Medal raises $133.7M to build general AI agents using gaming data for spatial-temporal reasoning

Summary

Medal has closed a $133.7 million seed round to develop general AI agents built on gaming data, targeting environments that require spatial-temporal reasoning. The round size, down to the $0.7 million, is a deliberate nod to the gaming reference 1337 ("leet"), a signal of the company's culture.

Pim (Medal's founder) built his earliest commercial instincts on gaming infrastructure, generating roughly $1.5 million in revenue running the largest private server on RuneScape before age 18. He subsequently spent three years working with a doctor, an experience that shaped the company's early focus on search and rescue drones as a flagship real-world application.

The Core Technical Bet

Medal's thesis is that gaming environments, particularly as physics engines grow more realistic, offer sufficient behavioral diversity to train foundation models that can transfer to novel physical environments with minimal additional data. The key insight is that drones, robotic arms, and other hardware already ship with game controllers as standard interfaces, meaning Medal's models don't need to learn new action spaces when deployed in the real world.

The current priority is general navigation for novel environments rather than broader task completion. Pim frames the near-term problem bluntly: preventing agents from running into things. More complex autonomous behavior, including search and rescue without human oversight, is the longer-horizon target.

Simulation Strategy and Deployment Flywheel

Medal is actively partnering with game developers to deploy its agents inside commercial video games, treating in-game performance against human players as a verifiable benchmark before transferring capabilities to physical hardware. The company holds an internal view that simulated realities may ultimately represent a larger total addressable market than the physical world, given there is only one physical reality but an unbounded number of simulated ones.

On the question of simulation architecture, Pim argues for a hybrid approach. High-fidelity game engines like Unreal remain essential for staying in a verifiable, deterministic domain, while generative world models like Genie are better suited for phenomena that lack sufficient real-world video footage, such as cellular or microscale environments. Generative world models, despite being a genuine technical advance, introduce non-determinism that creates safety and reliability risks in physical deployments, making full reliance on them premature.