As of: 2026-10-12 07:22 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/v/ryansael-opus-5-5-inside-data-center-3d-x/

# Opus 5.5 one-shots an interactive 3D data center in 1h53m for $38.99

Ryan Sael (@RyanSael), 23 September 2026, X. 79,566 views as of 12 October 2026.
Kind: Made by AI.
Watch: https://x.com/RyanSael/status/2102740041621762166

## Why it is here

Listed in the awesome-claude-5-5-videos library; ~80k views. Gives a cost breakdown (99% of input was cache hits).

## Description (written by Gemini from the video)

**Summary**  
Ryan Sael (@RyanSael) demonstrates an interactive 3D WebGL data center digital twin titled "Inside an AI Data Center," reportedly generated by Claude Opus 5.5 in a single agentic session in 1 hour 53 minutes at a cost of $38.99. The video tours the browser-based simulation, toggling between compute workloads (Idle, Inference, Training), triggering thermal overloads on individual server racks, and inspecting facility cooling and electrical substations.

**What is shown**  
- [00:00] Overview of the isometric 3D facility at idle load (0.87 MW power in, 46 °C chip temperature).
- [00:01] Switching compute load to "Inference," increasing facility power to 1.25 MW with bursting prompt workloads.
- [00:03] Switching compute load to "Training," driving power to 2.18 MW across 1,140 GPUs and igniting glowing heat convection plumes above server pods.
- [00:08] Zooming into the rack aisle and inspecting Rack 8-10 (72 GPUs drawing 139 kW at 85 °C).
- [00:10] Clicking "Overload" on Rack 8-10, pushing power to 141 kW and temperatures to 89 °C until thermal throttling is triggered.
- [00:16] Clicking "Reroute cooling" to flood emergency coolant into Rack 8-10, bringing down temperatures.
- [00:21] Camera tour of roof chiller assemblies and exhaust fans discharging 2.18 MW of heat into the atmosphere.
- [00:24] Inspection of outdoor transformers and illuminated power distribution conduits.
- [00:28] Interior view of the operations control room, featuring telemetry monitors, thermal heatmaps, and a 3D volumetric thermal plume block.
- [00:32] Zoom out to the full facility overview.

**Claims & numbers**  
- Title claims Claude Opus 5.5 generated the complete interactive 3D application "one-shot" in 1 hour 53 minutes for $38.99.
- Facility simulation claims:
  - Idle state draws 0.87 MW base power with chips at 46 °C.
  - Inference workload draws 1.25 MW (equivalent to 310 homes).
  - Training workload runs 1,140 GPUs drawing 2.18 MW (equivalent to 1,730 homes).
  - Individual rack thermal threshold reaches throttling at 89 °C / 141 kW.

**Notable quotes**  
- [00:00] UI text: *"Every watt a GPU consumes comes back out as heat. Follow the power in, and the heat out."*
- [00:03] UI text: *"Training: 1,140 GPUs flat out. 2.18 MW, as much as 1,730 homes, all of it back out as heat."*
- [00:13] UI text: *"Rack 8-10 throttling: 141kW, more than its cooling can carry. Multiprocess at 89°C and climbing. Reroute the cooling, or they will blow themselves down."*

**Assessment**  
A real screen recording demonstrating a working, interactive WebGL/Three.js browser application written via an AI coding agent. The simulation operates smoothly with responsive UI interactions, camera pathing, particle animations, and state logic, though it serves as an educational simulation rather than live telemetry from a real physical data center.

**Lyrics & themes**  
The video contains no sung vocals or spoken dialogue; audio consists of ambient electronic music and interface interaction sounds. The simulation focuses on the physics and thermodynamics of frontier AI infrastructure, emphasizing that every watt of computational electrical power must ultimately be rejected as heat.

**Lore & references**  
- **Data center energy crunch**: Reflects widespread industry discussions around the gigawatt-scale power consumption and cooling challenges of training frontier models.
- **"One-shot" agentic coding**: References the AI development practice of letting autonomous coding models write, debug, and assemble complex, multi-component 3D web applications entirely via prompts.
- **Thermal throttling & liquid cooling**: Directly models liquid-to-chip cooling, pod derating, and emergency rerouting when rack power density exceeds dissipation capacity.

**Visual style & craft**  
A stylized, clean low-poly isometric 3D architectural aesthetic built with WebGL/Three.js. Features include dynamic lighting, toggleable particle layers (airflow, thermal plumes, electricity conduits), smooth camera tweening between preset focus points, and custom overlay HUD components with charts and thermal color gradients.

_Described by gemini-3.8-flash on 2026-10-12 from the video's audio and frames._

## Made by AI

- Made with: Claude Opus 5.5
- How we know: X post (2026-09-23): 'I asked Opus 5.5 to show me what's inside the buildings it runs in /  / it came up with this after 1 hour 53 minutes in one shot, $38.99 API cost /  / https://datacenter.lab.sael.net /  / it's fully interactive 3D: overload a rack to watch its GPUs throttle, then follow t…'
- Human role: Asked Opus 5.5 to show what is inside the buildings it runs in; one shot.
- Pipeline: One prompt → Opus 5.5 builds an interactive 3D site (datacenter.lab.sael.net); the video is a screen capture
- Series: One-shot interactive 3D (LLM writes the scene code)

## Related

- 2026-09-22: [Anthropic releases Claude Opus 5.5](https://postcutoff.com/e/2026-09-22-claude-opus-5-5/)