Odyssey hits $1.45B valuation with $310M raise for world models

Key Takeaways

• Odyssey raised $310 million in Series B funding at a $1.45 billion valuation, with Amazon and AMD Ventures participating.
• World models simulate real-world physics in real-time, positioning them as potential replacements for traditional game engines.
• AWS becomes Odyssey's preferred cloud provider, with models optimized for Trainium chips instead of Nvidia hardware.

Odyssey, a two-year-old startup building AI systems that simulate real-world physics, just closed a $310 million Series B at a $1.45 billion valuation. Natural Capital led the round, with Amazon, AMD Ventures, and GV joining. The company now claims $337 million in total funding.

The founders bring serious self-driving credentials. CEO Oliver Cameron co-founded Voyage, an autonomous vehicle company GM's Cruise acquired. CTO Jeff Hawke was an engineer at Wayve, the UK self-driving startup that recently raised over $1 billion. Their bet: the same AI techniques that teach cars to understand physical environments can power a new category of simulation tools.

What are world models and why do they matter?

World models represent the next step beyond text-generating large language models. Where ChatGPT predicts the next word in a sentence, world models predict the next state of a physical environment. Feed them visual data from the real world, and they learn to simulate accurate physics: how objects fall, collide, and interact.

To gather training data, Odyssey borrowed a page from Google Earth's playbook. Instead of driving camera-equipped cars, they sent people hiking with cameras strapped to their backs. The approach captures environments cars can't reach and builds richer datasets for simulation.

40 milliseconds to generate a frame

The company's current models can generate and stream frames with just 40ms latency. That's fast enough for interactive, real-time experiences. A viral demo of Odyssey-1 showed users walking through a simulated environment with playable physics, no traditional game engine required.

Odyssey now offers several world models targeting different applications. Video game creation is the most visible use case, but robotics training may prove more valuable. Teaching robots in simulation is cheaper and safer than physical trial and error. Companies like Tesla already rely on simulation for their humanoid robot programs.

AWS over Nvidia: a strategic hardware shift

Amazon's participation comes with strings. Odyssey named AWS its preferred cloud provider and will optimize models for Amazon's Trainium chips. This is a direct challenge to Nvidia's dominance in AI infrastructure.

The move makes financial sense. Nvidia's H100 chips remain scarce and expensive. AWS Trainium offers a cheaper alternative, and Amazon is eager to prove its custom silicon can handle demanding AI workloads. Odyssey gets favorable cloud pricing; Amazon gets a high-profile customer to showcase Trainium's capabilities.

Hacker News discussions have picked apart this partnership skeptically. Some developers question whether Trainium can match Nvidia's performance for the specific computations world models require. Others see it as a pragmatic bet: diversifying away from a single supplier is smart risk management, even if Nvidia's chips are technically superior.

The angel investor list reads like an AI who's-who

Beyond the institutional money, Odyssey assembled a notable roster of angel investors. Jeff Dean, Google's former AI lead, participated. So did Elad Gil, Y Combinator president Garry Tan, Vercel CEO Guillermo Rauch, and Cruise founder Kyle Vogt.

Vogt's involvement stands out. He founded Cruise, the company that acquired Cameron's previous startup Voyage. He's now backing Cameron's next venture. That kind of repeat investment from someone who saw Cameron's work up close says something about the founders' execution ability.

Can world models actually replace game engines?

This is the billion-dollar question. Traditional game engines like Unity and Unreal have spent decades perfecting physics simulation, rendering pipelines, and developer tools. World models offer a fundamentally different approach: instead of coding rules for how objects behave, you train an AI on examples and let it learn the rules.

The advantage is flexibility. A trained world model can generalize to situations the developers never explicitly programmed. The disadvantage is unpredictability. Game designers want precise control over player experiences. An AI that approximates physics might introduce glitches or unexpected behaviors.

Cameron has described the company's goal as achieving a "GPT-3 moment" for world models. GPT-3 proved that language models could be useful for practical applications, not just research demos. Odyssey wants to prove the same for physics simulation.

The competitive landscape

Odyssey isn't alone in this space. Google DeepMind has published research on world models. Runway, the AI video company, has explored similar techniques. Nvidia itself is investing in simulation for robotics through its Omniverse platform.

What differentiates Odyssey is focus. The company has built its entire product around world models rather than treating them as one feature among many. The founding team's autonomous vehicle background also gives them deep experience with the specific challenges of real-time physical simulation.

Frequently Asked Questions

What is a world model in AI?

A world model is an AI system that learns to simulate the physics and dynamics of real-world environments. Unlike text-based language models that predict the next word, world models predict the next state of a physical scene.

How does Odyssey collect training data for its models?

Odyssey sends people out with cameras strapped to their backs, similar to how Google collected Street View imagery. This captures real-world visual data from diverse environments that the models use to learn physics.

Why is Amazon investing in Odyssey?

Amazon wants to promote its Trainium chips as an alternative to Nvidia hardware. Odyssey agreed to make AWS its preferred cloud provider and optimize its models for Trainium, giving Amazon a showcase customer for its custom AI silicon.

Can world models replace traditional game engines?

Potentially for some use cases. World models can generate interactive environments without explicit physics programming, but they trade precision for flexibility. Traditional engines still offer more control for developers who need predictable behavior.

Who are the founders of Odyssey?

CEO Oliver Cameron previously co-founded autonomous vehicle startup Voyage, which GM's Cruise acquired. CTO Jeff Hawke was an engineer at Wayve, a UK self-driving company.

Source: Startups | TechCrunch / Julie Bort