‹ rbtfl.

General Intuition raises $320M to train AI agents on video-game footage

Khosla-led Series A at a $2.3B valuation bets that gameplay clips teach machines to act in the physical world

스타트업·AI· active 누구의 돈인가·장기전 ·8 시각 · ·rbtfl 업데이트 2026년 6월 28일

Summary

General Intuition raised a $320M Series A at a $2.3B valuation on 25 June 2026, led by Khosla Ventures with General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg and researchers from Google Deepmind and MIT participating. The startup trains large action foundation models on action-labelled gameplay clips, records of exactly which buttons a player pressed and when, drawn from Medal's 17 million monthly users. CEO Pim de Witte says the same model that plays Fortnite can drive a quadruped robot after eight minutes of real-world fine-tuning. Most of the cash funds compute, via a Coreweave deal, to pre-train the next model and to build world models that generate training environments. Total disclosed funding reaches $454M after a $134M launch round in October 2025.

The split

US coverage (TechCrunch, PitchBook) reads this as the embodied-AI and world-model thesis pulling the capital that LLM labs commanded, gameplay as a cheaper route to robotics than teleoperation. The gaming press (GamesBeat) foregrounds the data moat in Medal's user base. A European desk (TechFundingNews) frames it through Khosla's conviction, two cheques in three months, signalling that world models, not chatbots, are the next venture frontier.

By the numbers

  • $320M, Series A round size.
  • $2.3B, post-money valuation.
  • $454M, total disclosed funding to date.
  • 17M, monthly active users on Medal supplying gameplay data.
  • 8 minutes, real-world data to fine-tune the model onto a quadruped robot.
  • $134M, October 2025 launch round.

Why it matters

The round is a marker that world models and embodied agents, not text LLMs, are where frontier venture money is rotating. If gameplay footage really transfers to robots cheaply, it undercuts the teleoperation and simulation pipelines that robotics labs rely on, and turns a consumer gaming platform's exhaust into a defensible training-data asset.

What to watch

  • A public model or API release by late summer, the stated CoreWeave-funded milestone.
  • Independent demos of game-to-robot transfer beyond the eight-minute quadruped claim.
  • Whether other gameplay platforms (Twitch, Discord, Roblox) monetise action-labelled data the same way.