Odyssey, the Palo Alto AI lab building world models that simulate physics in 3D, announced a $310 million Series B on June 17 at a $1.45 billion valuation. Natural Capital led the round, with Amazon, AMD Ventures, Alphabet's GV, EQT, and CIA-linked fund In-Q-Tel taking part.
The investor who isn't there
Plenty of breathless coverage credited Amazon, Nvidia, and AMD with backing this round. Nvidia didn't back this round. Its venture arm, NVentures, put money into Odyssey's Series A back in February, and then went missing from the Series B entirely. What replaced it is the interesting part: AMD Ventures, Nvidia's biggest chip rival, plus a deal naming AWS as Odyssey's preferred cloud provider and Amazon's Trainium chips as the silicon of choice.
Read the round as a hardware story and it reads like a defection. Whether that reflects real conviction in Amazon's silicon or just better terms in a crowded funding market, nobody outside the cap table can say. Amazon's Annapurna Labs will work with Odyssey to optimize the models for Trainium, which is Amazon's bid to pull heavy AI workloads off Nvidia's GPUs.
What Odyssey actually builds
Founders Oliver Cameron and Jeff Hawke came out of self-driving. Cameron ran autonomous startup Voyage until GM's Cruise bought it; Hawke engineered at UK self-driving outfit Wayve. The pitch is that world models are the next foundation-model category after text, gathering data from the physical world and replaying it with accurate physics, body language, and object dynamics. Things, Cameron argues, that language models simply can't capture.
The data collection is charmingly low-tech. Per TechCrunch, Odyssey strapped cameras to people's backs and sent them walking, a deliberate echo of how Google Earth was built, except Google used cars. It fits a 55-person team spread across London, Zurich, and Palo Alto, with a research bench poached from DeepMind, Tesla, Waymo, and Wayve.
The GPT-3 moment claim
"This round provides the compute, infrastructure, and partners to push the frontier of general world models, and to achieve a GPT-3 moment for the field."
That's Cameron, and it's a big swing for a two-year-old company that hasn't shipped a benchmark anyone outside the lab can independently check. Odyssey points to a string of internal milestones, including Odyssey-2 Max for physics accuracy and Agora-1 for multi-agent simulation, but the latency and multi-agent claims remain self-reported. The angel list is genuinely heavy though: Google chief scientist Jeff Dean, Elad Gil, Y Combinator's Garry Tan, Vercel's Guillermo Rauch, and Cruise founder Kyle Vogt.
One detail worth chewing on: In-Q-Tel's presence, and Odyssey's own mention of defense as a target market. World models that can simulate physical and social environments at scale are exactly the kind of thing intelligence agencies fund early.
Odyssey has now raised about $337 million total. The next real test is whether it publishes anything reproducible, because right now the round is the news and the technology is still a promise.
