Figure, the humanoid robot company, has signed a multi year agreement with the AI cloud provider Nscale for an initial 3.5 billion dollars of compute, with stated intent to scale past 6 billion, on up to 100,000 Nvidia GPUs of the Vera Rubin generation to be deployed in Barstow, Texas from the second half of 2027.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source Nscale becomes a Figure shareholder and its preferred compute provider through a strategic investment of undisclosed size.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source 2 Unite.AI 2026-09-03 Nscale sites in the UK, Norway and Europe and early Vera Rubin deployment; Figure at 39 billion dollars after a September 2025 Series C with NVIDIA investing; BotQ producing Figure 03; Index at about 35 minutes of training data per second. Open source Figure has raised about 1.9 billion dollars in its life, so the initial commitment alone is close to twice its total funding.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source Our assessment, with high confidence, is that this is the largest dedicated compute commitment by a robotics company on record; with moderate confidence, that the equity component means Nscale is financing part of its own customer; and with moderate confidence that the deal marks the point where humanoid robotics stops being a hardware business with a software problem and becomes a frontier model business with a hardware product.

The shape of the deal

Three features define it. The size, 3.5 billion dollars rising to more than 6 billion, is spread over multiple years and is not a lump sum payment, as eWeek stresses.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source The hardware, up to 100,000 GPUs on Vera Rubin, is Nvidia's next rack scale platform, and Nscale describes itself as among the first global providers to deploy it, with existing sites in the United Kingdom, Norway and Europe.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source 2 Unite.AI 2026-09-03 Nscale sites in the UK, Norway and Europe and early Vera Rubin deployment; Figure at 39 billion dollars after a September 2025 Series C with NVIDIA investing; BotQ producing Figure 03; Index at about 35 minutes of training data per second. Open source And the equity, which turns a vendor relationship into a cross holding: Nscale invests in Figure, Figure commits spend to Nscale, and the two say they will explore using Figure's humanoids in Nscale's own supply chain.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source

The circularity is the point, not a flaw to hide. Figure's chief executive Brett Adcock said the company is largely constrained by data and compute.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source A company that has raised 1.9 billion dollars cannot pay 3.5 billion in cash; it can pay in a mix of future revenue, future fundraising and equity in itself, which is what this structure looks like from the outside. The undisclosed size of Nscale's stake is the number that would settle how much of the compute Figure is paying for and how much it is being given.

Why a robot company needs a gigawatt class cluster

Figure's Helix models learn from video of humans and robots performing tasks. Its Index program had collected more than 16 million videos by late August and ingests about 35 minutes of footage per second.2 Unite.AI 2026-09-03 Nscale sites in the UK, Norway and Europe and early Vera Rubin deployment; Figure at 39 billion dollars after a September 2025 Series C with NVIDIA investing; BotQ producing Figure 03; Index at about 35 minutes of training data per second. Open source 3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source That data volume is a frontier lab's training problem, and the deal is priced like one. The company is also producing robots: its BotQ facility turns out about one Figure 03 unit per hour, and units are deployed at BMW's Spartanburg plant.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source Every deployed robot generates more data, which needs more compute, which needs more money. The flywheel Nvidia's Jensen Huang described for robotics, combining training, validation and deployment, is a description of this feedback loop.2 Unite.AI 2026-09-03 Nscale sites in the UK, Norway and Europe and early Vera Rubin deployment; Figure at 39 billion dollars after a September 2025 Series C with NVIDIA investing; BotQ producing Figure 03; Index at about 35 minutes of training data per second. Open source

Who gains and who loses

Nvidia gains whichever way the deal goes: it is an investor in Figure at the 39 billion dollar September 2025 valuation, its GPUs fill the Barstow site, and its Vera Rubin platform gets a marquee physical AI customer.2 Unite.AI 2026-09-03 Nscale sites in the UK, Norway and Europe and early Vera Rubin deployment; Figure at 39 billion dollars after a September 2025 Series C with NVIDIA investing; BotQ producing Figure 03; Index at about 35 minutes of training data per second. Open source Nscale gains an anchor tenant for a large US site, an equity position in a 39 billion dollar company and a story that distinguishes it from the many clouds selling the same chips. Figure gains committed capacity two years out, at a moment when frontier labs are locking up the same generation of hardware.

The losers are Figure's competitors that did not sign a comparable deal. If Helix quality scales with compute the way language models did, a rival with a smaller cluster is a rival with a worse robot, and eWeek's framing that humanoid winners may need infrastructure operations rivaling frontier labs describes a bar most robotics startups cannot clear.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source Figure's existing shareholders lose some clarity: an undisclosed equity issuance to a vendor dilutes them by an amount they cannot yet compute.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source And the Texas grid takes on another large load that will compete with the AI data centers already sited there.

The counter case

The thesis that this is a frontier scale commitment could be wrong because commitments are not spending. A multi year agreement with stated intent to scale is a ceiling, not a floor, and if Figure's fundraising stalls the contract will be renegotiated, as compute contracts in this cycle routinely are.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source The deployment does not begin until the second half of 2027, which leaves two years for Vera Rubin delivery schedules, Figure's revenue and Nscale's own financing to change.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source A second objection is that robotics may not scale like language. Sixteen million videos is a large dataset, but whether Helix improves smoothly with another 100,000 GPUs is an open research question that this deal assumes rather than answers.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source If it does not, the cluster is a very expensive way to discover that the bottleneck was elsewhere.

What to watch

  • Nscale's stake disclosed. If Figure's next funding round or a filing reveals the size of Nscale's equity by mid 2027, the real cash cost of the compute becomes visible; a stake above a few percent would mean Nscale is materially financing its customer.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source
  • Figure raises again before Barstow opens. A new round at or above 39 billion dollars before the second half of 2027 would fund the commitment; a down round or no round would put the 3.5 billion figure at risk.2 Unite.AI 2026-09-03 Nscale sites in the UK, Norway and Europe and early Vera Rubin deployment; Figure at 39 billion dollars after a September 2025 Series C with NVIDIA investing; BotQ producing Figure 03; Index at about 35 minutes of training data per second. Open source
  • First GPUs powered in Barstow. Vera Rubin capacity live in Texas in the second half of 2027 confirms both Nvidia's delivery and Nscale's buildout; a slip into 2028 would suggest the platform, not the customer, is the constraint.1 Nscale 2026-09-03 Initial 3.5 billion dollars scaling to more than 6 billion; up to 100,000 NVIDIA GPUs on Vera Rubin; second half of 2027 in Barstow, Texas; Nscale takes equity and preferred provider status; Helix models; Payne and Adcock quotes. Open source
  • A published scaling result for Helix. If Figure shows, within a year, that Helix task success improves predictably with training compute, the deal is rational on its own terms; if not, it is a bet on a curve nobody has drawn.3 eWeek 2026-09-03 Multi year commitment, not a lump sum; about 1.9 billion dollars raised in total; 675 million in 2024; 39 billion post money; BMW Spartanburg deployment; BotQ at about one robot per hour; more than 16 million Index videos by late August. Open source

Figure has priced its future like a frontier lab. Whether a robot improves the way a language model does is the question the 3.5 billion dollars is meant to answer.