Figure’s Compute Deal Makes Humanoid Training Look Like an AI Lab Play
Figure AI’s new video is called “Scaling Compute,” and the name undersells it. The company has signed a multi-year, multi-billion-dollar strategic partnership with AI cloud provider Nscale to deploy up to 100,000 GPUs on NVIDIA’s next-generation Vera Rubin platform — an initial compute commitment of $3.5 billion with explicit plans to scale beyond $6 billion. Deployments begin in the second half of 2027 at Nscale’s facility in Barstow, Texas.
One hundred thousand GPUs. For context, that’s the scale national AI labs talk about — pointed, in this case, at a single problem: making the Figure 03 useful at work.
What the compute is for
The capacity is earmarked to train Helix, Figure’s proprietary vision-language-action foundation model — the stack we track as the leading commercial VLA and the system running Figure’s BMW pilot fleet today. The deal also folds into Figure’s recently unveiled Index initiative, a data-crowdsourcing program aimed at feeding training pipelines at a scale one company’s robots can’t generate alone.
And the partnership isn’t just a customer contract: Nscale is taking a strategic equity stake in Figure, and the two companies are exploring putting humanoid units directly into Nscale’s data-center supply chain. Figure 03 units appeared in the announcement imagery carrying Nscale branding — a hint that the first “customer” of the Figure fleet may be the company powering its training runs.
Why this is the real dividing line
The humanoid industry’s quiet split is now out in the open. Figure ended its OpenAI collaboration in early 2025 and bet on owning the whole stack; this deal is that bet monetized at infrastructure scale. Tesla can lean on its own AI training fleet; Unitree just went the opposite direction and open-sourced its model family. Figure is buying the third path: vertical integration through raw capital — raise billions, own the models, own the data flywheel, and now own a guaranteed share of next-generation American compute to train on.
For buyers and competitors alike, the signal is the same: the moat in humanoids is no longer who has the best hardware demo. It’s who can afford to train on 100,000 GPUs and keep the models proprietary. Figure just made itself the most capitalized bet in the industry — and the Barstow facility gives it a distinctly American supply chain story at exactly the moment foreign-made robots are being squeezed out of the US market.
The sober footnote
Compute commitments of this size are announced more often than they are fully deployed, and Vera Rubin hardware doesn’t exist at volume yet — 2027 is the honest start date. But even as a signed commitment, $3.5–6 billion of dedicated training compute puts Figure in a class no humanoid peer, public or private, has matched. The video is titled “Scaling Compute.” The subtitle might as well be “scaling away from the pack.”