Odyssey’s $310 million round is not a clean break from Nvidia. It is a better story than that: Amazon, AMD and Nvidia are all now crowded around the same young world-model lab.
Odyssey has raised $310 million at a $1.45 billion valuation, and the interesting part is not only the size of the check. According to the Financial Times, Amazon joined the round as Odyssey made AWS its preferred cloud provider and agreed to use Amazon’s Trainium chips. Nvidia’s venture arm also participated, alongside AMD, IQT, GV, Natural Capital, Jeff Dean and Elad Gil. So don’t read this as Nvidia being pushed out of the room. Read it as a sign that even a startup backed by Nvidia wants more than one way to pay for serious AI compute.
That matters if you’re building anything that depends on video, robotics or simulation. Odyssey is not chasing another chatbot. The company is trying to build world models: AI systems that can simulate how physical environments behave, with objects, motion, physics and interaction carrying the work rather than text alone. Amazon wants Trainium to be part of that workload. AMD wants a place in the same conversation. Nvidia, still the chip company everyone else is trying to measure themselves against, is not walking away either.
The company was founded in late 2023 by Oliver Cameron and Jeff Hawke, both of whom came out of self-driving technology. Cameron previously led Voyage before Cruise acquired it, while Hawke worked on autonomous driving systems at Wayve. That background is not a biographical flourish. Self-driving forced engineers to deal with the hard question inside world models long before the term became fashionable: what happens next in the real world, and can a machine predict it fast enough to act?
Odyssey’s bet is that the same problem reaches far beyond cars. If a model can simulate the next state of a street, a room, a factory floor or a game environment, you can use it for robotics training, autonomous vehicle testing, scientific simulation and interactive media. You don’t need to buy the grandest version of that claim to see the direction. Language models learned to predict words. World models are trying to predict environments.
The company’s own system, Odyssey-2, is described as a causal, autoregressive video model. It generates video frame by frame, rather than handing over a finished clip from a prompt. The company has said it can produce continuous streams while responding to inputs, a very different job from making a short, pretty video that falls apart once the camera moves or an object needs to behave consistently. Frankly, that is where much of today’s video AI still feels thin. It can look convincing for a moment. Holding the world together is harder.
The FT reported one demo in which Cameron showed Odyssey’s model creating a multiplayer version of GoldenEye from pixels, actions and sounds, rather than from the game’s physics engine. That is the kind of example you should pay attention to, because it is specific enough to test the claim. A simulated game world has rules, movement, cause and effect. If the model can preserve those while users interact with it, the result is closer to an environment than a clip.
Amazon’s role in the round is the hardware story. AWS has spent heavily on Trainium because it does not want every large AI workload to default to Nvidia GPUs forever. The FT reported that Odyssey will use AWS as its preferred cloud provider and deploy Amazon’s latest chips, while also noting that Odyssey’s tools currently run on Nvidia H200 and B200 chips. That combination is the real point. Startups are not making moral choices between chip vendors. They’re trying to get enough compute, at a price and supply level that lets them keep building.
You can see why Amazon wants this customer. World models are compute hungry in a different way from a text model answering prompts. Real-time simulation means continuous generation, low latency and expensive serving if the system is used by many people at once. Cameron told the FT that usage costs for Odyssey’s tools run between $2 and $4 per person per hour on Nvidia’s H200 and B200 chips, before training costs. That is not a rounding error. If Trainium can bring those costs down or give Odyssey more predictable capacity, Amazon gets a proof point it badly needs.
IQT’s participation gives the round another edge. The CIA-affiliated investor backs companies with potential national security uses, and a model that can simulate real environments from video is plainly relevant to defense, autonomous systems and intelligence work. Odyssey has not announced defense contracts in this round, and you should not pretend it has. But IQT’s presence tells you the technology is being watched for more than gaming demos.
The valuation also deserves a clear eye. A $1.45 billion price tag for a company founded in 2023 is a large vote of confidence in a category that is still early. Natural Capital partner Jay Zaveri described world models to the FT as part of a second wave of AI. Maybe that is right. But you do not get a second wave just by naming it. You get it when the models work reliably, the costs come down, and customers can use the system without treating every hour of generation like a luxury item.
Odyssey now has money, chip partners and a cleaner story than most young AI labs. It also has a harder test than most of them. If world models become foundational, the company’s AWS and Trainium deal will look like an early infrastructure decision that mattered. If they stay as impressive demos with punishing costs, the investor list will not save the category.
Also read: ASML's CEO says Terafab is an opportunity but only if the company doesn't run out of machines to sell • Pramaana Labs is building the proof layer that AI needs before it touches your tax return or your health • Canada's pension giant bets $741 million on India's data center boom as US and European pipelines stall
