← Back to feed

Odyssey AI closed a $310 million Series B at a $1.45 billion valuation, bringing its total funding to $337 million since founding in late 2023. The round was led by Natural Capital with participation from Amazon, AMD Ventures, GV (Google Ventures), EQT, and In-Q-Tel, the fund affiliated with the US intelligence community.

Nvidia backed Odyssey’s Series A. It is not in this round.

The Investor Swap

The Nvidia absence is the most significant signal in the deal structure. Nvidia’s venture arm has been active across the AI ecosystem, putting $40 billion in equity deals in 2026 alone. Stepping aside from a follow-on in a portfolio company it backed at Series A is not standard operating procedure.

In its place, Amazon entered as a strategic backer. Odyssey simultaneously committed to AWS Trainium as its preferred cloud compute platform, which is how these deals work: Amazon provides capital and Trainium access, the portfolio company builds workloads on Trainium. AMD Ventures also joined, adding a second Nvidia competitor to the cap table.

The practical message is that Odyssey has chosen a path that runs on non-Nvidia silicon, at least for its primary training infrastructure.

What Odyssey Builds

Odyssey builds general-purpose world models that simulate physical environments with realistic physics. The applications span robotics, gaming, autonomous driving, and defence. Earlier this year the company shipped Agora-1, a multi-agent world model supporting up to four simultaneous players operating in a single generated world.

World models are a distinct category from standard LLMs and diffusion models. Where language models compress the statistical structure of text and image models compress visual distributions, world models attempt to compress causal dynamics: what happens next given initial conditions. The training data is video and sensor logs rather than text corpora, and the inference task is simulation rather than generation.

That makes compute requirements unusual. World model training is GPU-intensive in a different way from language model training: longer sequences of correlated frames, heavier reliance on temporal convolutions, and different memory access patterns than transformer forward passes. Whether Trainium handles those workloads efficiently at scale is now something Odyssey will demonstrate under real training conditions.

Context

The $1.45 billion valuation puts Odyssey in a tier well below the largest world model competitors. Google’s Genie connected 280 billion Street View images to its world model and feeds Waymo’s training data pipeline. Physical Intelligenceraised at a $4 billion valuation before being acquired. Manus was blocked from a $2 billion acquisition by Chinese regulators.

At $337 million raised and $1.45 billion valued, Odyssey is the most-capitalised independent world model lab in the West that is not already part of a larger platform. The Intel investment thesis implies that compute flexibility, not just model quality, determines which labs survive to the next training run.

Key Numbers

• Round size: $310M Series B
• Post-money valuation: $1.45B
• Lead: Natural Capital
• Co-investors: Amazon, AMD Ventures, GV, EQT, In-Q-Tel
• Total funding: $337M since late 2023
• Preferred cloud: AWS Trainium
• Previous investor absent: Nvidia (backed Series A)