Figure AI Bets $6 Billion That Robots Need More GPUs, Not More Data
Chips & GPUs
Figure AI signed a partnership with Nscale to deploy up to 100,000 GPUs on Nvidia's Vera Rubin platform for training its next-generation Helix robot models, an initial 3.5 billion dollar commitment that could scale beyond 6 billion.
AP
Staff
September 8, 2026 · 4 min read
Figure AI has signed a strategic partnership with cloud provider Nscale to deploy up to 100,000 GPUs on Nvidia’s Vera Rubin platform, starting with an initial 3.5 billion dollar commitment and stated intent to scale beyond 6 billion dollars over time. The deal, aimed squarely at training the next generation of Figure’s Helix models, the AI system controlling its humanoid robots, targets first deployments in the second half of 2027 at a facility in Barstow, Texas. The scale of the commitment says something important about where the actual bottleneck in humanoid robotics now sits: not in hardware design, and not primarily in data collection, but in raw access to compute.
Why a robotics company needs this much compute
Training a policy that lets a humanoid robot generalize across real homes, factories, and unstructured environments is a fundamentally different computational problem than training a language model on internet text. Our own explainer on how robots learn to walk using AI covers why: most training happens in simulation first, running thousands of parallel attempts simultaneously, before policies get validated against messy real-world physics that never perfectly matches the simulated environment they were trained in. That combination, massive parallel simulation plus the real-world validation loop layered on top of it, requires a scale of compute that has increasingly become the actual limiting factor on how fast a company like Figure can iterate, more than any single breakthrough in model architecture or robot hardware design.
Figure’s own framing of the deal makes this explicit: the company describes the compute capacity as essential specifically for training next-generation Helix models, not for manufacturing more robots or improving hardware components. That’s a meaningful signal about where Figure believes its competitive edge actually needs to come from over the next several years. The company has effectively concluded that whoever controls the largest, most efficient training compute for embodied AI, not necessarily whoever builds the most refined physical robot, will end up leading the category.
Why Nvidia’s newest platform specifically
Vera Rubin is Nvidia’s next-generation compute platform, and Figure’s decision to build its next Helix generation on it rather than current-generation hardware reflects a broader pattern across the AI industry: frontier labs and robotics companies increasingly commit to unreleased or newly released hardware generations years in advance, betting that the performance gains justify the risk of building around infrastructure that isn’t yet fully proven in large-scale production. That’s a similar dynamic to the compute commitments Anthropic has made with Google’s TPUs and other frontier labs have made with next-generation Nvidia silicon, extending a pattern already well established among language model labs into the physical robotics space.
Figure’s position in a genuinely crowded field
Figure reached a 39 billion dollar private valuation after raising 1 billion dollars in September 2025, a figure that, at the time, exceeded some analysts’ entire projected valuation for the whole humanoid robotics market nearly a decade out. That valuation puts real pressure on Figure to demonstrate genuine, differentiated progress rather than incremental hardware updates, and a compute commitment of this scale is one of the clearest ways a company can signal serious intent to investors and competitors alike, well before the resulting model improvements are actually visible in a deployed robot.
The broader humanoid robotics field remains genuinely competitive and, per independent analysis, considerably less mature than headline deployment numbers often suggest. One detailed review of the sector found that widely circulated claims about tens of thousands of deployed humanoid units do not survive contact with actual company filings, noting that neither Tesla nor Figure has published an audited production count for their respective robots. Figure’s own most concrete public deployment result to date comes from an 11-month deployment of two Figure 02 units at BMW’s Spartanburg, South Carolina plant, which reportedly contributed to producing more than 30,000 BMW X3 vehicles and loaded over 90,000 sheet metal components across roughly 1,250 operational hours. That’s a genuinely useful, verifiable data point, but it’s a narrow, single-task industrial deployment, not evidence of the kind of general-purpose capability the compute investment in next-generation Helix models is explicitly aimed at producing.
The compute arms race extending into physical AI
Figure’s deal is best understood as part of a broader pattern rather than an isolated bet. Boston Dynamics has committed its entire 2026 electric Atlas production allocation to Hyundai and Google DeepMind. AgiBot scaled from 1,000 humanoid units in 2025 to 10,000 within months in early 2026. Across the sector, the actual constraint increasingly isn’t whether a company can build a capable-looking robot, prototypes and demos have become common, it’s whether that company can train a genuinely generalizable AI system controlling it, and that training bottleneck now runs directly through the same compute infrastructure competition already reshaping the language model industry. A 3.5 to 6 billion dollar GPU commitment from a robotics company that has not yet published an audited production count for its existing robots is, in that light, less an outlier than a signal of where the entire humanoid robotics industry believes the real competitive battle now sits.
Whether that bet pays off depends heavily on something no compute commitment alone can guarantee: whether more training compute genuinely translates into a robot that generalizes reliably across real, unstructured environments, or whether the harder unsolved problems in humanoid robotics, balance, safe physical contact, and reliable performance outside tightly controlled settings, require breakthroughs that scale doesn’t straightforwardly buy.
See Humanoid Press’s original coverage of the Nscale partnership for further technical detail.
