Veeda AI

What Veeda AI is building

Veeda is building multimodal foundation world models: systems that take language, image, video and 3D together and produce a simulated environment with enough physical and spatial consistency that an embodied agent can learn inside it. The company's own framing is that it is building the Matrix for Physical AI. The premise is a constraint, not an aspiration. Robot policies improve through interaction, but the real world is a terrible place to run trial-and-error at scale — it is slow, unsafe, and, as the company puts it, robotic hardware does not scale the way compute does. You cannot buy 10,000 more robot arms the way you buy 10,000 more GPUs. So the training ground has to be synthesised, and the bottleneck moves to whether the synthesis is faithful enough that a policy learned inside it survives contact with a real actuator. The job postings show the actual shape of the work: designing and training foundation models for image, video and 3D generation; post-training alignment with RLHF, DPO and reward modelling; large-scale data curation pipelines; and evaluation methodology that goes past standard generative metrics. That last one is the hard part and the company says so — for world models, the metric that matters is not perceptual quality, it is whether the physics holds up under a policy that is actively trying to exploit it. Sanja Fidler's position is not that this is a promising direction but that it is imminent: "World models is where the next breakthrough lies, and it is around the corner." Zan Gojcic's stated conviction is that simulated reality "will become the critical infrastructure layer for all areas of robotics."

Why this matters

The credibility here is unusually concentrated. Fidler spent eight years at Nvidia after Jensen Huang personally invited her in May 2018 to start what became the Spatial Intelligence Lab in Toronto, and the lab's output is the lineage of this company: GameGAN, DriveGAN, Video LDM, Cosmos, OmniDreams. Cosmos is Nvidia's flagship world model for developers building physical AI. Veeda is the same core team choosing to rebuild that outside a chip company's roadmap. That matters because a world model built inside a GPU vendor is inevitably shaped by what sells GPUs. An independent one can optimise for whatever actually closes the sim-to-real gap, and can sell to robotics companies that would rather not route their training substrate through a hardware supplier. The round's speed is its own signal: incorporated in June, $90M+ committed by late July, announced in August. Khosla and Radical co-led, the same pair that recently co-led Jeff Dean's Discovery Loop. Both took board seats. Nobody underwrites a three-month-old company at that size on the product — they underwrite the team's ability to compress years of research into a running system, which is exactly what the Toronto lab did before.

Investors: Khosla Ventures (Sven Strohband, board), Radical Ventures (Tomi Poutanen, board)