Interview

Moonlake AI raises $28M seed to build interactive AI-generated worlds and games

Oct 1, 2025 with Fan-Yun Sun

Key Points

  • Moonlake AI, a Stanford spinout, raises $28 million seed from Nvidia Ventures, AX Ventures, and over ten unicorn founders to build AI-generated interactive worlds where player input shapes environments in real time.
  • The company combines code generation and diffusion models to reduce inference costs versus pure video generation, positioning itself as an application layer atop existing graphics tools rather than a foundation model lab.
  • Early studio interest includes an F1 IP holder seeking real-time ghost-racing experiences, signaling demand for interactive worlds that would historically be cut for cost and complexity.
Moonlake AI raises $28M seed to build interactive AI-generated worlds and games

Summary

Moonlake AI, a Stanford AI lab spinout founded roughly six months ago, has raised a $28 million seed round led by Ventures, AX Ventures, and Nvidia Ventures, alongside more than ten unicorn founders and prominent AI researchers. The company's founder and CEO, Fannon Sun, is building a platform for AI-generated interactive worlds — games and simulations where user actions shape the environment in real time, not passive video or image output.

The distinction matters commercially. Most world-model companies generate convincing visuals but can't hold state or respond to player input. Sun argues the current moment in generative worlds is roughly where ChatGPT was at GPT-3 — capable but not yet given the tool-use layer that unlocks real utility. Moonlake's technical answer is a hybrid representation that combines code generation and diffusion models, letting reasoning models choose how to represent a given environment. Sun says that architecture also reduces inference cost meaningfully compared to pure video generation approaches, since rendering state changes in code is cheaper than generating each frame from scratch.

GTA cost $2 billion to develop. That number is Sun's shorthand for the problem Moonlake is trying to solve: world-building has always required enormous resources and deep domain expertise. The pitch is that reasoning models, applied to existing graphics tooling, can make interactive world creation accessible to anyone — vibe coding for games rather than vibe coding for software.

Moonlake is positioning itself as an application-layer company, not a foundation model lab. Sun draws the analogy to Cursor, which never trained a frontier model but built proprietary modules — tab completion, context retrieval — that combined into a product users love. Moonlake plans a similar approach: train what the product requires, buy or use what it doesn't.

Early studio interest is real, at least anecdotally. Sun says a team that owns the F1 IP approached Moonlake about building real-time ghost-racing experiences with F1 cars — the kind of feature that would historically get cut for cost and complexity. Roblox, a $100 billion business built on conventional graphics, is the reference point for what the market ceiling looks like if the technology works.

Sun expects consumer reception to be more positive than the backlash against AI video tools like Sora. His argument is that interactive experiences feel different from generated video because users are authoring rather than consuming — and inconsistency, the main complaint about AI video, is far less jarring when the user is actively steering the world.