Figure Just Signed a Deal for Up to 100,000 Nvidia GPUs — More Than It Has Ever Raised

A humanoid robotics company just made a purchase this week that looks less like buying motors and batteries and more like the compute deals frontier AI labs sign.

The Deal Itself

Initial deployment is targeted for the second half of 2027 in Barstow, Texas. Nscale will become both a Figure shareholder and its preferred compute provider, supplying the power, data center capacity, GPU compute, and orchestration platform needed to train and run future generations of Figure's Helix models.

Why a Robotics Company Needs This Much Compute

Figure CEO Brett Adcock put it plainly: "To bring humanoid robots to every home in the world, we are largely constrained by data and compute. Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute."

Behind that is what Adcock calls a "physical AI flywheel": train Helix on Nvidia's Vera Rubin infrastructure, validate it in Nvidia's Isaac Sim simulation environment, then deploy the trained model into Figure's Nvidia-chip-equipped robots — and as those robots work in the real world, they generate new data that feeds the next generation of models. Training, simulation, deployment, and data collection form a self-reinforcing loop.

This compute deal lines up with another move Figure made earlier this month: the launch of Index, a data-collection app designed to capture video of ordinary people performing real-world tasks — because the training data a general-purpose robot needs simply isn't available to buy at the volume, diversity, or quality Helix requires. Before its formal launch, Index already had 44,000 weekly active users across 108 countries, with the pipeline processing more than 35 minutes of new task video every second. In other words, Figure is now spending heavily on two fronts simultaneously: buying compute to train the model, and paying to collect real-world data — both have to scale in lockstep for the model to actually get smarter.

Interestingly, while the training side is piling on compute without limit, Figure has kept on-robot inference extremely lean: through 4-bit quantization combined with model parallelism split across two onboard GPUs, it cut computational overhead to 1/23rd of cloud-hosted inference while keeping total power draw under 60 watts — because the robot's battery budget still has to cover motors, sensors, and safety systems. That detail says a lot: you can throw unlimited money at training-side scale, but once a model has to fit inside a physical robot body, the constraints of batteries and heat dissipation in the real world can't be talked around — and that's one of the most fundamental differences between "physical AI" and purely software-based AI.

A Breakup That Pushed Figure Into Buying Its Own Compute

There's a less-discussed backstory here: not long before signing this compute deal, Figure had publicly split from OpenAI. According to The Decoder, Adcock's stated reason was that solving embodied AI requires vertical integration: "We can't outsource AI for the same reason we can't outsource our hardware." The partnership ended in under a year, largely because OpenAI's language-model expertise didn't transfer directly to the unique challenges of learning in the physical world — Figure's internal team includes researchers with over a decade of robot-learning experience who understand how to train and debug models on real hardware in ways OpenAI's team didn't. A subtler factor: OpenAI had reportedly signaled its own ambitions to enter humanoid robotics directly, even building out a data-collection team of roughly 100 people — which made Figure realize its "partner" could turn into a direct competitor at any moment.

That breakup pushed Figure straight into a position where it had to solve its own compute problem — if it could no longer lean on OpenAI's model capability, it had to train its own frontier models in-house, and training frontier models in-house means securing your own compute supply chain. In a sense, this Nscale deal is the first concrete proof point Figure has delivered since the OpenAI split that it's serious about going it alone.

Robotics Companies Are Starting to Look Like Frontier AI Labs With Arms and Legs — But Capital Is Clearly Outrunning Commercialization

What actually matters here isn't the 100,000-GPU number itself — it's what the deal signals about how competition in humanoid robotics is being redefined. The old assumption was that robotics companies competed on hardware engineering: how precise the motors are, how dexterous the hands are, how long the batteries last. But Figure's move here suggests the real moat is shifting toward who can secure more GPUs, more data center capacity, and more real-world human behavioral data — a competitive logic nearly identical to that of pure software AI labs like OpenAI and Anthropic. Nvidia is playing the same game from its own angle: through its Cosmos world-model family, Isaac GR00T robot foundation models, and Jetson Thor onboard compute chips, Nvidia is folding the entire training-simulation-deployment pipeline into its own stack, positioning itself to become the "operating system" of embodied AI — much as Android became the operating system for smartphones. Figure, Boston Dynamics, Agility Robotics, and other robotics companies are, in effect, apps running on top of that operating system; whoever secures more underlying compute and plugs into that stack earliest has the best odds of surviving the race.

But there's a side to this story that's hard to ignore: capital is flowing into the sector noticeably faster than actual commercialization is catching up. Figure closed a funding round earlier this year that valued the company at $39 billion, while it has only a few hundred units actually deployed in real commercial settings — a gap between valuation and real-world deployment scale that's common across the sector, not unique to Figure. As a point of contrast, Elon Musk himself admitted back in January that not a single Tesla Optimus robot was performing "useful work" in any Tesla facility — more than 1,000 Gen 3 units sitting in Texas and California factories were there mainly for "learning and data collection," not production, a striking gap from his own prediction a year earlier that units would be mass-produced and put to real use within the year. Figure, by comparison, has its second-generation robot actually working on BMW's production line at its Spartanburg plant — one of the few humanoid robots genuinely operating in a commercial production environment, which is part of why Figure could confidently sign a multi-billion-dollar compute deal and get Nscale to take an equity stake. Even so, veteran roboticists remain broadly skeptical: iRobot co-founder Rodney Brooks has bluntly called the vision of humanoid robots as catchall general-purpose assistants "pure fantasy thinking."

For anyone tracking embodied AI — operators and investors alike — this story carries a dual signal. On one hand, evaluating a robotics company's competitiveness can no longer stop at how gracefully its robot folds laundry in a demo video; the real question is whether it has enough compute and data supply behind it to sustain continuous iteration — that's becoming the new price of admission in this race. On the other hand, when valuations and compute commitments keep growing faster than actual commercial deployment, the industry as a whole may be further from a working business model than capital markets' optimism would suggest. Whoever manages to turn "burning money to train models" into "robots that reliably do paid work in factories, warehouses, and homes" first is the one who'll actually come out ahead.

Sources: Figure / Nscale / The Decoder / WOWTALE / Interesting Engineering