On September 3, Figure and the AI cloud provider Nscale signed a multi-year partnership built around a single resource: GPUs. The deal covers up to 100,000 Nvidia GPUs on the new Vera Rubin platform, with an initial compute commitment of $3.5B and stated intent to scale past $6B. First systems land in the second half of 2027, at Nscale’s site in Barstow, Texas.
Nscale is also taking an undisclosed equity stake in Figure and becoming its preferred compute provider. Figure gets guaranteed capacity to train its Helix models; Nscale gets a customer it partly owns.
The number that matters is the one above Figure’s own balance sheet
Figure’s Series C closed a year earlier, in September 2025, at $1B-plus and a $39B valuation, led by Parkway Venture Capital with Nvidia, Intel Capital, Salesforce and Qualcomm Ventures. Before that, a $675M Series B in 2024 brought in Microsoft, Nvidia, Amazon’s Industrial Innovation Fund and Jeff Bezos personally. Add it up and Figure has raised a little under $2B in its history.
The Nscale commitment alone is nearly double that, before it even reaches the $6B ceiling both companies have flagged. A humanoid robotics company that has never shipped a consumer product just signed a compute bill bigger than everything investors have ever put into it.
Nscale can afford the bet. It closed a $2B Series C at a $14.6B valuation in March, added roughly $3B in debt financing weeks before this deal to build GPU campuses in Texas and North Carolina, and is preparing a U.S. IPO after telling investors it has $51B in contracted revenue on the books. For Nscale, Figure is one customer inside a much larger buildout.
What the money is actually for
Nvidia CEO Jensen Huang described the arrangement as a closed loop: train Figure’s models on Vera Rubin through Nscale’s cloud, validate them in Nvidia’s Isaac Sim, deploy them on Nvidia GPUs inside Figure’s own robots. Nvidia sits underneath all three steps.
Figure CEO Brett Adcock’s framing was narrower and more specific: “To bring humanoid robots to every home in the world, we are largely constrained by data and compute.” The company has spent the past four months proving the data half of that sentence isn’t a bottleneck anymore. Its Index platform, which pays contributors to record themselves doing household chores, logged 16 million video uploads and 30 minutes of footage arriving every second, close to five years of human labor uploaded per day. Figure has already committed more than $1B to data collection and compute over Index’s first 12 months alone.
Raw footage does not train a model by itself. Turning 16 million videos into a robot that can hold a shift folding laundry or clearing shelves takes training runs at a scale Figure has said it could not previously afford to run. That is the gap the Nscale deal is meant to close, and it lines up with the company’s own production ramp, a monthly output curve that has been doubling but was always going to hit a compute ceiling before it hit a demand ceiling.
The financing pattern Wall Street already has a name for
Nscale isn’t just selling Figure compute. It’s buying equity in the company that will spend $3.5B–$6B renting that compute back from it. That is the same structure Nvidia itself has used with the AI labs it backs, and it draws the same criticism: when an infrastructure provider owns a slice of its own customer, the revenue the deal generates says less than it appears to about independent demand. Wall Street has already flagged these circular arrangements as a way to make a market look bigger than the number of people actually buying the finished product would support.
That question matters more in humanoid robotics than in cloud software, because the finished product here is not a subscription. It’s a machine that has to physically complete a shift on a warehouse floor or in someone’s kitchen, at a cost a customer will actually pay. A recent shipment count put real-world humanoid placements at roughly half of what industry projections had implied for the year. The compute race and the deployment race are not running at the same speed.
Figure isn’t raising alone
The rest of the field is spending at a similar clip. Skild AI raised $1.4B in January at a valuation above $14B, led by SoftBank with Nvidia and Bezos also in. Apptronik closed a $520M Series A extension in February, pushing its round past $935M. Physical Intelligence was reported in March to be in talks for roughly $1B at above $11B, though that round had not been confirmed closed months later. Across the sector, humanoid robotics startups have raised $8.6B so far in 2026, 1.8 times everything raised in all of 2025.
The labor angle: the clock starts at 2H 2027, not today
Nothing about this deal puts a robot in a job tomorrow. The GPUs land in Barstow in the second half of 2027, and everything before that date is training, not deployment. But the number attached to that date is the one worth tracking. A company betting $3.5B–$6B of someone else’s compute on a specific 18-month runway is making a public claim about when general-purpose humanoid labor becomes commercially real — not a five-year hand-wave, a dated infrastructure buildout with a site address attached.
The occupations in the blast radius are the ones Figure has already targeted with its own robots and its own Index task list: warehouse tote-handling and package sorting, retail restocking, and, if Adcock’s “every home in the world” framing holds, a category of household labor that has never before had a capital market pricing its automation. None of that shows up in a jobs report yet. The compute bill is the leading indicator; the headcount numbers come after the 2027 systems go live, not before.
