As of: 2026-10-13 07:38 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/p/2026-10-12-cheeky-pint-zuckerberg-wang-muse-wonder/

# Cheeky Pint: 'Muse Wonder: Zuckerberg and Wang's $100B AI' (interview transcript)

John Collison (Cheeky Pint), with Mark Zuckerberg and Alexandr Wang (@collision), Substack, 2026-10-12. Importance: major (4 of 5).
Source: https://cheekypint.substack.com/p/muse-wonder-zuckerberg-and-wangs

## Why it matters

The only long first-hand account of how Meta built Muse Wonder: codenames, team size, training cost, the $100–200B compute bet made before the models existed, and what Meta plans next.

## Summary

Published Oct 12, 2026, the day Muse Wonder launched; just over two hours (the last chapter starts at 02:08:11). Chapters: Muse's breakout use cases; Why the model is the product; Meet Muse Wonder; How frontier models are actually built; Betting $100B; Agentic commerce; OpenRouter; Beyond the smartphone; Funding; Why AI won't be winner-take-all; Meta Superintelligence Labs (MSL); Working for Mark Zuckerberg. The video is in the dataset as `cheeky-pint-zuckerberg-wang-muse-wonder`.

Main first-hand claims (short quotes from the published transcript, which is machine-made in places):
- Codenames: "Avocado was the code name for the Muse Spark line, and then Watermelon was the code name for this one. And we're out of fruits now." (Zuckerberg)
- Timeline: "we've planned Muse Wonder since last August" (Wang); Wonder is "the next pre-train", "significantly bigger" than Spark; pre-training "has been done for a bit" (Zuckerberg).
- Team: "The research team is less than 200 people ... It's closer to 100, right? And then there's a bunch of infra folks." (Zuckerberg)
- Cost: "to operate a gigawatt a year, call it on the order of $15 billion"; building it "$50 or $60 billion" (Zuckerberg). Wang: "tens of billions of dollars is a good estimate" for training a frontier model.
- Compute: Prometheus in Ohio is "scaling to larger than a gigawatt"; Hyperion in Louisiana "is going to scale to 5 gigawatts".
- The bet: at the end of 2025, "we haven't seen the models yet ... we're going to write this check for, like, I don't know, like $100 billion, $200 billion. I'm like, Alex, are you sure this is going to work?" Wang says "Susan Li, our CFO" kept pulling him aside: if it failed, "it's going to be very, very bad in financial terms". Zuckerberg: "in order for it not to work out, it's not just that we have to fail, AI overall has to fail."
- Availability: Wonder "will also be available via API, on the Meta Model API and OpenRouter ... And then Muse Wonder Full will be available to subscribers of Muse Code" (Wang); in Muse, "we're going to start off primarily in paid subscriptions" (Zuckerberg).
- Next: "at some point not too far from now, have a model that is as good as Wonder, but in the Spark form factor"; "the time between each subsequently bigger class of models, I think will now start to compress"; a version number comes "when we have our first dot release".
- Safety: Wang says RL environments were curated to drive "more aligned behavior"; Zuckerberg prefers "a diversity and ... checks and balances" over "a couple of good models that are only in a couple companies' hands". Wang: "distillation is quite a good force for the entire ecosystem".
- Agentic commerce: Wang calls Muse "one of the first products where ... people are fully delegating purchases to the agents".
- No user numbers, prices or dates for the next model are given.

## Archived text

Summary plus short quotes only; full transcript at the URL

## Cited in

- 2026-10-12: [Meta releases Muse Wonder and Muse Wonder Full, its first frontier-class models](https://postcutoff.com/e/2026-10-12-meta-muse-wonder/)

## People in this post

- [Mark Zuckerberg](https://postcutoff.com/person/mark-zuckerberg/), Chairman and CEO, Meta