Odyssey is a world model lab based in Palo Alto. Its co-founder and CEO, Oliver Cameron, has spent a decade on machines that learn the physical world. He built the self-driving car course at Udacity with Sebastian Thrun, founded Voyage, a self-driving company that Cruise acquired, and then served as VP Product at Cruise. In late 2023, he started Odyssey with Jeff Hawke, a founding engineer at Wayve. This June, Odyssey raised a $310mn Series B at a $1.45bn valuation, with Amazon, AMD Ventures and GV in the round. It recently released Odyssey-3, its foundation world model.
I have been looking forward to this interview with Oliver for a while. I thought this would be a conversation about world models, yet the key point we kept coming back to was about compute scarcity. Oliver recounted a time when compute deals agreed on a handshake disappeared the next day. That caused him to run his Series B as a search for compute partners. If you lead a robotics company that trains its own models, this changes how you plan your next raise.
In this episode, we cover
(00:24) What is Odyssey-3, a foundation world model?
(02:05) Why task-specific robot data hits a ceiling
(04:50) Driving in India: one researcher, results of a 500-person team
(06:57) 20 hours of blocky simulation vs 30 hours of real driving
(09:21) Flexion, post-training and why robot companies stop building everything
(11:45) Poke & Wiggle: real-robot evals as a test suite
(15:15) World models vs language models: 14B parameters vs a 5T rumor
(18:48) What it takes to scale world models
(21:37) Egocentric video and the data mixture
(23:33) Is any robot data still hard to find?
(25:26) World models as the ultimate RL environment
(27:30) When compute became existential
(29:29) Fundraising as a compute partnership search
(31:44) A hybrid fleet: NVIDIA, Trainium and AMD
(33:13) Coding agents and CUDA lock-in
(34:18) What is a robot company’s moat?
(36:17) A short memory for pain
(37:55) The Voyage lesson: solve the hardest thing first
Below are the 7 ideas that stayed with me, in Oliver’s words. You can read this without watching. The full video is above.
1. Compute is now a strategy problem. And it’ll get harder
Oliver dates the shift to around March 2026, when the frontier labs raised enormous rounds and moved quickly to lock up compute. By May and June, it was existential for Odyssey.
“What we saw was really deals that we had handshaken were the next day disappearing.”
Odyssey treated it as a code red. It signed deals with Amazon and AMD while, in his words, “the terms are changing every day, the deposit amounts are going up, everything’s changing, the compute availability, the locations of the compute.”
He also pushed back on the hope that this is temporary:
“I think there’s equally a case to be made that it will actually get dramatically worse, that compute will become more scarce, will become more expensive, more constrained.”
A robot company that trains models is buying from the same hyperscalers and neoclouds as the frontier labs. Compute belongs in the plan from the first raise.
2. The fundraise became a compute partnership search
Oliver’s point is that hyperscalers have so much demand that they choose which companies to work with. Money alone stops being enough.
“The question is, well, what other value beyond the money you have are you providing to that hyperscaler in this dynamic? And so we treated our fundraise, our Series B, as a sort of partnership search process.”
Odyssey partnered closely with Amazon, AMD and NVIDIA. Amazon and AMD Ventures came into the Series B, and NVentures invested earlier in 2026. For Oliver, the partnership plus the compute commitment was about access “not just now but in the future as well.”
So here’s a key question robotics founder needs to consider before a raise. What do you offer a compute provider beyond cash? Odyssey’s answer was strategic investors who also supply the chips.
3. Coding agents make chips portable
Odyssey trains on what Oliver calls a hybrid fleet: NVIDIA, AWS Trainium and AMD. Each chip is used for the parts of model production where it performs best.
“The time where you could be rigid about your model or your engineering, purely optimizing for one chip, good luck.”
Language models changed the cost of switching. Oliver calls the idea of CUDA lock-in “a bit overblown”:
“The idea that now we can simply have coding agents help us port our models to different chip sets, that means that some of the bottleneck that used to exist in saying we can’t possibly think about adopting this exotic new chip type.”
He adds that newer chip companies know porting has to be easy for them to win customers.
What stood out to me is the importance of portability. For model companies, they need to be able to use whatever capacity that shows up. Portability is a hedge against scarcity. Thanks to coding agents, the CUDA lock-in is less of the case and model companies can increasingly benefit from other chip supplies.
4. Data is abundant now, and world models may become a data source
In early 2024, Odyssey sent people out with camera backpacks to film the world. It stopped, because a market of data vendors now does that work.
“There is a huge, massive, growing ecosystem of companies that are collecting physical AI data and then selling it to companies like us and to other labs.”
When I asked whether any type of data is still hard to find, his answer was “Not anymore.” Odyssey works with 10 to 12 data companies, and for an unusual request, “three of them would raise their hand and say, I’ll go get this tomorrow.”
He also described a dystopian example of data collecting: grandmothers in a care home sewing with GoPros on, footage that ends up in foundation models.
Egocentric video matters, but it is a little less than half of Odyssey’s data mixture. A model that only sees the world from a human’s point of view risks becoming narrow. He compares it to coding in language models: the main application, but probably well under half of the pre-training data.
He also gives the data vendors more credit than they usually get:
“I think this ecosystem of data companies has driven more progress than we give them credit for.”
Then he flips it. A world model can become the environment where other AI learns:
“Maybe placing a language model inside a world model is the ultimate RL environment, is the ultimate data generator for a language model.”
The driving test supports part of that view. Odyssey post-trained Odyssey-3 on 30 hours of real driving, then ran a comparison with 20 hours of simulation from a simulator Oliver calls “kind of like Roblox.” Performance dropped, but far less than the data gap suggested. Oliver’s explanation: the model learns concepts like steering and stopping, and those look the same in a blocky simulator.
I find this training loop fascinating: world models take in data to train, then as it gets better, become the data provider for next generation of models. What does that mean to the future of egocentric data collection companies?
