Figure will spend $3.5 billion on Nscale compute to train its humanoid robots

Figure announced on September 3, 2026 that it has signed a strategic partnership with AI infrastructure provider Nscale to deploy up to 100,000 Nvidia GPUs on the new Vera Rubin platform, backed by an initial $3.5 billion compute commitment that the companies say they intend to scale beyond $6 billion. The GPUs land at a Nscale facility in Barstow, Texas, with initial deployment targeted for the second half of 2027. As part of the deal, Nscale is also making an undisclosed equity investment in Figure and becoming what the companies describe as Figure's preferred compute provider.

What the compute is actually for

Figure has been building out its Helix robot-control model using Index, a dataset of real-world human activity that the company says has generated 30 minutes of new video footage every second since launch, downloaded 264,000 times across 108 countries. Turning that volume of physical-world video into a usable robot policy is the compute-bound step the Vera Rubin deployment is meant to unblock -- Nvidia CEO Jensen Huang, in Figure's own announcement, called the deal evidence that "humanoid robots extend physical AI into the world designed for people."

Nvidia is on every side of this deal

Nscale CEO Josh Payne, quoted separately by The Next Web, put the pitch more simply: "physical intelligence is AI's next frontier." But Nvidia's role here is bigger than a chip vendor's applause line. Nvidia is an investor in both Figure and Nscale, meaning the company that manufactures the GPUs Nscale is buying to supply Figure also holds equity stakes in the customer receiving the compute and the intermediary selling it -- what Interesting Engineering describes Nvidia itself characterizing as a "robotics flywheel": Figure trains on Vera Rubin through Nscale's cloud, validates in Nvidia's Isaac Sim, then deploys the resulting models on Nvidia silicon inside the robots themselves. None of that makes the deal less real, but it's worth naming plainly rather than treating the transaction as three independent parties reaching an arm's-length price.

The bill nobody's itemizing yet

A $3.5-billion, multi-year compute commitment is exactly the kind of AI spend that's simple to announce in a press release and hard to attribute to results afterward -- there's no invoice line item for "humanoid robots that actually work" the way there's one for GPU-hours. Figure and Nscale get to describe this as infrastructure for the future of physical AI today; whether that compute converts into robots doing useful work at a cost that justifies $3.5 billion, rather than into an expensive training run with a good demo reel, is a question that won't have an answer until well after the 2027 deployment date this announcement points to.