Amazon is partnering with a start-up building AI systems that can simulate the physical world, betting on the emerging technology in an effort to expand the reach of its Trainium chips.
The Seattle-based group contributed to a $310 million (€267 million) funding round for Odyssey ML, which values the company developing so-called world models that can create 3D environments at $1.45 billion including the new capital.
Other participants included the investment arms of chip groups Nvidia and AMD, the CIA-affiliated fund IQT, venture firm Natural Capital, Google’s chief scientist Jeff Dean and investor Elad Gil.
Odyssey is among a handful of early-stage companies developing models that are trained on physics and the relationships between objects in order to go beyond the language-based models underpinning today’s most popular AI tools.
READ MORE
World models are part of a “second wave of AI” set to revolutionise industries from robotics to gaming, said Jay Zaveri, a partner at Natural Capital. “We have taken a human brain-like construct and only taught it language.”
Odyssey co-founders Oliver Cameron and Jeff Hawke previously worked on self-driving technology.
Cameron, the group’s chief executive, said Odyssey’s models “will have a much more complete understanding of the world . . . physics, body language, dynamics — all these things that exist in the world that language doesn’t really capture”.
The company’s staff of 55 is spread between London, Zurich and Palo Alto. Many were recruited from AI labs and self-driving start-ups.
As part of the deal, Odyssey will use Amazon Web Services as its preferred cloud provider and deploy the company’s latest chips, boosting Amazon amid fierce competition from market leader Nvidia and Google.
Ron Diamant, a vice-president at AWS, said Odyssey’s input would help further Trainium’s development. “My goal is to be able to tell my son that I built the best AI accelerator in the world,” he said.
Amazon chief executive Andy Jassy has bet heavily on AI, including a commitment to invest as much as $83 billion in Anthropic and OpenAI. Earlier this year, he told investors the company had some $225 billion in outstanding contracts for Trainium chips. Executive chair Jeff Bezos, meanwhile, is pursuing his own AI start-up, called Prometheus.
Odyssey’s partnerships with chip companies will offset some of the heavy costs of training and running models that it claims push the frontier in world simulation.
In one demonstration, Cameron showed the model build a multiplayer version of the 1997 Nintendo game Goldeneye based solely on “pixels, actions and sounds”, rather than being trained on the game’s specific physics.
The company’s tools run on Nvidia’s H200 and B200 chips, with usage costs per person ranging between $2 to $4 per hour, said Cameron. The cost does not include training the model.
Luna Schmid, a London-based partner at GV, formerly Google Ventures, said the heavy spending on training and running world models would help unlock a “GPT-3” moment for world models, referring to OpenAI’s language model breakthrough in 2020 which set the stage for today’s AI boom. GV also participated in the funding round. – Copyright The Financial Times Limited 2026
