runs Qwen3.5, Gemma 4 and others on an FPGA card

**; FeSens (independent); Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-06-opentpu-ai-developed-accelerator/

openTPU (github.com/FeSens/openTPU, Apache-2.0) is a full inference accelerator stack (SystemVerilog RTL, ISA, bit-exact simulator, kernel compiler and host software) that its author says was developed by AI agents through an automated improvement loop. It runs ten small open models with real weights on a Kintex-7 FPGA PCIe card, up to ~86 tokens/s on LFM2.5-230M, with the card producing the same tokens as the simulator bit for bit. It reached the Hacker News front page on Oct 6, 2026 (~259 points). The models used are not named.

- Repo: github.com/FeSens/openTPU, created 24 Sept 2026, Apache License 2.0; ~270 stars on 7 Oct; tagline 'An open-source AI accelerator, developed by AI'
- Hardware: Inspur YPCB-00338 card (Xilinx Kintex-7 xc7k480t, two DDR3 channels) at 133.33 MHz; four-column systolic matrix unit; decode is DRAM-bound at 82–94% of DDR3-1066 peak
- Results (README): LFM2.5-230M up to 85.8 tok/s decode (4-bit); Qwen3-0.6B 31.3 tok/s; Qwen3.5-0.8B 24.5 tok/s; Gemma 4 E2B ~12 tok/s; Phi-4-mini 6.6 tok/s; MoE models streamed over PCIe (Qwen3.5-35B-A3B at 3.95 tok/s, LFM2.5-8B-A1B at 10.6 tok/s)
- Stated goal (README): 'how far can AI agents go at hardware design, and can they build the chip that runs their own inference?'; it applies lessons from the author's earlier auto-arch-tournament project (AI-developed RISC-V cores)
- Author on HN (fsbonetto): 'The TPU started able to produce only a few tokens per second and trough a recursive self improvement loop got to 80+ tok/sec on the smallers models'; a 'tournament of Vivado runs' keeps optimising timing and area
- Not stated: which AI models or agents did the work, or how much human design input there was
- Hacker News: ~259 points, ~316 comments (6 Oct 2026)

#### What happened
On 6 October 2026 an open-source project called openTPU reached the Hacker News front page. In one repository it contains an inference accelerator design in SystemVerilog, its instruction set, a bit-exact Python simulator that serves as the specification, a kernel language and compiler, and host tools (`otpu-chat`, `otpu-smi`) that drive an FPGA PCIe card. The README reports measured decode speeds for ten small open models, from Qwen3 and Qwen3.5 to Gemma 4, LFM2/2.5, SmolLM3 and Phi-4-mini, with exact token agreement between card and simulator.

The author (GitHub FeSens, posting as fsbonetto) presents it as hardware "developed by AI". He says the same agent technique had earlier been used to develop RISC-V cores, and that a "recursive self improvement loop" took the design from a few tokens per second to over 80. The repository does not name the models or agents used, and the human share of the design is not documented.

#### Why it matters
It is a small but concrete public example of AI agents carrying a whole hardware stack, from RTL to drivers, to a working FPGA card that runs current open models, with an automated performance loop. On HN, one top comment stressed how much is gained from "an experienced user pointing an LLM in a tasteful direction", and others joked about "recursive self-improvement" arriving as a hobby FPGA project. The claim about AI authorship is self-reported.

Sources: [GitHub: FeSens/openTPU](https://github.com/FeSens/openTPU); [GitHub: FeSens/auto-arch-tournament (earlier AI-developed CPU cores)](https://github.com/FeSens/auto-arch-tournament); [Hacker News discussion](https://news.ycombinator.com/item?id=49980715)

### 2026-10-07: WSJ: Broadcom seeks $50B+ in private credit to finance OpenAI's custom chips; Oracle and SpaceX chase chip debt too

***; Broadcom, OpenAI, Oracle, SpaceX; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-07-broadcom-50b-financing-openai-chips/

On Oct 7, 2026 the Wall Street Journal reported that Broadcom has been working to arrange more than $50B of financing, in early talks with Apollo and Blackstone, to pay for several gigawatts of OpenAI's custom AI chips (Jalapeño and the second-generation Serrano, under a program called Nexus) deployed from late 2026 through 2029. Oracle was in talks with Apollo and Goldman Sachs on an off-balance-sheet vehicle to buy chips and lease them back, and SpaceX was lining up $40B for Nvidia chips.

- Broadcom: more than $50B of private financing for OpenAI's custom chips; early-stage talks with Apollo Global Management and Blackstone; would fund several gigawatts of OpenAI capacity; aimed to close by the end of 2026 (WSJ via investingLive/AI Weekly)
- Chips named in the report: Jalapeño (first generation) and Serrano (second generation), deployed under a program called Nexus between the second half of 2026 and the end of 2029
- Oracle: talks with Apollo and Goldman Sachs to fund a 1 GW data center through a separate entity that buys chips and leases them to Oracle, to bridge the gap between paying for hardware and collecting cloud revenue; early stage
- SpaceX: a $40B package (about $10B bank loans and $30B investment-grade bonds, led by Apollo) to buy Nvidia chips (see 2026-10-06-spacex-40b-debt-nvidia-chips-apollo)
- Context in the report: hyperscalers fund AI with bonds, but OpenAI and Anthropic lack the balance sheets to buy chips outright; one gigawatt of Nvidia chips costs tens of billions of dollars

#### What happened
The WSJ described a wave of bespoke private-credit deals to pay for AI chips. The largest is Broadcom's effort to finance the custom
accelerators it builds with OpenAI, so that OpenAI can deploy several gigawatts of its own silicon without paying up front.

#### Why it matters
OpenAI's compute plans now rely on lenders financing the chips themselves, which ties private-credit funds to OpenAI's revenue. The
report ran a day before the FT's lower OpenAI revenue figure and was cited among the reasons for that day's fall in AI stocks.

Note: the WSJ article is paywalled; details come from secondary summaries (investingLive, AI Weekly, Yahoo Finance). The names
Serrano and Nexus are not otherwise confirmed in this dataset.

Sources: [WSJ: Oracle, Broadcom and SpaceX seek blockbuster debt deals to pay for AI chips (paywalled)](https://www.wsj.com/tech/oracle-broadcom-and-spacex-seek-blockbuster-debt-deals-to-pay-for-ai-chips-848e8032); [investingLive: WSJ: Broadcom seeks over $50 billion for OpenAI chips as Oracle, SpaceX chase AI debt deals](https://investinglive.com/stocks/wsj-broadcom-seeks-over-50-billion-for-openai-chips-as-oracle-spacex-chase-ai-debt-deals/); [AI Weekly: Broadcom lines up $50B+ private-credit deal for OpenAI chips](https://aiweekly.co/alerts/broadcom-lines-up-50b-private-credit-deal-for-openai-chips)

### 2026-10-07: Microsoft and Nvidia launch RTX Spark Windows PCs for local AI; Microsoft Execution Containers for sandboxing AI agents reach general availability

***; Microsoft, Nvidia; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-07-microsoft-windows-rtx-spark-event-mxc/

At an Oct 7, 2026 event in San Francisco, Satya Nadella, Windows chief Pavan Davuluri and Jensen Huang opened preorders for the first Windows PCs built on Nvidia's RTX Spark superchip, aimed at running large AI models locally. These are the Surface Laptop Ultra (from about $2,600) and the Surface RTX Spark Dev Box (about $6,000, 128GB unified memory). Microsoft also made Microsoft Execution Containers (MXC), an MIT-licensed policy layer that sandboxes AI agents on Windows, generally available. It added a Copilot for Windows that manages local files with on-device models.

- Event: Oct 7, 2026, San Francisco; speakers Satya Nadella, Pavan Davuluri (EVP Windows and Devices) and Nvidia CEO Jensen Huang; theme 'how local AI will shape the next chapter of the PC'
- Surface Laptop Ultra: Nvidia RTX Spark (N1X: 18-core CPU, 5,120-core RTX Blackwell GPU, 24GB unified memory in the base model); eight configurations from $2,599 to $5,899; preorders open, shipping from Oct 16 (search summaries of Windows Central/TweakTown; TechCrunch rounds to $2,600–$5,900)
- Surface RTX Spark Dev Box: ~$6,000 ($5,999), 128GB unified memory, up to one petaflop of AI compute, ships with VS Code, GitHub Copilot CLI, WSL and PowerShell 7 (TechCrunch; search summaries)
- Other RTX Spark PCs: Dell XPS 16 Creator Edition at $3,800, preorder at Best Buy for October delivery (TechCrunch); Nvidia first announced RTX Spark in June 2026
- Microsoft Execution Containers (MXC) GA on Oct 7: SDK and JSON policy schema that confines agents (or code they generate) with declared limits on files, network, processes and desktop; four isolation modes: process container (Windows 11, macOS, Linux), session container, WSL container and MicroVM; repo github.com/microsoft/mxc, MIT license, public since Feb 2026
- MXC adopters: Nvidia (integrated OpenShell), GitHub Copilot, OpenAI Codex, Replit, LM Studio, Unsloth; support coming from Anthropic's Claude Code, Box, Egnyte, Heidi Health, Nous Research's Hermes Agent, Manus, Perplexity, Raycast and Simular (Windows Developer blog)
- Still to come: Microsoft Entra attribution of agent vs. user activity, Agent 365 controls for local agents and Intune policy for MXC process containers
- Davuluri: agent actions 'should be attributed to the agent not to you'; Microsoft also showed a Copilot for Windows that searches, renames and moves files using local models and asks before anything leaves the device, plus a redesigned Windows Search
- Nadella interview (Joanna Stern, The New Things, Oct 15): 'I think of all agents as insider risk and treat them as such—or all models as insider risk'; agents 'need to be contained separately', with their own identities, isolated environments, restricted permissions and monitoring, and Windows Defender tracking agent behaviour and shutting down unauthorized actions; on the $2,600+ RTX Spark Surface Laptop Ultra (up to 128 GB): 'last year's frontier is running now on this laptop'

#### What happened
Microsoft and Nvidia promoted "local AI" PCs: laptops and a desktop dev box built on Nvidia's Arm-based RTX Spark chip with large unified
memory, so that sizable models can run on the device. Alongside the hardware, Microsoft shipped MXC, an OS-level containment layer for AI
agents. Coding agents (Codex, GitHub Copilot) already use it, and Anthropic's Claude Code and others have promised support.

Confidence is medium on specs and prices: the main hardware articles (Windows Central, TweakTown, Ars Technica) could not be read in full,
and the figures come from TechCrunch and search summaries, which round differently. The MXC facts come from Microsoft's own developer blog.

#### Why it matters
Agents with computer access have caused real incidents (sandbox escapes, data leaks). Microsoft now ships standard OS sandboxing and
per-agent attribution for agents, and major agent vendors have signed on. The RTX Spark PCs also mark Nvidia's entry into Windows laptop CPUs.

Sources: [Windows Developer blog: Microsoft Execution Containers: Policy-driven containment for AI agents](https://blogs.windows.com/windowsdeveloper/2026/10/07/microsoft-execution-containers-policy-driven-containment-for-ai-agents/); [GitHub: microsoft/mxc](https://github.com/microsoft/mxc); [TechCrunch: Microsoft releases new Nvidia-chip AI PCs with revamped Windows 11](https://techcrunch.com/2026/10/07/microsoft-releases-new-nvidia-chip-ai-pcs-with-revamped-windows-11/); [Ars Technica: Microsoft event debuts new AI-friendly hardware and Windows changes](https://arstechnica.com/gadgets/2026/10/microsoft-event-debuts-new-ai-friendly-hardware-and-windows-changes/); [Engadget: Microsoft Windows and Surface event 2026 live blog](https://www.engadget.com/2279642/microsoft-windows-surface-event-2026-live-blog-nvidia-rtx-spark-laptop-ultra/); [Windows Central: Surface Laptop Ultra with RTX Spark preorders are live](https://www.windowscentral.com/hardware/surface/surface-laptop-ultra-preorders-live); [The New Things (Joanna Stern): Microsoft CEO on rogue AI agents, super laptops and Clippy (Oct 15)](https://thenewthings.com/p/microsoft-ceo-on-rogue-ai-agents-super-laptops-and-clippy); [Joanna Stern on X: interview with Satya Nadella (Oct 15)](https://x.com/joannastern/status/2110729668907376698)

### 2026-10-08: Drones hit Yandex's Sasovo data center, home to two of its AI supercomputers: first major attack on a Russian data hub of the war (followed by hits on Kaluga, Oct 9, and Vladimir, Oct 11)

***; Yandex; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-08-drones-strike-yandex-sasovo-data-center/

On Oct 8, 2026 drones struck a Yandex data center in Sasovo, Ryazan region, starting a fire. Yandex fully suspended the center's operations. It houses two of Yandex's three supercomputers used to train neural networks ("Lyapunov" and "Chervonenkis"). Reuters called it the first major attack on a Russian data hub since the 2022 invasion. It came weeks after Russian strikes on at least eight Ukrainian data centers. Drones then hit Yandex data centers in Kaluga (Oct 9) and Vladimir (Oct 11); Yandex fully suspended the Vladimir site.

- Location: Sasovo, Ryazan region; fire confirmed (NASA fire monitoring, per The Moscow Times); no casualties reported
- Yandex said operations at the center were fully suspended and some services might be unavailable; core consumer services were not affected (Reuters via US News/NBC)
- Downdetector: 1,300+ reports of problems with Yandex Cloud, Documents and Tables; Meduza also lists Yandex Disk and sites such as Cian and Russian Railways
- Supercomputers at Sasovo: Lyapunov and Chervonenkis (Chervonenkis was No. 19 on the TOP500 in 2021); the third, Galushkin, is in Vladimir (Meduza)
- Yandex shares fell 3.4% in morning trading (The Moscow Times)
- Ukraine had not officially commented (Meduza). Since late September, Russian strikes damaged or destroyed data centers of at least eight companies in Ukraine; Russia's Defense Ministry said they processed intelligence data for Ukraine's armed forces
- Second hit within 48 hours (Oct 9): TASS reported that 'several modules of the Yandex data center in Kaluga have been completely taken out of operation as a result of a UAV attack'. The Kaluga site (Grabtsevo industrial park), operating since late 2023, is Yandex's largest Russian data center: about 130,000 m² and 63 MW (UNITED24 Media, citing TASS; HN 165 points)
- Third hit in four days (Oct 11): Yandex said operations at its Vladimir data center were 'completely suspended' after a drone attack and that remaining resources were 'unstable', advising cloud customers to activate disaster-recovery plans (Al Jazeera); its status page reported 'power interruptions in the ru-central1-a availability zone' (Kyiv Independent). Disrupted: Yandex Go, Yandex Pro, smart-home services and internal support, with reports from Russia, Serbia, Kazakhstan, Belarus and Armenia. On Oct 10 Yandex had added drone and missile attacks to its user agreement as force majeure. Zelensky: 'We always respond in mirror-like fashion… they have been hitting our data centres.' Meduza had earlier placed Yandex's third AI supercomputer (Galushkin) in Vladimir; no report seen confirms that the struck building hosts it
- Vladimir strike, more detail (Kyiv Post, Oct 11): the site in Vladimir's Energetik district is a 40-to-50 MW facility designed for up to 2,880 server racks (per ASTRA and Crimean Wind Telegram channels); more than 80 services were disrupted, including the Alice voice assistant, Yandex Music, Telemost, Yandex Smart Home, Yandex Cloud compute, storage and databases, and the AI interfaces YandexGPT API, SpeechKit and Vision OCR; users reported failures in banking apps, ride-hailing and retail chains (Magnit, Fix Price). Governor Aleksandr Avdeev said drones damaged electrical substations. After the first strikes Yandex removed its data-center locations in four regions from its maps
- Traffic data: Cloudflare Radar showed bot traffic from Yandex's network at 0.4% of its two-week maximum after the third strike (about 57% after Sasovo); the Yandex Cloud status page listed at least 84 services offline; NetBlocks put traffic at 30% of normal for Yandex Cloud and 63% for Yandex LLC (Tom's Hardware, Oct 11). NetBlocks' post confirming 'a second collapse' had ~167k views and an Osinttechnical post ~157k (fxtwitter, Oct 12)
- Zelensky (Oct 12, quoted by Futurism): after the strikes 'it suddenly became a little easier to breathe on many social media platforms'; 'there was less hate, less division and fewer provocations because there were fewer bots'. His office then told the BBC he was 'talking about the atmosphere on social media', not a measured fall in bot activity. Cloudflare said its 99% drop in Yandex traffic covers all automated requests, including crawlers and monitoring tools
- Unverified pattern noticed on X (Futurism, Oct 13): likes on Elon Musk's posts averaged 17,000-81,000 a day earlier in October and about 11,000 and 8,000 on the two days after the strikes; Futurism says no causal link can be established. Euronews found no independent verification of claims that AfD-related activity fell
- Ars Technica (Oct 12): the strikes followed a Russian drone campaign against Ukrainian data centers and mobile networks from late September, which cut service for up to 100,000 Kyiv households; Zelensky called the Yandex strikes a 'symmetrical response'. Outages hit Russian payment and banking apps, e-commerce, streaming, ride-hailing and maps. Pro-Kremlin bloggers (translated by the researcher ChrisO_wiki) blame Yandex's 'hyper-centralization' and the 2019 sovereign-internet law, which keeps backups from being hosted abroad

#### What happened
The strike was part of an overnight wave in which Russia said it intercepted 399 Ukrainian drones over 18 regions; refineries in
Bashkortostan and Omsk were also hit. Meduza notes it is unclear whether the data center or a nearby plant was the intended target.

#### Why it matters
AI compute has become a military target on both sides of the war. Losing Sasovo takes most of Yandex's training supercomputing offline,
which affects Russia's main domestic model developer.

Sources: [Jeremiah Fowler on X: 'russian bot traffic has dropped to 49%' (unverified claim, ~2.75M views)](https://x.com/yoda69/status/2108280211863421355); [Cloudflare Radar on X: Yandex Cloud (AS200350) traffic about 33% below the previous week after the Sasovo strike (Oct 8)](https://x.com/CloudflareRadar/status/2108122893028671604); [Kyiv Post: Drones hit Yandex data center in Vladimir, disrupting over 80 services (Oct 11)](https://www.kyivpost.com/post/86708); [Tom's Hardware: Yandex bot traffic nearly eliminated after strikes on its data centers (Oct 11)](https://www.tomshardware.com/tech-industry/data-centers/russian-disinformation-bot-traffic-nearly-eliminated-overnight-after-ukraine-drone-strikes-cripple-multiple-of-russias-google-data-centers-ukraine-intensifies-bombing-of-yandex-data-centers-in-response-to-moscows-strikes-on-kyiv-data-centers); [NetBlocks on X: second collapse in Yandex networks after the Vladimir strike (Oct 11)](https://x.com/netblocks/status/2109151636098818212); [Al Jazeera: Russia's Yandex says data centre in Vladimir shut down after drone attack (Oct 11)](https://www.aljazeera.com/news/2026/10/11/russias-yandex-says-data-centre-in-vladimir-shut-down-after-drone-attack); [Kyiv Independent: Third Yandex data center in Russia shut down following reported drone attack (Oct 11)](https://kyivindependent.com/third-yandex-data-center-in-russia-shut-down-following-reported-drone-attack/); [Yandex Cloud status: incident 2136 (Oct 11)](https://status.yandex.cloud/en/incidents/2136); [Hacker News: 3rd Yandex Cloud Data Center Was Hit (239 points, Oct 11)](https://news.ycombinator.com/item?id=50040547); [UNITED24 Media: Yandex takes a second data center hit in 48 hours (Kaluga)](https://united24media.com/war-in-ukraine/yandex-takes-a-second-data-center-hit-in-48-hours-now-its-biggest-russian-site-is-damaged-23277); [Hacker News discussion (165 points)](https://news.ycombinator.com/item?id=50020927); [Reuters: Drones hit Yandex data centre in first major attack on Russian data hub](https://www.reuters.com/world/drones-hit-yandex-data-centre-first-major-attack-russian-data-hub-2026-10-08/); [US News (Reuters): Drones hit Yandex data centre in first major attack on Russian data hub](https://www.usnews.com/news/world/articles/2026-10-08/drones-hit-yandex-data-centre-in-first-major-attack-on-russian-data-hub); [Meduza: Ukraine strikes Yandex data center housing two of the company's supercomputers](https://meduza.io/en/feature/2026/10/08/ukraine-strikes-yandex-data-center-housing-two-of-the-company-s-supercomputers); [The Moscow Times: Ukrainian drones hit Yandex data center after wave of Russian attacks](https://www.themoscowtimes.com/2026/10/08/ukrainian-drones-hit-yandex-data-center-after-wave-of-russian-attacks-a93915); [Ars Technica: Ukraine's drones knock out AI data center belonging to 'Russia's Google' (Oct 9; headline only, page not readable by the fetcher)](https://arstechnica.com/gadgets/2026/10/ukraines-drones-knock-out-ai-data-center-belonging-to-russias-google/); [Hacker News: Yandex bot traffic down 99% after data center attack (219 points, Oct 12; links Cloudflare Radar AS13238)](https://news.ycombinator.com/item?id=50051532); [Cloudflare Radar: bot traffic for Yandex (AS13238)](https://radar.cloudflare.com/bots/as13238?dateRange=7d#bot-traffic); [Ars Technica: Russia was not ready for the data center war it started with Ukraine (Oct 12, 2026)](https://arstechnica.com/gadgets/2026/10/russia-was-not-ready-for-the-data-center-war-it-started-with-ukraine/); [Futurism: Eyebrows raise when Elon Musk's Twitter engagement shrivels after destruction of Russian data centers linked to bot farms (Oct 13, 2026)](https://futurism.com/future-society/elon-musk-engagement-shrivels-destruction-russian-bot-farm)

### 2026-10-08: Samsung guides to a record ₩107.4T (~$80B) Q3 operating profit, up 783% on AI memory demand, above Nvidia's latest quarter

***; Samsung Electronics; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-08-samsung-q3-record-107-trillion-won-profit/

On Oct 8, 2026 Samsung Electronics' preliminary guidance put Q3 operating profit at ₩107.4 trillion (about $80B) on ₩195T revenue, up 782.5% and 126.6% year on year: the first quarter above ₩100T for any Korean company. Analysts attribute roughly ₩110T to memory (HBM4 for Nvidia's Vera Rubin, double-digit DRAM price rises) while the phone division posted a loss.

- Preliminary Q3 2026: operating profit ₩107.4T (~$80.3B), revenue ₩195T; +782.5% and +126.6% YoY; +20.1% QoQ (Korea Times, Seoul Economic Daily)
- Operating margin ~55.1% (52.2% in Q2); consensus was ₩106.64T profit and ₩200.64T revenue (profit beat, revenue slight miss)
- Memory alone estimated at ~₩110T operating profit; drivers: HBM4 for Nvidia Vera Rubin, server DRAM and NAND shortages, long-term supply agreements
- Device Experience (phones/consumer) expected to post a second straight quarterly operating loss (₩1.1–2.64T) as memory costs raise prices and cut demand
- Seoul Economic Daily: the quarter's operating profit is roughly 20% above Nvidia's latest reported quarterly operating profit ($63.7B); nine-month total ₩254T
- Same week: IDC said global PC shipments fell ~20% in Q3 and Nikkei reported Apple cut iPhone 18 Pro orders, both blamed partly on memory costs (Techmeme Oct 9)

#### What happened
Samsung's quarterly guidance showed how far the AI build-out has pushed memory prices: a hardware maker posting a larger operating profit
than Nvidia, with margins above 50%, while its own phone business lost money because of the same memory costs.

#### Why it matters
Memory (HBM and server DRAM) is now one of the tightest bottlenecks and biggest profit pools of the AI boom, and its cost is spilling
over into consumer electronics (PC shipments down ~20%, iPhone order cuts). Final results are due later in October.

Sources: [Korea Times: Samsung Electronics estimates $80.3 bil. in Q3 operating profit (Oct 8)](https://www.koreatimes.co.kr/business/companies/20261008/samsung-electronics-estimates-803-bil-in-q3-operating-profit); [Seoul Economic Daily: Samsung Electronics Q3 operating profit tops 100 trillion won for first time](https://en.sedaily.com/finance/2026/10/08/samsung-electronics-posts-record-107-trillion-won-quarterly); [CNBC: Samsung forecasts record third-quarter profit of $80 billion as AI boom fuels chip demand](https://www.cnbc.com/2026/10/08/samsung-q3-earnings.html); [IDC: PC tracker 3Q26 press release](https://www.idc.com/resource-center/press-releases/idc-pc-tracker-3q26/)

### 2026-10-08: Qualcomm CEO Cristiano Amon: AI companies ask for phones that run 100B-parameter models all the time by 2028

**; Qualcomm; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-08-qualcomm-amon-ai-labs-100b-parameter-phones/

In an interview with Alex Heath's Sources, published Oct 8, 2026 and recorded at Qualcomm's Snapdragon Summit in Hawaii, CEO Cristiano Amon said AI companies are already asking Qualcomm to support models of at least 100 billion parameters on phones by 2028, with AI running all the time. He expects a smartphone "supercycle" driven by AI agents. The 100B figure is the publisher's paraphrase; we read only the episode page, not a full transcript.

- Request from AI companies, as summarised by Sources: support for at least 100-billion-parameter models on phones by 2028, running continuously. For scale: Qualcomm's September 2026 flagship chip runs a 30B-parameter mixture-of-experts model locally
- Techmeme's headline (Oct 12) adds that Amon said AI companies are 'starting to do phones'; that wording is not on the episode page we read
- Amon expects an upgrade 'supercycle' from agents that use apps, know personal context and act for the user; buyers have delayed upgrades because of high memory prices
- On devices beyond phones and glasses: 'It feels like an iPod moment.' He is most bullish on smart glasses; camera earbuds, jewelry, pendants and badges are in testing, some with Qualcomm technology
- Qualcomm chips will power coming devices from 'AI companies like OpenAI' (Sources); they already power Meta's smart glasses
- Data centers: 'Our DNA is building chips, assuming there's a battery on the other side.' Qualcomm bought Modular (Chris Lattner's AI-software startup; all-stock deal of about $3.92B announced June 24, 2026, per Quartz and Reuters figures) to offer an alternative to Nvidia's software, and Amon said it will 'celebrate every single Modular deployment on Nvidia'
- 6G: demos expected at the 2028 Los Angeles Olympics, early deployments around 2029–2030; he described glasses and other devices as 'walking cameras on this world' feeding context to AI

#### What happened
Alex Heath interviewed Cristiano Amon live at Snapdragon Summit. The episode page says AI companies are asking Qualcomm for phone chips
that can run models of at least 100 billion parameters by 2028, with AI active all the time. Amon also talked about AI wearables, Qualcomm's
move into data-center chips after buying Modular, and 6G.

#### Why it matters
It is a chip supplier's view of what AI labs plan for consumer devices: models several times larger than today's on-device ones, running
continuously on the phone. The claim comes from one interview and is not confirmed by any lab.

Sources: [Alex Heath on X: the Sources interview with Cristiano Amon (Oct 8)](https://x.com/alexeheath/status/2108275202174157044); [Fireside Alpha on X: clip and transcript of Amon's remarks (~641k views)](https://x.com/firesidealpha/status/2108930562794864912); [Elon Musk on X: 'Problem is RAM supply' (reply, ~509k views, Oct 11)](https://x.com/elonmusk/status/2109260288314806434); [Sources (Alex Heath): Qualcomm's CEO on the coming AI phone supercycle (Oct 8, 2026)](https://sources.news/p/qualcomm-ceo-ai-phone); [Quartz: Qualcomm acquires Modular in an all-stock deal (June 24, 2026)](https://qz.com/qualcomm-acquires-modular-ai-software-stock-deal-062426)

### 2026-10-09: Nuvacore, a six-month-old data-center CPU startup from Nuvia founder Gerard Williams, raises at about $2.5B (Reuters)

**; Nuvacore, Sequoia Capital; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-09-nuvacore-raising-2-5b-valuation/

Reuters reported on Oct 9, 2026, citing sources, that Nuvacore, a six-month-old San Jose chip startup with no product yet, is raising hundreds of millions of dollars at about a $2.5B valuation to build a data-center CPU. It was founded by Gerard Williams (who sold Nuvia to Qualcomm for $1.4B in 2021), John Bruno and Ram Srinivasan, and Sequoia led its seed round. The deal reflects how AI agents and accelerator clusters have revived demand for general-purpose server CPUs.

- Reported Oct 9, 2026 by Reuters (exclusive, unnamed sources); round not closed, so valuation and size could change; Nuvacore declined to comment
- Valuation ~$2.5B; amount: 'hundreds of millions of dollars'; Sequoia Capital led an earlier seed round
- Founders: Gerard Williams (ex-Apple; founded Nuvia, sold to Qualcomm for $1.4B in 2021), John Bruno, Ram Srinivasan; HQ San Jose, California; company about six months old, no product yet
- Approach: design the CPU core before committing to an instruction set (x86, Arm or RISC-V), to optimize for data-center and AI workloads (Reuters). AI Weekly names the core 'WarpCore' and says the company claims 'over 95% of the foundational IP' is done before choosing an ISA, and that ex-Apple CPU leader David Williamson joined as SVP of hardware engineering (not confirmed from a primary source)
- Reuters' context: CPUs are in demand as traffic controllers for AI accelerators and to run autonomous AI agents behind chatbots such as Anthropic's Claude

#### What happened
Reuters reported that Nuvacore, founded about six months earlier, is raising a large round at roughly $2.5B before having any product.
Its founders are the team that built Nuvia, whose Arm server-core design Qualcomm bought in 2021. Written from the Reuters syndication
(bdnews24) and AI Weekly; Reuters itself could not be fetched.

#### Why it matters
Agentic AI workloads run much of their tool use, orchestration and sandboxes on general-purpose CPUs, and investors are pricing new server
CPU designers at unicorn valuations before tape-out. It is another sign that AI demand is spreading from GPUs to the rest of the data-center stack.

Sources: [Reuters: Sequoia-backed chip startup Nuvacore raising funds at about a $2.5 billion valuation, sources say](https://www.reuters.com/legal/transactional/sequoia-backed-chip-startup-nuvacore-raising-funds-about-25-billion-valuation-2026-10-09/); [bdnews24 (Reuters syndication)](https://bdnews24.com/business/073przu41q); [AI Weekly: Nuvacore raising funds at $2.5B for data-center CPU design](https://aiweekly.co/alerts/nuvacore-raising-funds-at-25b-for-data-center-cpu-design)

### 2026-10-09: Sabi raises a $50M seed led by Khosla Ventures for a non-invasive 'brain-AI interface' baseball cap that turns imagined speech into prompts

**; Sabi, Khosla Ventures; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-09-sabi-50m-seed-brain-ai-baseball-cap/

On Oct 9, 2026 Sabi announced a $50M seed round led by Khosla Ventures for a baseball cap with up to ~100K non-contact neural sensors. A "brain foundation model" trained on 100,000+ hours of neural recordings decodes imagined speech and intent into text and AI prompts. An early prototype is due at CES 2027. Performance claims (about 30 words per minute targeted) have not been independently verified.

- Seed: $50M led by Khosla Ventures; Accel, Initialized Capital, DST Global, Collaborative Fund, Ascend and Kevin Weil (former OpenAI CPO) participated (company press release, Oct 9, 2026)
- Hardware: custom EEG chip; sensors of 1–5 mm read signals through hair without gel or surgery; Sabi claims 50x the power efficiency of TI's ADS1299; Forbes/Techmeme: 100K non-contact neuro-sensors
- Data and model: 100,000+ hours of labeled neural recordings (from about 100 volunteers, per earlier coverage), which Sabi calls the largest disclosed dataset of its kind, used to train a 'brain foundation model'
- Functions: imagined speech-to-text, thought-to-prompt AI commands, and prediction of keystrokes and on-screen actions; ~30 words per minute targeted (earlier coverage)
- Founders: CEO Rahul Chhabra; CTO Atmadeep Banerjee (co-first author of MindEye, NeurIPS 2023). Sabi came out of stealth in April 2026 with a 'beanie' form factor
- Vinod Khosla: 'You cannot perform a billion brain surgeries to replace a billion keyboards.' Prototype demo planned for CES 2027

#### What happened
Sabi, a California neurotech startup, raised a $50 million seed round to build a wearable, non-invasive brain-computer interface for
everyday AI use. The device is a cap with dense dry EEG-style sensors. Its brain foundation model is trained on a large proprietary set
of neural recordings and aims to turn imagined speech into text and commands for AI assistants. The company has not published accuracy or
speed results for independent review.

#### Why it matters
As AI assistants become the main interface to computers, investors are backing non-surgical "thought-to-prompt" input. Whether scalp EEG
can carry enough signal for real imagined-speech decoding is still an open scientific question.

Sources: [Forbes: Vinod Khosla backs an AI startup building a mind-reading baseball cap](https://www.forbes.com/sites/rashishrivastava/2026/10/09/vinod-khosla-backs-an-ai-startup-building-a-mind-reading-baseball-cap/); [Sabi press release: $50M seed led by Khosla Ventures to put a brain-AI interface inside a baseball cap](https://pinionnewswire.com/press-release/sabi-raises-50-million-seed-round-led-by-khosla-ventures-to-put-a-brain-ai-interface-inside-a-baseball-cap/); [TestingCatalog: How Sabi's brain-reading cap turns thoughts into text](https://www.testingcatalog.com/how-sabis-brain-reading-cap-turns-thoughts-into-text/); [Boing Boing: A Silicon Valley startup wants to sell you a brain-reading beanie (April 2026)](https://boingboing.net/2026/04/16/a-silicon-valley-startup-wants-to-sell-you-a-brain-reading-beanie.html)

### 2026-10-10: WSJ: Dario Amodei personally asked Meta's Alexandr Wang for compute this year; Meta said no

***; Anthropic, Meta; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-10-wsj-ai-compute-hunt-amodei-asked-meta/

A Wall Street Journal feature on the scramble for AI computing power (Oct 10, 2026) reported that Anthropic CEO Dario Amodei spoke with Meta's chief AI officer Alexandr Wang earlier in 2026 hoping to source more compute, and Meta declined. The WSJ says rivals are forming alliances and executives are personally intervening to get capacity. In July CNBC had reported early talks for Anthropic to lease about $10B of Meta compute over two years.

- WSJ (Oct 10, 2026; paywalled, written here from Techmeme's headline and secondary summaries): Amodei spoke with Wang 'earlier this year' to source more compute; Meta declined
- WSJ framing: 'Bitter rivals are forming alliances and executives are having to personally intervene over the scramble for resources to fuel the AI boom'
- Not stated in the summaries: how much capacity Anthropic sought or why Meta said no
- Background (CNBC, July 17, 2026, per secondary reports): Anthropic approached Meta in June about leasing roughly $10B of compute over two years, with monthly payments and an early-exit clause; both sides called the talks preliminary
- Follow-up (The Information, Oct 11–12, paywalled; from the Techmeme headline and the article's opening line only): Anthropic has signed about $530 billion of data-center leasing deals and is pushing partners to speed them up; analysts estimate a 245 MW site could yield about $1.3B a month; sources say Anthropic signed a cloud deal with IREN. Terms of the IREN deal were not readable
- IREN follow-up (Oct 12): neither company confirmed the deal; IREN shares rose about 4% to $36.70. In August IREN had disclosed a multiyear cloud contract with 'a leading frontier AI lab' it could not name. The Information said Anthropic's build-out increasingly relies on current and former bitcoin miners (IREN, Hut 8, TeraWulf, Nscale), per GuruFocus and 24/7 Wall St
- More from the same The Information report, as summarised by MT Newswires (Oct 12; a 'Market Chatter' item that warns it 'may include rumor and speculation'): Anthropic has committed to leasing 15 gigawatts of data-center capacity from AWS, Google, SpaceXAI and other providers; it is to supply Google AI server chips to Hut 8's Louisiana data center from Q2 2027, and opening that site one month earlier 'could add at least $1.3 billion in revenue'; faster capacity could lift revenue and margins before the IPO. Anthropic, Hut 8, IREN, Google and SpaceX did not respond to MT Newswires

#### What happened
The Wall Street Journal published a feature on how the shortage of AI computing power is reshaping Silicon Valley. Its most quoted detail:
Anthropic CEO Dario Amodei asked Meta's Alexandr Wang for compute earlier in 2026, and Meta turned him down. The article is paywalled;
this entry is written from Techmeme's summary and secondary reports. Re-read it if fuller coverage appears.

#### Why it matters
Anthropic has said publicly that compute limits its growth, and Meta is its rival in frontier models. That Amodei asked Meta directly, and
was refused, shows how scarce frontier compute is and how much it shapes who can train and serve the largest models.

Sources: [The Information: The attractive economics behind Anthropic's data center sprint (paywalled)](https://www.theinformation.com/articles/attractive-economics-behind-anthropic-data-center-sprint); [WSJ: The desperate hunt for AI computing power is upending Silicon Valley](https://www.wsj.com/tech/ai/ai-computing-power-demand-a63da9b9); [Techmeme (Oct 9/10): Sources: Dario Amodei spoke with Meta's Alexandr Wang ... Meta declined](https://www.techmeme.com/261009/p33); [AI Weekly: Amodei asked Meta's Wang for compute this year; Meta said no](https://aiweekly.co/alerts/amodei-asked-metas-wang-for-compute-this-year-meta-said-no); [Invezz (July 17, 2026): Meta could soon lease computing power to Anthropic (citing CNBC)](https://invezz.com/in/news/2026/07/17/meta-could-soon-lease-computing-power-to-anthropic/); [Yahoo Finance / 24/7 Wall St: IREN Climbs 4% on Report of Unconfirmed Anthropic Cloud Deal](https://finance.yahoo.com/markets/stocks/articles/iren-climbs-4-report-unconfirmed-123016202.html); [GuruFocus via TradingView: IREN Stock Gains as Report Names Anthropic as Cloud Client](https://www.tradingview.com/news/gurufocus:520a1020b094b:0-iren-stock-gains-as-report-names-anthropic-as-cloud-client/); [MT Newswires via Yahoo Finance: Market Chatter: Anthropic calls on data center partners to speed up capacity buildout (Oct 12)](https://finance.yahoo.com/technology/ai/articles/market-chatter-anthropic-calls-data-125043695.html)

### 2026-10-12: Bloomberg: building and equipping a 1 GW AI data center in Abilene, Texas costs about $30B

***; Bloomberg, Cushman & Wakefield; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-12-bloomberg-1gw-ai-data-center-costs-30b/

A Bloomberg feature published Oct 12, 2026 puts the high-end cost of a one-gigawatt AI data center in Abilene, Texas at about $30 billion: roughly $17B for chips, $3B for power and $2.4B for construction, among other items. The building costs come from Cushman & Wakefield and the hardware estimates from Bloomberg Intelligence. Meta, Amazon, Microsoft and Alphabet expect about $750B of capital spending in 2026.

- Total: about $30B at the high end for a 1 GW site; includes about $17B on chips, $3B on power and $2.4B on construction (Techmeme's summary of Bloomberg)
- Sources of the estimate: Cushman & Wakefield gave Bloomberg a cost breakdown for the shell of a 1 GW data center in Abilene; Bloomberg Intelligence estimated the hardware (LavX summary)
- Meta, Amazon, Microsoft and Alphabet together expect about $750B of capital expenditure in 2026, more than half of what all S&P 500 companies spent on capital projects in 2025
- Data-center memory sells for more than five times its price a year earlier (Bloomberg Intelligence)
- Backlogs for gas turbines, transformers and switchgear extend past 2030
- The US plans more than 60 GW of new data-center capacity over four years, tripling its current total
- Oracle sent a force majeure notice in September 2026 on a large New Mexico data center after a gas pipeline for the site ran into regulatory trouble
- John McWilliams (Cushman & Wakefield) on four companies spending this much in a year: 'We've never seen it'

#### What happened
Bloomberg (Lynn Doan) combined a construction-cost model from the property firm Cushman & Wakefield with hardware estimates from
Bloomberg Intelligence for a hypothetical one-gigawatt site in Abilene, the West Texas town where the first Stargate campus stands.
Hardware is the larger share of the bill, and memory prices have risen more than fivefold in a year. Electrical equipment is the
bottleneck: suppliers of turbines, transformers and switchgear are booked beyond 2030.

#### Why it matters
"Dollars per gigawatt" is the unit in which AI compute deals are now quoted. A figure near $30B per gigawatt, most of it for
hardware that ages within a few years, explains the size of the debt packages in this dataset and why chip and memory prices
drive the economics more than buildings do.

Written from Techmeme's headline and a secondary summary (LavX); the Bloomberg feature is paywalled and was not read. The split of
the remaining roughly $7.6B (networking, memory, cooling and other items) is not given in the sources read.

Sources: [Bloomberg: AI Data Centers Cost Billions to Build. This Is Why](https://www.bloomberg.com/features/2026-ai-data-center-build-costs/); [LavX News: A single AI data center now costs about $30 billion to build (summary of the Bloomberg feature)](https://news.lavx.hu/article/a-single-ai-data-center-now-costs-about-30-billion-to-build)

### 2026-10-12: Microsoft announces Maia 300, its third in-house AI chip, and plans multi-gigawatt Maia deployments from 2027

***; Microsoft; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-12-microsoft-maia-300/

On Oct 12, 2026 Microsoft announced Maia 300, the third generation of its Azure Maia AI accelerator, built for inference (long-context reasoning and agentic tasks). A rack holds 96 accelerators within 167 kW. Microsoft expects 40% more tokens per megawatt than "leading GPU-based systems" and says it will scale Maia to multi-gigawatt deployments starting in 2027, making it the primary inference system for its products. Production deployments of Maia 300 begin in 2027.

- Announced Oct 12, 2026 by Scott Guthrie, Microsoft EVP Cloud + AI; production deployments of Maia 300 begin in 2027
- Plan: 'We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services'
- Microsoft's claims (not independently measured): 40% more tokens per megawatt than 'leading GPU-based systems'; up to 7x the rack-scale token throughput of Maia 200 in the full 96-accelerator configuration; more than 2.5x Maia 200's performance per watt
- Chip: multi-chiplet design; compute chiplets on TSMC's N2P process; eight HBM4 stacks with 288 GB and 18 TB/s of memory bandwidth per accelerator within a 950-watt TDP; support for enhanced MXFP4 and sub-3-bit storage formats
- Rack: 96 accelerators in 24 four-accelerator trays within 167 kW (double Maia 200's density), 27.6 TB of HBM; liquid cooling removes more than 80% of the heat
- Network: Ethernet-based custom transport; up to 128 racks form a 'superpod' of 12,288 accelerators
- Maia 200 (announced January 2026) now serves Microsoft Copilot and the MAI model team; GitHub Copilot and Microsoft Foundry follow by the end of 2026. It runs OpenAI models, Microsoft's MAI models and open-source model families
- Microsoft says Maia 200 delivers up to 1.7x more tokens per watt than leading GPU infrastructure for AI features in Teams, Copilot and Cowork, and is deployed in the US, Europe and Australia
- Mustafa Suleyman (Microsoft AI CEO), to The Verge: 'At that scale we need a platform that we own, and that we co-design with our models.'
- Reactions on X (Oct 12): Mustafa Suleyman: 'Our new silicon is 🔥 40% more tokens per watt!'; Ed Zitron argued that 'multi gigawatt deployments' is worded to suggest gigawatts of Maia chips when it may mean installing them across data centers with that combined capacity, and claimed 'Microsoft only has 2gw of ai data centers' (his figure, not verified)
- Aaron Holmes of The Information said the announcement confirms his August report that Microsoft was 'aiming for >300k chip productions next yr' and hoped to 'win Anthropic and other big customers over' (his X post; the article is paywalled and was not read)
- The 40% baseline is unnamed: Bloomberg's Ed Ludlow noted that 'Microsoft hasn't identified the specific GPU system used for the 40% comparison', only 'leading GPU-based systems'. Analyst Austin Lyons inferred that, since Maia 300 uses HBM4, the comparison is with an HBM4 system such as Nvidia's Vera Rubin (his reading, not Microsoft's statement)
- Analysts (X, Oct 12): Patrick Moorhead: 'We haven't tested it, and GPUs won't stand still before 2027. At this scale, Maia is primarily a margin lever, not a benchmark story.' Ben Bajarin relayed further Microsoft projections: 1.4 times more tokens per gigawatt and 30% better tokens per dollar (he does not state the baseline for these two), with 'Current deployments … centered on 64- to 128-accelerator clusters, with 12,000 accelerators representing a current plan of record'
- Brendan Burke (Futurum) described how Maia is used: 'agentic kernel generation with Maia silicon in the loop', and Microsoft's Superintelligence team running RL rollouts and synthetic data generation on Maia 200 while pretraining stays on NVIDIA (analyst's account, not checked against Microsoft's documents)
- Scott Guthrie, in a live interview with Deirdre Bosa on Oct 12: 'In a world where power is constrained, land is constrained, and materials are constrained... everything you can do to achieve more with less ultimately helps in the build out phase that we're in right now'
- Reach on X (fxtwitter, Oct 13): Chamath Palihapitiya's 'everyone competes with everyone' post about 335k views; Elon Musk replied that the constraint is 'Scaling power harnessed, along with super fast construction, cooling, networking and batteries. This is hard.' (about 189k views). ETH Zurich's Torsten Hoefler linked the architecture paper, arXiv 2608.24664 ('Our SDLA is a great architecture for LLMs'; paper not read for this entry)

#### What happened
Scott Guthrie, who runs Microsoft's Cloud + AI group, announced Maia 300 on the company blog, and the Azure hardware team published an
architecture paper. Maia 300 is an inference system, not a training chip. Microsoft describes it as "a complete intelligent system
designed for the new types of AI workloads emerging today: Long-context reasoning, latency-sensitive conversational experiences and
sustained, multi-step agentic tasks."

The design splits compute, memory-controller and I/O functions into separate chiplets, with compute on TSMC's N2P node. Each
accelerator has eight HBM4 stacks (288 GB, 18 TB/s). Compared with Maia 200, Microsoft states 36% more compute FLOPs, 2.5 times the
memory bandwidth and 33% more memory capacity. A rack holds 96 accelerators in 167 kW, and an Ethernet-based network joins up to 128
racks into one scale-up domain of 12,288 accelerators.

Microsoft says the two Maia generations will serve GitHub coding agents, Microsoft Foundry model serving, inference and data generation
for its MAI model team, and Dragon Copilot. The Reuters report was read only as a headline (the site blocks automated access); the
numbers here come from Microsoft's two posts and The Verge.

#### Why it matters
Microsoft is the largest buyer of Nvidia GPUs among cloud companies, and it now states that its own chip will be its main inference
system from 2027. All performance figures are Microsoft's own and compare against unnamed GPU systems. The announcement came nine months
after Maia 200, which shows a faster release rhythm for hyperscaler chips, alongside Google's TPUs, Amazon's Trainium and the custom
chips Broadcom builds for OpenAI and Anthropic.

Sources: [Official Microsoft Blog: Microsoft to expand Azure Maia system to multi-gigawatt deployments (Scott Guthrie)](https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/); [Azure Blog: Deep dive into Azure Maia 300 architecture](https://azure.microsoft.com/en-us/blog/deep-dive-into-azure-maia-300-architecture/); [The Verge: Microsoft's new AI chip makes it less reliant on Nvidia's GPUs](https://www.theverge.com/news/1009465/microsoft-maia-300-chip-improvements-rollout); [Reuters: Microsoft plans huge deployment of its own AI chips with an eye on saving costs](https://www.reuters.com/business/microsoft-plans-huge-deployment-its-own-ai-chips-with-an-eye-saving-costs-2026-10-12/); [Satya Nadella on X: 'our biggest bet yet on our own silicon' (1.39M views on Oct 13)](https://x.com/satyanadella/status/2109631106404487523); [Microsoft on X: 'Introducing Maia 300'](https://x.com/Microsoft/status/2109634232163471782); [Mustafa Suleyman on X: 'Our new silicon is 🔥'](https://x.com/mustafasuleyman/status/2109655335250366752); [Ed Zitron on X: on the 'multi gigawatt deployments' wording](https://x.com/edzitron/status/2109674961569009915); [Aaron Holmes (The Information) on X: announcement confirms August report](https://x.com/aaronpholmes/status/2109639265504358491); [Ed Ludlow on X: the GPU baseline for the 40% claim is not named](https://x.com/EdLudlow/status/2109648408973807874); [Patrick Moorhead on X: 'a margin lever, not a benchmark story'](https://x.com/PatrickMoorhead/status/2109635430769475668); [Ben Bajarin on X: rack, networking and tokens-per-dollar details](https://x.com/BenBajarin/status/2109655386727108698); [Brendan Burke (Futurum) on X: software-configured dataflow and agentic kernel generation](https://x.com/BrendanBurkeX/status/2109657608995188890); [Deirdre Bosa on X: Scott Guthrie on the 'cone of uncertainty'](https://x.com/dee_bosa/status/2109783379625513328); [Deirdre Bosa's live interview with Scott Guthrie (YouTube)](https://www.youtube.com/watch?v=4dVWeo6VKyI); [Chamath Palihapitiya on X: 'everyone competes with everyone' (about 335k views)](https://x.com/chamath/status/2109658962807386330); [Elon Musk on X: the constraint is power, construction, cooling, networking and batteries](https://x.com/elonmusk/status/2109668152754020609); [Torsten Hoefler on X: 'Our SDLA is a great architecture for LLMs'](https://x.com/thoefler/status/2109657493341372497); [arXiv 2608.24664: Maia 200: A Software Defined Dataflow System for Large-scale AI Acceleration (25 Aug 2026; linked by Hoefler, body not read)](https://arxiv.org/abs/2608.24664); [Andreas Schilling on X: Maia 300 chiplet configuration](https://x.com/aschilling/status/2109633438672458119)

### 2026-10-12: AMD announces Helion, an open rack design for AI agents with up to 32,768 CPU cores; Intel shows Crescent Island GPU and 46,000-core Xeon racks at OCP

**; AMD, Intel, Open Compute Project; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-12-amd-helion-agentic-ai-rack-ocp/

At the OCP Global Summit on Oct 12, 2026 AMD announced Helion, an open rack blueprint for running AI agents: a CPU-only rack with up to 32,768 cores of 6th-generation EPYC, and an all-in-one rack that adds Instinct MI350P GPU nodes. Configurations are planned for November 2026. Intel announced Xeon 6+ rack designs with more than 46,000 cores and gave specifications for its Crescent Island inference GPU (up to 480 GB of memory, 350 W).

- AMD Helion: 'an open blueprint for agentic AI infrastructure' based on an OCP spec-design and the Helios rack platform, with an integrated token router that directs agents' model calls across frontier and open-source models (AMD)
- CPU-only rack: 6th Gen AMD EPYC, up to 32,768 cores, aimed at hyperscalers that run agent execution next to separate inference systems; all-in-one rack: EPYC CPU nodes plus Instinct MI350P GPU nodes, aimed at enterprises; liquid- and air-cooled versions
- Availability: configurations planned for November 2026; AMD plans to contribute the Helion specification to OCP
- Partners named by AMD: Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn
- Dan McNamara (AMD): 'Agentic AI is not a single workload.'
- Intel GPU D-series (Crescent Island): Xe3P architecture, up to 480 GB LPDDR5X (160 GB on the Intel-branded card), 1.5 TB/s memory bandwidth, 350 W, air-cooled PCIe card; Intel says one card supports models up to 300 billion parameters; customer sampling under way, no release date
- Intel Xeon 6+ racks on Open Rack v3: selected configurations exceed 46,000 cores per rack at 133.45 kW; Intel also announced an AI for Enterprise Agent Toolkit and day-0 support for Meta's Muse Glimmer
- Intel GPU D270 product page (the first named D-series card): 32 Xe3P cores and 256 XMX matrix engines, up to 480 GB LPDDR5X per card, 350 W air-cooled, standard PCIe 5.0, vLLM and PyTorch support; Intel's news release adds 1.3 petaflops of FP8 compute and 'more than one trillion parameters in an eight-GPU system'. No price. 299 points on Hacker News, where commenters compared the 1.5 TB/s memory bandwidth with about 1.8 TB/s for an RTX 5090 and regretted the lack of a consumer card
- Reach: Intel's D270 product page reached the Hacker News front page on Oct 12 (300 points, 172 comments by Oct 13), mostly for the 480 GB memory figure

#### What happened
Both x86 vendors used the Open Compute Project summit to present racks built around CPUs for agent workloads, on the argument that
agents spend much of their time running tools, code and orchestration, which need processor cores more than GPUs. AMD's Helion
sits beside its GPU rack Helios. Intel paired dense Xeon racks with a memory-heavy inference card.

#### Why it matters
Both companies now sell rack designs specifically for "agentic" workloads. They expect a growing share of AI data-center
spending to go to CPU capacity for agents, separate from model inference. No customers or prices were
announced.

The Register's article is linked but was not read.

Sources: [AMD: Introducing AMD Helion, Agentic AI Compute at Rack Scale](https://newsroom.amd.com/news/open-rack-blueprint-scale-agentic-ai/); [Intel: Intel Accelerates Open Infrastructure for the Agentic AI Era](https://www.intel.com/content/www/us/en/newsroom/news/artificial-intelligence/intel-accelerates-open-infrastructure-agentic-ai-era.html); [The Register: AMD unveils Helion racks for AI agents at OCP summit](https://www.theregister.com/systems/2026/10/12/amd-unveils-helion-racks-for-ai-agents-at-ocp-summit/5302652); [Intel: GPU D270, inference-optimized data center GPU (product page)](https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/d-series.html); [Hacker News: Intel D270 GPU AI Accelerator, up to 480 GB LPDDR5X (299 points)](https://news.ycombinator.com/item?id=50063365); [Hacker News: Intel D270 GPU AI Accelerator, up to 480 GB LPDDR5X](https://news.ycombinator.com/item?id=50063365)

### 2026-10-12: OCP Global Summit: Nvidia contributes Vera Rubin rack designs; hyperscalers publish open data-center and grid-resilience specs

**; Nvidia, Open Compute Project, Meta, Google, Microsoft, AMD, Cisco; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-12-nvidia-ocp-vera-rubin-rack-designs-open-specs/

At the Open Compute Project Global Summit in San Jose on Oct 12, 2026, Nvidia contributed its Vera Rubin rack and compute-tray designs to OCP through the open MGX architecture (MGX 1.2 supports racks up to 250 kW). Nvidia, AMD, Google and Meta published an Open Data Center Specification v0.7, and Google, Meta, Microsoft and Nvidia published a Grid Resilience Specification for how AI data centers behave during grid disturbances. Nvidia and Cisco announced a new pluggable optics format, USFP.

- Nvidia contributed 'key NVIDIA Vera Rubin rack and compute tray designs' to OCP via MGX; MGX 1.2 adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000A busbar and support for racks up to 250 kW
- Open Data Center Specification v0.7 (Nvidia, AMD, Google, Meta): a common blueprint for AI data-center power, cooling and layout
- OCP Grid Resilience Specification (Google, Meta, Microsoft, Nvidia): large AI sites should stay connected during grid disturbances, respond automatically and restore to near-full load in under two seconds
- Meta adopts Nvidia Spectrum-6 102.4 Tb/s Ethernet switches (its N-IceCube switch, running FBOSS); 'multiplane' networking is aimed at clusters of more than 100,000 GPUs
- Nvidia and Cisco: Ultra-Dense Small Form Factor Pluggable (USFP), described as the first module for 400 Gb/s-based networking fabrics
- Nvidia and Google: Falcon Transport support on future ConnectX-10 SuperNICs and BlueField DPUs
- New open-source DSX OS components from Nvidia: Config Manager, Cluster Readiness Engine, Rack Management Service, Cluster API provider

#### What happened
Nvidia used the OCP summit to open parts of its next rack platform and to publish facility standards together with its largest
customers. Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, presented the announcements. Visionbay
also showed plans for "K-1", a data-center design built around Nvidia's 800-volt DC power architecture.

#### Why it matters
Gigawatt-scale sites need suppliers of racks, power gear and cooling to build to one target. Open rack and facility specifications let
manufacturers start before each customer issues its own design. The grid specification answers a concern of utilities: very large AI
loads that disconnect suddenly during a fault can destabilise the grid. The same day Microsoft detailed its own 167 kW Maia 300 rack,
which shows the power density these standards are written for.

Sources: [NVIDIA Blog: OCP Global Summit: Amid Ecosystem Adoption, NVIDIA Opens Vera Rubin Architecture to Help Scale AI Factories Worldwide](https://blogs.nvidia.com/blog/ocp-global-summit-vera-rubin-scale-ai-factories/); [NVIDIA Newsroom: NVIDIA and Cisco Introduce New Pluggable Optics Form Factor to Scale AI Factories](https://nvidianews.nvidia.com/news/nvidia-and-cisco-introduce-new-pluggable-optics-form-factor-to-scale-ai-factories); [Cisco Blogs: Cisco and NVIDIA Announce USFP: Expanding the Path to AI Connectivity](https://blogs.cisco.com/news/cisco-and-nvidia-announce-usfp-expanding-the-path-to-ai-connectivity)

### 2026-10-13: IDC: Nvidia's share of China's AI data-center chips fell to 34% in H1 2026; Huawei reached 28%

***; IDC, Nvidia, Huawei; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-13-idc-nvidia-china-ai-chip-share-34-percent-h1-2026/

IDC data reported by Reuters on Oct 13, 2026 show Nvidia shipped 678,000 data-center AI chips in China in January to June 2026, down 44% from a year earlier, for a 34% market share. Huawei shipped 572,000 (28%), up from just over 300,000. Non-GPU accelerators (ASICs), almost all from Chinese vendors, overtook GPUs in volume for the first time in IDC's count.

- Nvidia, H1 2026: 678,000 chips shipped in China, -44% year over year; share 34% (IDC via Reuters)
- Huawei, H1 2026: 572,000 chips, share 28%; up from 'more than 300,000' in H1 2025
- GPU shipments fell 26% to 991,000 units; non-GPU accelerators rose nearly 80% to 1.016 million and took the volume lead
- Huawei, Alibaba's T-Head, Baidu's Kunlunxin and Cambricon made up 95% of the non-GPU segment; Kunlunxin, Cambricon and Hygon gained share
- Implied total market: about 2.0 million accelerators in the half (sum of the two segments; our arithmetic)
- Comparison: for full-year 2025 IDC had Nvidia at about 2.2 million cards and 55% of roughly 4 million, Chinese vendors at 41%, Huawei at about 812,000 (Reuters, April 1, 2026)
- Causes named: US export controls limit what Nvidia can sell, and China often requires import approvals even for compliant parts

#### What happened
Reuters reported new figures from the research firm IDC on China's market for AI accelerators used in data centers. In the first
half of 2026 Nvidia's shipments fell by 44% and its share dropped to 34%, from 55% for the whole of 2025. Huawei nearly doubled its
shipments and now stands six points behind. The fastest-growing part of the market is non-GPU accelerators, where four Chinese
suppliers hold 95%.

#### Why it matters
China was once a market Nvidia dominated with over 90%. The IDC numbers show that export controls and Beijing's own import
approvals have moved a majority of the volume to domestic chips within about 18 months. Unit counts say nothing about performance
per chip, so Nvidia's share of computing power may differ from its share of units.

Written from Finimize's summary of the Reuters report and Techmeme's headline; reuters.com blocks automated access. The shares of
T-Head, Kunlunxin, Cambricon and Hygon were cut off in the readable copy.

Sources: [Reuters: Nvidia's China AI-chip share falls as Huawei closes gap, report finds](https://www.reuters.com/business/retail-consumer/nvidias-china-ai-chip-share-falls-huawei-closes-gap-report-finds-2026-10-13/); [Finimize: Nvidia loses ground in China's AI chip market (summary of the Reuters report)](https://finimize.com/content/nvidia-loses-ground-in-chinas-ai-chip-market); [Reuters via Yahoo Finance: Chinese chipmakers claim nearly half of local market as Nvidia's lead shrinks (IDC, 2025 data)](https://finance.yahoo.com/sectors/technology/articles/chinese-chipmakers-claim-nearly-half-091441324.html)

### 2026-10-13: Nikkei: NTT Data in final talks to build a 500 MW, $6.3B data center in Osaka for Anthropic

***; NTT Data, Anthropic; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-13-ntt-data-anthropic-osaka-data-center-talks/

Nikkei reported on Oct 13, 2026, citing multiple sources, that NTT Data Group is in the final stages of talks to build a data center in Osaka for Anthropic. It would be Anthropic's first dedicated data center in Japan, with about 500 MW of power capacity, among the largest in the country. Construction could cost about 1 trillion yen ($6.3B), with a partial opening by 2028.

- Nikkei: NTT Data Group 'is in the final stages of talks to build a data center in western Japan for U.S. AI company Anthropic, multiple sources have told Nikkei'; the center 'would have one of the largest processing capacities in the country'
- Cost: could reach $6.3 billion (about 1 trillion yen) to construct; servers and GPUs would cost extra (W.Media, citing Nikkei)
- Capacity: up to 500 MW (Techmeme's summary of Nikkei; W.Media)
- Timeline: the companies aim to conclude talks by the end of NTT's fiscal 2026 (March 2027); partial operation by 2028
- Power: getting 500 MW from the grid is a challenge; on-site generation and fuel cells are being considered as a start (W.Media, citing Nikkei)
- Neither company has confirmed the talks (W.Media); Reuters carried the report as 'NTT Data in final talks to build Anthropic's first data centre in Japan, Nikkei reports'

#### What happened
Nikkei's report (by Kosuke Shimizu, Katsuyuki Tanimoto and Shugo Tamura) says the deal is not signed yet. NTT Data would build and
operate the site in Osaka for Anthropic. The open part of the Nikkei article gives the $6.3B cost and the 2028 partial opening;
the 500 MW figure and the power plans come from Techmeme's summary and W.Media's account of the full article.

#### Why it matters
It would be Anthropic's first dedicated capacity in Japan and one of its largest sites outside the United States, next to the
Norway (Volta/Bitdeer) contract and the Australian tender. For Japan, a single 500 MW AI site is close to one-seventh of the
country's installed data-center power (3.57 GW in 2026, per Mordor Intelligence as quoted by W.Media).

Sources: [Nikkei Asia: NTT to build first data center for Anthropic in Japan](https://asia.nikkei.com/business/technology/artificial-intelligence/ntt-to-build-first-data-center-for-anthropic-in-japan); [W.Media: NTT Data in talks with Anthropic to build US$6.3 billion data center in Osaka](https://w.media/ntt-data-in-talks-with-anthropic-to-build-us-6-3-billion-data-center-in-osaka/); [Reuters: NTT Data in final talks to build Anthropic's first data centre in Japan, Nikkei reports](https://www.reuters.com/business/media-telecom/ntt-data-final-talks-build-anthropics-first-data-centre-japan-nikkei-reports-2026-10-13/)

### 2026-10-13: Bastrop County, Texas asks the state for a public hearing on a 710 MW gas plant that would power a data center reportedly leased by Anthropic

**; Anthropic, Pacifico Energy, BlackChamber; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-13-bastrop-county-pacifico-gas-plant-anthropic-data-center-hearing/

On Tuesday Oct 13, 2026 Bastrop County commissioners unanimously asked the Texas Commission on Environmental Quality to hold a public meeting on the air permit for Pacifico Energy's proposed 48-turbine, 710 MW gas plant in Cedar Creek. Public filings say it would power a $2.2B+ data-center campus developed by BlackChamber whose primary tenant is reportedly Anthropic. Residents have opposed the project for months; the vote does not force TCEQ to hold the meeting.

- Plant: 48 gas turbines, 710 MW generating capacity; Bastrop County's 44,700+ homes need about 224 MW at peak (Austin American-Statesman)
- Campus: over $2.2B, at least two data-center buildings, on 2,800+ acres off Earl Callahan Road; developers say about 600 acres will be developed. Developer BlackChamber; 'primary tenant is reportedly Anthropic', according to public filings
- First building 'Project Longhorn': filed with the Texas Department of Licensing and Regulation in October, ~500,000 sq ft, ~$366M; construction to start late November 2026 and finish January 2028
- Water: the utility allows up to 6,900 gallons of potable water a day; the developers say they plan about 2,000 gallons a day
- Process: the developers dropped a tax-abatement request after months of public comment, which removed the requirement for a public hearing; the county's resolution asks TCEQ to hold one anyway

#### What happened
After months of resident opposition in Cedar Creek, Bastrop County's commissioners voted unanimously on Tuesday to ask the state
environmental regulator for a formal public meeting on the gas plant's air permit. The plant would power a large data-center campus
near family neighbourhoods, farmland and a middle school. Anthropic has not publicly confirmed it is the tenant; the link comes from
public filings as reported by the Statesman.

#### Why it matters
It is one of the first local fights over on-site gas generation tied by name to a frontier AI lab, as data-center opposition grows into
a bipartisan political issue before the midterms.

Sources: [Austin American-Statesman: Controversial Bastrop gas plant for Claude maker Anthropic faces new pressure](https://www.statesman.com/business/technology/article/bastrop-claude-anthropic-data-center-gas-plant-22472993.php)

### 2026-10-14: ASML: planned lithography supply for 2027 and 2028 will not fully meet AI-driven demand; customers ask about 2029–2030 orders

***; ASML; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-14-asml-q3-2026-supply-short-of-ai-demand-2027-2028/

Reporting Q3 2026 results on Oct 14, ASML said that "our supply as currently planned for both 2027 and 2028 will not fully meet the demand of our customers", because chipmakers keep raising equipment forecasts on the size and length of AI investment, and that some customers want to place orders for 2029 and even 2030. Q3 sales were a record €11.7B; Q4 guidance is €14.2–15.2B.

- Q3 2026: net sales €11.7B (€11,662M, a record), gross margin 57.5%, net income €5.0B (€4,960M); installed-base management sales €3.3B
- Q4 2026 guidance: net sales €14.2–15.2B, gross margin about 57%; full-year 2026: €44–45B, gross margin about 56%
- CEO Christophe Fouquet: 'we now expect that our supply as currently planned for both 2027 and 2028 will not fully meet the demand of our customers'; 'some customers are expressing an interest in placing orders with us for 2029 and even 2030'
- Earlier plan (Q2 release, July): add 30% to 2026 low-NA EUV capacity of about 65 systems for 2027, and study another 30% for 2028
- Q3 net bookings were not in the press-release text we read

#### What happened
ASML, the only maker of EUV lithography machines, reported a record quarter and warned that it cannot build enough tools for its
customers' 2027–2028 plans, which are driven by AI. Some customers are already asking about slots for 2029–2030.

#### Why it matters
Lithography capacity is one of the slowest-moving limits on how much AI compute can be built. ASML saying it will be short for two years
is a signal from the bottom of the supply chain that chipmakers expect AI capex to last, a day after an Asian chip sell-off driven by
bubble worries.

Sources: [ASML Q3 2026 results press release (SEC Form 6-K, Oct 14)](https://www.sec.gov/Archives/edgar/data/0000937966/000162828026065813/pressreleasefinancialresul.htm); [Reuters: ASML forecasts Q4 sales above expectations on AI chip demand](https://www.reuters.com/business/asml-forecasts-q4-sales-above-expectations-ai-chip-demand-2026-10-14/); [CNBC: ASML Q3 earnings, AI demand, orders and capacity](https://www.cnbc.com/2026/10/14/asml-3q-earnings-ai-demand-orders-capacity.html)

### 2026-10-15: TSMC posts record Q3 2026 profit on AI chip demand, guides Q4 revenue to $52–54B and confirms talks over Musk's Terafab on condition that TSMC runs any fab

***; TSMC; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-15-tsmc-q3-2026-record-profit-terafab-talks/

On Oct 15, 2026 TSMC reported Q3 revenue of NT$1,494.24B (US$46.72B) and net income of NT$763.06B, a sixth straight quarterly profit record driven by AI chips. It guided Q4 revenue to US$52.3–54.3B. CEO C.C. Wei said TSMC is discussing a role in Elon Musk's Texas Terafab project but that "any fab should be run by TSMC". The record came as AI-funding worries hit chip stocks.

- Q3 2026: revenue NT$1,494.24B (US$46.72B); net income NT$763.06B; diluted EPS NT$29.43 (US$4.60 per ADR); gross margin 66.4% (TSMC)
- CNBC: revenue +50.9% YoY and net income +68.7% (~US$23.7B), both above estimates; sixth straight quarterly profit record
- Wafer revenue by node: 3nm 33%, 5nm 28%, 7nm 10%, 2nm 8%; 7nm and below 79%
- Q4 guidance: revenue US$52.3–54.3B, gross margin 65–67%, operating margin 56.5–58.5%
- Terafab: talks at 'the discussion stage'; Wei: 'any fab should be run by TSMC to control production'; Musk said TSMC could sublease part of Terafab, 'nothing more than that' (Focus Taiwan)
- Capacity: 2nm capacity to more than double in 2027; A14 trial production 2027, mass production 2028; nearly 20 fab projects including three new 3nm fabs (Tainan H1 2027 and H1 2028); second Arizona 2nm fab mass production Q3 2027
- Same day: SK Hynix and Samsung fell about 10% in Seoul on AI-funding and rate worries (CNBC)

#### What happened
TSMC's Q3 results beat estimates again. Advanced nodes made up 79% of wafer revenue, with 3nm the largest single node. On the
earnings call C.C. Wei said the company is discussing a possible role in Musk's Terafab chip-plant project in Texas, but any
fab it joins must be run by TSMC. He also laid out a large expansion: 2nm capacity more than doubling in 2027, A14 in 2028 and
new 3nm fabs in Tainan.

#### Why it matters
TSMC makes nearly all leading AI accelerators, so its numbers are the clearest read on AI hardware demand. They stayed at records
even as Asian memory stocks sold off on AI-funding fears the same day.

Note: CNBC's US$46.36B revenue figure differs from TSMC's US$46.72B, probably because of exchange-rate conversion; we use TSMC's.

Sources: [TSMC: Third Quarter 2026 earnings release](https://pr.tsmc.com/english/news/3346); [CNBC: TSMC third-quarter profit beats on global AI demand](https://www.cnbc.com/2026/10/15/tsmc-third-quarter-profit-global-ai-.html); [Focus Taiwan: TSMC on Terafab, US expansion and new fabs](https://focustaiwan.tw/business/202610150023); [CNBC: SK Hynix, Samsung lead Asia tech rout on AI funding worries](https://www.cnbc.com/2026/10/15/sk-hynix-samsung-softbank-stocks-asia-tech-rout-ai-funding-worries.html)

### 2026-10-15: Bloomberg: ByteDance in advanced talks to buy data-center AI ASICs from MediaTek, which would become MediaTek's second big ASIC customer after Google

**; ByteDance, MediaTek; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-15-bytedance-mediatek-data-center-asic-talks/

Bloomberg reported on Oct 15, 2026 (unnamed sources) that ByteDance is in advanced talks to buy custom data-center chips (ASICs) from Taiwan's MediaTek, which would make it MediaTek's second major customer for such chips after Google. Volumes and terms were not given; Reuters, which relayed the report, could not verify it. ByteDance is diversifying away from export-restricted Nvidia GPUs.

- Bloomberg (via Techmeme headline): 'ByteDance is in advanced talks to buy data center ASICs from MediaTek, potentially becoming its second major customer for such chips after Google'
- Reuters relay (Business Recorder): 'ByteDance is close to an agreement with MediaTek to secure semiconductors for its data center operations'
- Not disclosed: quantity, chip generation, foundry or price. Google's TPU 8i was reported in April to be designed with MediaTek
- Context (Crypto Briefing, unverified secondary): earlier 2026 reports of ByteDance talks with Iluvatar CoreX (50,000 chips) and a Qualcomm ASIC deal

#### What happened
Bloomberg's sources say ByteDance and MediaTek are close to a supply agreement for application-specific AI chips for ByteDance's data
centers. The Bloomberg article is paywalled; this entry relies on the Techmeme headline and the Reuters relay.

#### Why it matters
Custom ASICs from a Taiwanese designer are one route for China's largest AI spender to get compute while Nvidia's best GPUs are restricted.
For MediaTek it would add a second hyperscale ASIC customer next to Google.

Sources: [Bloomberg: ByteDance in talks to secure MediaTek AI chips for data centers (paywalled)](https://www.bloomberg.com/news/articles/2026-10-15/bytedance-in-talks-to-secure-mediatek-ai-chips-for-data-centers); [Techmeme summary](https://www.techmeme.com/261014/p57); [Business Recorder (Reuters): ByteDance nears chip supply deal with MediaTek for data centres](https://www.brecorder.com/news/40444327/bytedance-nears-chip-supply-deal-with-mediatek-for-data-centres-bloomberg-news-reports)

### 2026-10-15: Taiwan plans state-backed US industrial parks for AI suppliers, starting with a Texas complex of up to $4 billion

**; Taiwan Ministry of Economic Affairs, Nvidia, TSMC, Foxconn; Chips & compute; confidence medium; https://postcutoff.com/e/2026-10-15-taiwan-texas-industrial-parks-ai-supply-chain/

On Oct 15, 2026 Taiwan's Minister of Economic Affairs Kung Ming-hsin outlined plans for Taiwanese-funded industrial parks in the US, each costing $2–4 billion, to help smaller suppliers to Nvidia, TSMC and Foxconn set up there. The first, in Texas and costing up to $4 billion, already has land secured. The plan sits inside the US–Taiwan trade deal, in which Taiwan offered up to $250 billion of credit guarantees for US chip investment.

- Announced by Minister of Economic Affairs Kung Ming-hsin (Oct 15; Bloomberg per Techmeme, Crypto Briefing)
- First park: Texas, up to $4B, land secured; later parks $2B–$4B each
- Aim: smaller supply-chain firms that lack financing and regulatory capacity to expand to the US on their own
- Context: Wistron opened a $700M Texas plant producing Nvidia AI systems in July 2026; Taiwan's credit-guarantee pledge of up to $250B under the US–Taiwan trade agreement
- No construction start date or tenant list given

#### What happened
Taiwan's government will build and finance industrial parks in the US so that its smaller component makers can follow Nvidia, TSMC
and Foxconn to American AI-server production. Texas comes first. This entry was written from Crypto Briefing and the Techmeme
headline, because Bloomberg's article is paywalled.

#### Why it matters
AI server demand is pulling the supply chain behind the big Taiwanese firms to the US, with state money covering the risk smaller
suppliers cannot carry. It also deepens Taiwan's economic ties with the US as AI hardware becomes a security matter.

Sources: [Bloomberg: Taiwan to fund $4 billion Texas complex for AI, Nvidia partners (paywalled; not read)](https://www.bloomberg.com/news/articles/2026-10-16/taiwan-to-fund-4-billion-texas-complex-for-ai-nvidia-partners); [Crypto Briefing: Taiwan plans up to $4 billion Texas industrial park for Nvidia suppliers](https://cryptobriefing.com/taiwan-texas-industrial-parks-nvidia-suppliers/)

### 2026-10-16: Anthropic signs its first Canadian data-centre lease: EverWind's ~$10B wind-powered campus at Point Tupper, Nova Scotia

***; Anthropic, EverWind Fuels; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-16-anthropic-everwind-nova-scotia-data-centre/

On Oct 16, 2026 Halifax-based EverWind Fuels named Anthropic as the tenant of a data-centre campus at Point Tupper, Nova Scotia. Anthropic signed a long-term lease for the first phase (95 MW, targeted for late 2027); a second phase of 325 MW is planned for 2028. With about $2B of new wind farms, the project is put at about $10B, which EverWind calls the largest private investment in the province's history. Reuters described it as Anthropic's first Canadian data-centre lease.

- Developer: EverWind Fuels (Halifax; CEO Trent Vichie), with Indigenous equity partners Membertou First Nation and Mi'kmaq Nation communities
- Site: Point Tupper, Nova Scotia, about 200 km east of Halifax
- Capacity: phase 1 of 95 MW, start-up targeted for late 2027; phase 2 of 325 MW targeted for 2028, subject to commercial, technical and regulatory conditions (420 MW in total)
- Cost (The Globe and Mail): about $8B for the campus and about $2B for a provincial wind portfolio (Bear Lake, Windy Ridge, Kmtnuk, Upper Afton wind farms), about $10B in all; each phase financed separately with project debt
- Power and cooling: wind power, battery storage and closed-loop water cooling; EverWind and Anthropic say they will pay all grid connection and upgrade costs so existing customers' rates do not rise; grid impact studies are still under way
- Jobs: about 2,000 construction jobs and 215 permanent positions (EverWind)
- Anthropic's Jacob Glick (head of public policy in Canada) spoke of 'bringing new clean energy onto the grid'
- EverWind had earlier described the tenant only as a 'leading frontier AI developer'; Anthropic had posted jobs for a Canadian compute lead and an Alberta community-engagement manager in August 2026 (BetaKit, The Logic)

#### What happened
EverWind Fuels, a Nova Scotia developer known for green-hydrogen and wind projects, revealed on Friday, Oct 16, 2026 that Anthropic is
the anchor tenant of its planned data-centre campus at Point Tupper on Cape Breton's Strait of Canso. Anthropic signed a long-term lease
for the first 95 MW phase. EverWind plans to have the data centre running before four linked wind farms are complete and will finance
each phase separately.

This entry is written from The Globe and Mail's report. Reuters' and CBC's articles were not accessible to us; their headlines say the
same (first Canadian lease; the centre will power Claude).

#### Why it matters
It is Anthropic's first data-centre lease in Canada (until now it used Google capacity in Ontario and Quebec and Fluidstack/Iren sites in
British Columbia, per The Logic). The pledge to pay for grid upgrades so ratepayers are not charged reflects the political pressure on
data centres' power costs in 2026.

Sources: [The Globe and Mail: Anthropic signs onto $10-billion Nova Scotia data centre project (Oct 16)](https://www.theglobeandmail.com/business/article-anthropic-signs-onto-10-billion-nova-scotia-data-centre-project/); [Reuters: Anthropic signs lease for first Canadian data center (Oct 16; not accessible to us)](https://www.reuters.com/legal/litigation/anthropic-signs-lease-first-canadian-data-center-2026-10-16/); [CBC: Nova Scotia to host massive data centre to power Anthropic's AI assistant Claude (Oct 16; blocked for us)](https://www.cbc.ca/news/canada/nova-scotia/nova-scotia-to-host-massive-data-centre-to-power-anthropic-s-ai-assistant-claude-9.7383129); [BetaKit: Anthropic job posting suggests it's eyeing Alberta for data centre buildout](https://betakit.com/anthropic-job-posting-suggests-its-eyeing-alberta-for-data-centre-buildout/)

### 2026-10-19: Amazon launches TorchNeuron, an open-source (Apache 2.0) layer that runs unmodified PyTorch code on Trainium chips

***; Amazon, AWS; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-19-amazon-torchneuron-pytorch-trainium/

On Oct 19, 2026 Amazon announced TorchNeuron, which lets developers run existing PyTorch code natively on AWS Trainium2 and Trainium3 "with a single configuration change and no rewrite", released as open source (Apache 2.0) at the PyTorch Conference North America on Oct 20. It targets Nvidia's biggest moat, its CUDA software ecosystem; TechCrunch called it Amazon's "biggest shot at Nvidia yet".

- What it does: native PyTorch execution on Trainium; Amazon says the same training script ran on a conventional chip and on Trainium with only the device setting changed, and the model learned the same things at the same rate
- Chips: Trainium2 and Trainium3; licence Apache 2.0; code at github.com/aws-neuron/torch-neuronx; public beta (Techmeme/GeekWire)
- Amazon on X: '85% of AI researchers and developers write their code in PyTorch'
- Early users named by Amazon: Cartesia (voice AI inference pipeline with minimal changes) and the Lieber Institute for Brain Development (drug discovery, molecular dynamics); Hugging Face enabled native Trainium support in its tools

#### What happened
Amazon presents TorchNeuron as making Trainium behave like any other PyTorch device: per its X post, "the tools developers
already use to test, debug, and scale their models all work on Trainium without modification".

#### Why it matters
GeekWire frames the launch as a challenge to Nvidia's software advantage (CUDA). Amazon, Huawei (with DeepSeek's TileLang tools) and others are each
trying to make their chips a drop-in target for the code labs already write.

Sources: [About Amazon: Amazon opens its AWS Trainium AI chips to every PyTorch developer, no rewrite needed](https://www.aboutamazon.com/news/innovation-at-amazon/torchneuron-pytorch-aws-trainium-chips-ai); [TechCrunch: Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool](https://techcrunch.com/2026/10/19/amazon-is-taking-its-biggest-shot-at-nvidia-yet-with-new-torchneuron-dev-tool/); [GeekWire: Amazon challenges Nvidia's software advantage with new tool for its AI chips](https://www.geekwire.com/2026/amazon-challenges-nvidias-software-advantage-with-new-tool-for-its-ai-chips/); [Amazon on X: Introducing TorchNeuron](https://x.com/amazon/status/2112182407256277095); [GitHub: aws-neuron/torch-neuronx](https://github.com/aws-neuron/torch-neuronx)

### 2026-10-19: Meta details Prometheus, the Ohio gigawatt supercluster that trained Muse Wonder: aiming at ~1M GPUs, 85%+ training goodput, 99% fewer interruptions

***; Meta, Nvidia; Chips & compute; confidence high; https://postcutoff.com/e/2026-10-19-meta-prometheus-supercluster-details/

On Oct 19, 2026 Meta Engineering published a technical account of Prometheus in New Albany, Ohio, the supercluster whose first completed phase trained the Muse models including Muse Wonder. Meta says it is built toward roughly 1 million GPUs working together, across GB200 and GB300 NVL72 racks with Vera Rubin NVL72 next, linked by an Ethernet fabric spanning buildings up to 100+ km apart.

- Scale target: ~1M GPUs working together; built in phases; two years earlier the site was a conventional web-serving data center; part of the capacity is leased from partners; 'rapid deployment structures' (industrial-grade buildings) used alongside conventional halls
- Racks: Catalina (GB200 NVL72, ~120 kW per two-rack pair), Clemente (GB300 NVL72, 288 GB HBM3e per GPU, ~135 kW per rack, 5–10% faster than GB200) and upcoming San Miguel (Vera Rubin NVL72, 288 GB HBM4 per GPU, 2.52 EFLOPS FP4 per rack, 170+ kW per rack)
- Network: Ethernet back-end, up to 20,736 GPUs per building fabric; Backend Aggregation links building fabrics 2 km to 100+ km apart (Arista 7800 / Broadcom Jericho3); 800G-ZR coherent optics beyond ~10 km; 600,000+ NVLink connections
- Reliability: training goodput above 85% for Muse Wonder pre-training and above 90% for RL runs; mean time between interruptions ~47K GPU-hours, raised to 54K+ with firmware fixes; interruptions peaked at nearly 1,000 per 8K GPUs per day in mid-2025, now down 99%; target ~2 per 8K GPUs per day
- Spares: each NVL72 rack runs a 64-GPU slice with 8 in-rack spares, plus ~2% spare racks; a 32,768-GPU run needs 523 racks (37,656 GPUs)
- Power: each GPU draws ~1,400 W under load; synchronized training swings load 30–40%, starting or pausing a job can move tens of megawatts in seconds; Meta smooths this with 'Dynamo Dimmer' power capping, rack batteries, capacitor trays and facility battery storage
- Next: Hyperion in Richland Parish, Louisiana, with closed-loop liquid cooling and a grid connection through Entergy that Meta says it pays for in full
- Authors: Kevin Lee, Yashar Bayani, Ahmed Sharif, Yee Jiun Song

#### What happened
Meta published the most detailed public description yet of a frontier training cluster: rack generations, network topology across
distant buildings, failure rates and power handling. The post confirms that Prometheus's first phase is complete and trained the Muse
family, and gives Meta's own goodput and reliability figures for the Muse Wonder run.

#### Why it matters
Frontier labs rarely publish failure and goodput numbers for their largest runs. The figures show what a gigawatt-class cluster takes in
practice: hundreds of failures a day at 2025 peaks, spare GPUs built into every rack, and power swings of tens of megawatts that the grid
has to absorb.

Sources: [Meta Engineering: Inside Prometheus, Meta's gigawatt-scale supercluster (Oct 19)](https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/)

## Videos

- 2026-10-15, Linus Tech Tips: [My First 15 Minutes with RTX Spark](https://www.youtube.com/watch?v=sjuisARFWn8) https://postcutoff.com/v/ltt-my-first-15-minutes-with-rtx-spark/
- 2026-10-15, Micro Center: [RTX Spark Benchmarks: AI, Gaming & More at Micro Center](https://www.youtube.com/watch?v=Rqtdx4V0TkQ) https://postcutoff.com/v/micro-center-rtx-spark-benchmarks-ai-gaming/
- 2026-10-15, NotebookcheckReviews: [Nvidia's RTX Spark Is Here: The Future of Windows Laptops?](https://www.youtube.com/watch?v=9tq71nn0u0c) https://postcutoff.com/v/notebookcheck-nvidia-rtx-spark-future-of-windows-laptops/
- 2026-10-15, Prodigeek: [NVIDIA RTX SPARK (N1X) | LA RECENSIONE DEFINITIVA](https://www.youtube.com/watch?v=eh5kWyNAaVI) https://postcutoff.com/v/prodigeek-rtx-spark-n1x-definitive-review/
- 2026-10-15, The Tech Chap: [Can RTX Spark Replace My MacBook Pro?](https://www.youtube.com/watch?v=MaumIMgtwGw) https://postcutoff.com/v/the-tech-chap-rtx-spark-replace-macbook-pro/
- 2026-10-14, LMGClips: [Who is a $6000 AI Laptop For?](https://www.youtube.com/watch?v=uiaEGWafEZU) https://postcutoff.com/v/lmgclips-who-is-a-6000-ai-laptop-for/
- 2026-10-14, NVIDIA: [The Road to NVIDIA RTX Spark](https://www.youtube.com/watch?v=vJvFta7e76E) https://postcutoff.com/v/nvidia-road-to-rtx-spark/
- 2026-10-12, Bloomberg Tech: [Microsoft Ramps Up Homegrown AI Chip Push](https://www.youtube.com/watch?v=cOO44Fn6Taw) https://postcutoff.com/v/bloomberg-tech-microsoft-homegrown-ai-chip-push/
- 2026-10-12, Bloomberg Tech: [Microsoft, Musk Race to Reinvent the AI Chip](https://www.youtube.com/watch?v=ri8ko_PTEgw) https://postcutoff.com/v/bloomberg-tech-microsoft-musk-race-to-reinvent-ai-chip/
- 2026-10-12, Deirdre Bosa: [Microsoft Unveils Maia 300 | Cloud Chief Scott Guthrie Live](https://www.youtube.com/watch?v=4dVWeo6VKyI) https://postcutoff.com/v/deirdre-bosa-scott-guthrie-maia-300-live/
- 2026-10-12, Yahoo Finance: [Microsoft unveils its Maia 300 AI chip to cut its reliance on Nvidia | The 8:30](https://www.youtube.com/watch?v=8W2agm_nmjo) https://postcutoff.com/v/yahoo-finance-microsoft-maia-300-ai-chip/
- 2026-10-11, Bloomberg Television: [AI's Next Big Risk Is Hiding in the Debt Market](https://www.youtube.com/watch?v=P6dtMlkY9ZY) https://postcutoff.com/v/bloomberg-tv-ai-next-big-risk-debt-market/
- 2026-10-11, Jake Broe: [Russian Internet CUT OFF after Yandex Drone Attacks](https://www.youtube.com/watch?v=sNLbvLtEUro) https://postcutoff.com/v/jake-broe-russian-internet-yandex-drone-attacks/
- 2026-10-11, Tom McCarty: [2 Strikes in 48 HOURS... Ukraine Is Wiping Out Russia's 'Google' Data Centers](https://www.youtube.com/watch?v=OvwOXdbhtbQ) https://postcutoff.com/v/tom-mccarty-two-strikes-48-hours-yandex/
- 2026-10-10, Fireside Alpha (@firesidealpha): [Qualcomm CEO: AI companies want phones running 100-billion-parameter models by 2028 (X clip)](https://x.com/firesidealpha/status/2108930562794864912) https://postcutoff.com/v/fireside-alpha-qualcomm-amon-100b-parameter-phones-x/
- 2026-10-09, Matt Talks Tech: [NVIDIA SPARK Surface Laptop Ultra VS MacBook Pro M5 Pro: Which One Will WIN?](https://www.youtube.com/watch?v=DMtkOXdAHsM) https://postcutoff.com/v/matt-talks-tech-rtx-spark-surface-laptop-ultra-vs-macbook/
- 2026-10-09, RFU News — Strategic Geopolitics: [Wow! 40% of Russian Yandex data centers destroyed in 2 days!](https://www.youtube.com/watch?v=kkbZskBCRaI) https://postcutoff.com/v/rfu-news-40-percent-yandex-data-centers/
- 2026-10-09, Windows: [Satya Nadella and Jensen Huang discuss Hybrid Intelligence and the future of PC](https://www.youtube.com/watch?v=lFyNu4eBrBE) https://postcutoff.com/v/windows-nadella-huang-hybrid-intelligence/
- 2026-10-08, Kanal13: [Ukrainian drones strike Yandex’s largest data center in Russia](https://www.youtube.com/watch?v=knQcbbX_lHs) https://postcutoff.com/v/kanal13-drones-strike-yandex-largest-data-center/
- 2026-10-08, New York Post: [Smoke Rises From Russian Yandex Data Hub Following Ukrainian Drone Strike](https://www.youtube.com/watch?v=6qo4E5Y1iuU) https://postcutoff.com/v/ny-post-smoke-yandex-data-hub-drone-strike/
- 2026-10-08, RFU News — Strategic Geopolitics: [Ukraine destroys a massive Yandex data center and cripples Russian tech](https://www.youtube.com/watch?v=2cECsed-478) https://postcutoff.com/v/rfu-news-ukraine-destroys-yandex-data-center/
- 2026-10-08, Silicon Curtain: [Ukraine Strikes Yandex in Huge Blow to Putin - Russia Back to Digital Dark Ages!?](https://www.youtube.com/watch?v=D3spMc7HfbY) https://postcutoff.com/v/silicon-curtain-ukraine-strikes-yandex/
- 2026-10-07, CNET: [Windows, Surface and Nvidia Event Highlights: Everything Announced in 16 Minutes](https://www.youtube.com/watch?v=K6QJp1edjmU) https://postcutoff.com/v/cnet-windows-surface-nvidia-event-highlights/
- 2026-10-07, Will Phillips: [24h Inside the $11M Startup Building a Personal AI Computer](https://www.youtube.com/watch?v=GqSXkWbYVZo) https://postcutoff.com/v/will-phillips-24h-inside-ghost-personal-ai-computer/
- 2026-10-07, Windows: [Something new is coming from Windows | October 7](https://www.youtube.com/watch?v=ilmBGeGldrI) https://postcutoff.com/v/windows-something-new-october-7-event/
- 2026-10-06, NVIDIA: [NVIDIA RTX Spark | Made to Create](https://www.youtube.com/watch?v=4XZzjY4X13A) https://postcutoff.com/v/nvidia-rtx-spark-made-to-create/
- 2026-09-30, Google: [Making sure AI chips don’t overheat in space](https://www.youtube.com/watch?v=ktdbUIZKeSE) https://postcutoff.com/v/google-suncatcher-chips-overheat-space/
- 2026-09-28, Google: [Testing AI chips to survive in space](https://www.youtube.com/watch?v=8NPgswgbTnE) https://postcutoff.com/v/google-suncatcher-testing-chips-space/
- 2026-09-24, Android Authority: [Snapdragon 8 Elite Gen 6 benchmarks: Unbelievable](https://www.youtube.com/watch?v=5blJhnOPLoo) https://postcutoff.com/v/android-authority-snapdragon-8-elite-gen-6-benchmarks/
- 2026-09-24, Google: [Google’s latest moonshot to put machine learning in space](https://www.youtube.com/watch?v=o1JK79jszqo) https://postcutoff.com/v/google-suncatcher-moonshot-2026/
- 2026-09-22, CNET: [Qualcomm Snapdragon Summit 2026: Everything Announced in 15 Minutes](https://www.youtube.com/watch?v=RCTYyo82UAA) https://postcutoff.com/v/cnet-snapdragon-summit-2026-everything-in-15-minutes/
- 2026-06-01, Linus Tech Tips: [NVIDIA Just Slapped Apple Silicon - RTX Spark](https://www.youtube.com/watch?v=OXneUrL3Fe0) https://postcutoff.com/v/ltt-nvidia-just-slapped-apple-silicon-rtx-spark/
- 2026-05-31, Dave2D: [Nvidia's New Laptops - RTX Spark](https://www.youtube.com/watch?v=0-VG9QBm8S8) https://postcutoff.com/v/dave2d-nvidias-new-laptops-rtx-spark/
- 2026-05-31, NVIDIA: [NVIDIA RTX Spark Reinvents Windows PCs for the Age of Personal AI](https://www.youtube.com/watch?v=H4nJo-oqAro) https://postcutoff.com/v/nvidia-rtx-spark-reinvents-windows-pcs-personal-ai/

## Posts

- 2026-10-19, Amazon: [](https://x.com/amazon/status/2112182407256277095)
- 2026-10-17, JustDario: [JustDario: why does Anthropic claim to be worth $2 trillion and behave like a company 'on its way to be worth close to zero'?](https://x.com/DarioCpx/status/2111259736528326880)
- 2026-10-17, Ed Zitron: [Ed Zitron: Anthropic paying 10% on billions of debt on top of $4.4B at 8.5% is 'an insane deal'](https://x.com/edzitron/status/2111262199872389375)
- 2026-10-15, Joanna Stern: [](https://x.com/joannastern/status/2110729668907376698)
- 2026-10-12, Satya Nadella: [Satya Nadella: 'our biggest bet yet on our own silicon' (Maia 300, multi-gigawatt Maia deployments)](https://x.com/satyanadella/status/2109631106404487523)
- 2026-10-12, Chamath Palihapitiya: [Chamath Palihapitiya on Maia 300: 'everyone competes with everyone'](https://x.com/chamath/status/2109658962807386330)
- 2026-10-12, Elon Musk: [Elon Musk replying to Chamath on AI constraints: power, construction, cooling, networking, batteries](https://x.com/elonmusk/status/2109668152754020609)
- 2026-10-12, Microsoft: [Microsoft: 'Introducing Maia 300, Microsoft's next-generation accelerator'](https://x.com/Microsoft/status/2109634232163471782)
- 2026-10-12, aaron holmes: [Aaron Holmes (The Information): Maia 300 announcement confirms his August report](https://x.com/aaronpholmes/status/2109639265504358491)
- 2026-10-12, Andreas Schilling 🇺🇦: [Andreas Schilling: Maia 300 chiplet configuration](https://x.com/aschilling/status/2109633438672458119)
- 2026-10-12, Ben Bajarin: [Ben Bajarin: Maia 300 rack and networking details](https://x.com/BenBajarin/status/2109655386727108698)
- 2026-10-12, Brendan Burke: [Brendan Burke (Futurum): Maia's gigawatt plan is a bet on software-configured dataflow and agent-written kernels](https://x.com/BrendanBurkeX/status/2109657608995188890)
- 2026-10-12, Deirdre Bosa: [Deirdre Bosa quoting Scott Guthrie: custom chips narrow the 'cone of uncertainty'](https://x.com/dee_bosa/status/2109783379625513328)
- 2026-10-12, Ed Ludlow: [Ed Ludlow (Bloomberg): Microsoft has not named the GPU baseline for Maia 300's 40% claim](https://x.com/EdLudlow/status/2109648408973807874)
- 2026-10-12, Ed Zitron: [Ed Zitron: 'multi gigawatt deployments' wording questioned; 'Microsoft only has 2gw of ai data centers'](https://x.com/edzitron/status/2109674961569009915)
- 2026-10-12, Mustafa Suleyman: [Mustafa Suleyman: 'Our new silicon is 🔥 40% more tokens per watt!'](https://x.com/mustafasuleyman/status/2109655335250366752)
- 2026-10-12, Patrick Moorhead: [Patrick Moorhead on Maia 300: 'primarily a margin lever, not a benchmark story'](https://x.com/PatrickMoorhead/status/2109635430769475668)
- 2026-10-12, Torsten Hoefler 🇨🇭: [Torsten Hoefler on Maia 300: 'Our SDLA is a great architecture for LLMs'](https://x.com/thoefler/status/2109657493341372497)
- 2026-10-11, Elon Musk: [Elon Musk on 100B-parameter phone models: 'Problem is RAM supply'](https://x.com/elonmusk/status/2109260288314806434)
- 2026-10-11, NetBlocks: [](https://x.com/netblocks/status/2109151636098818212)
- 2026-10-11, Slava 🇺🇦: [Viral claim: War Thunder down worldwide and Yandex not restored until 2028 (Community Note: false / speculative)](https://x.com/Heroiam_Slava/status/2109214421218144373)
- 2026-10-11, medoyid_ua: ['A third Yandex data center is now down': list of 80+ Yandex Cloud services lost after the Vladimir strike](https://x.com/letsarmukr/status/2109094114675261582)
- 2026-10-11, OSINTtechnical: [OSINTtechnical: Yandex apps down after the Vladimir strike, 'the third Yandex facility hit in four days'](https://x.com/Osinttechnical/status/2109160348133609590)
- 2026-10-10, Fireside Alpha: [Fireside Alpha clip: Amon says AI companies want 100B-parameter models running on phones by 2028 and 'are starting to do phones'](https://x.com/firesidealpha/status/2108930562794864912)
- 2026-10-10, Zephyr: [Zephyr: a 100B-parameter phone model likely means ~25 GB after quantization, late 2028 or early 2029](https://x.com/zephyr_z9/status/2108959348408143990)
- 2026-10-09, Rahul Chhabra: [Sabi co-founder Rahul Chhabra: $50M raised to build 'the world's most wearable BCI cap'](https://x.com/rahulchhabra07/status/2108558367258186220)
- 2026-10-08, Alex Heath: [Alex Heath: Sources interview with Qualcomm CEO Cristiano Amon on the 'AI phone supercycle'](https://x.com/alexeheath/status/2108275202174157044)
- 2026-10-08, Cloudflare Radar: [Cloudflare Radar: Yandex Cloud traffic about 33% below the previous week after the Sasovo strike](https://x.com/CloudflareRadar/status/2108122893028671604)
- 2026-10-08, Jeremiah Fowler Security Researcher: [Jeremiah Fowler: 'russian bot traffic has dropped to 49%' after the Yandex data-center strike](https://x.com/yoda69/status/2108280211863421355)
- 2026-10-07, Elon Musk: [Elon Musk: 'We will build and run the fab'; TSMC may at most sublease part of Terafab](https://x.com/elonmusk/status/2107721071571128709)
- 2026-10-05, Zain Javaid: [Ghost introduces Core, 'The Personal AI Computer'](https://x.com/zainmfj/status/2107146556176798081)
- 2026-10-03, Elon Musk: [Elon Musk confirms Terafab–TSMC talks: 'Just discussions, but something may come of it'](https://x.com/elonmusk/status/2106290920605790717)
- 2026-10-01, Rohan Paul: [](https://x.com/rohanpaul_ai/status/2105504789526409498)
- 2026-10-01, The Information: [](https://x.com/theinformation/status/2105693008985284875)
- 2026-09-24, Oracle: [Oracle: 'Project Jupiter remains on our planned schedule'](https://x.com/Oracle/status/2103105677874897334)
- 2026-08-25, tae kim: [](https://x.com/firstadopter/status/2092266927377039838)