Moonlake

Moonlake (also referred to as Moonlake AI) is a San Francisco-based artificial intelligence research lab focused on world modeling and simulation infrastructure. The company positions its work as closing the gap between AI model development and physical deployment, building models that generate real-time interactive content — including 3D and 2D simulations and games — from natural language or image input. Its stated thesis is that generative, interactive worlds will provide the environments and feedback loops that reinforcement learning and embodied AI agents need, in the same way that internet-scale text supported next-token-prediction models.

The publicly described platform exposes a REST API for generating 3D assets from text prompts and reference images (for example, an endpoint at app.moonlakeai.com/api/v1/assets, with a sample gallery of generated objects such as an office chair, a robot vacuum dock, and a stand mixer). At launch the company described a system in which users describe what they want and the platform compiles it into a working interactive scene, with instant edits and adjustable physics. Under the hood the company cites multi-modal reasoning for spatial layout, tool use over existing engines to build assets and scenes, program synthesis for logic, a simulation layer for agents, and a real-time diffusion model to reskin worlds. Stated initial use cases include user-generated interactive worlds and micro-fandom content, world creation for game developers, studios and IP holders, robotics and embodied-AI simulation for training and evaluation, 3D environments for film and animation, and immersive learning experiences for teachers.

Founding story

Moonlake AI was founded by Fan-Yun Sun and Sharon Lee, who were both working out of the Stanford AI Lab. Sun completed a computer science PhD at Stanford while working as a researcher at NVIDIA on the Learning & Perception Research Team and Metropolis Deep Learning (Omniverse), where he led efforts to generate large-scale 3D worlds for training AI agents and foundation models. Lee pursued a CS PhD focused on generative AI and computer graphics, combining diffusion models with 3D engines and developing foundation models trained to reason about space for richer interactivity and control in virtual worlds. The founders frame the company around the observation that building interactive worlds is slow and expensive — franchise games can consume large budgets and years of prototyping — leaving most creative ideas untested.

Business model

Moonlake provides its world-modeling and simulation capabilities as a developer-facing platform and API, with a documented HTTP endpoint for programmatic 3D asset generation from text and image references, alongside a demo and early-access request flow for creators, developers, studios and research partners.

Traction

As of the October 2025 launch the product was in private preview, with $28 million in seed funding disclosed. In May 2026 the company was included in CB Insights' AI 100 list of the most promising AI startups of 2026.

Latest developments

Company blog posts in 2026 covered efficient world models (March 2026), the launch of Moonlake's 3D Agent adding computer-use capabilities for world modeling (April 2026), and guidance on using Moonlake to build simulations and digital twins (May 2026). On May 5, 2026 the company was named to CB Insights' AI 100.