Nscale Commits $3.5B in AI Compute to Power Figure Humanoids

Nscale’s agreement with Figure commits an initial $3.5 billion of AI compute to humanoid robotics, with the parties intending to scale the relationship beyond $6 billion. The deal matters because it moves the industry’s bottleneck beyond robot mechanics: training general-purpose machines now requires infrastructure commitments approaching the scale of major foundation-model programmes.

Nscale Is Trading Compute for Strategic Position

Under the companies’ official agreement, Nscale will become a Figure shareholder and its preferred compute provider. The partnership could deploy the NVIDIA Vera Rubin platform with up to 100,000 GPUs. Initial deployment is targeted for the second half of 2027 at Nscale’s site in Barstow, Texas.

The structure is more consequential than a conventional cloud contract. Nscale gains equity exposure to a customer whose demand it will help satisfy, while Figure secures a route to scarce next-generation capacity. Nscale’s full-stack cloud platform competes on access to dedicated GPU infrastructure, making a flagship robotics workload commercially useful even before the full capacity is operating.

Figure Says Helix Is Compute-Constrained

Figure’s announcement says progress is increasingly constrained by the data and compute required to train Helix. The company describes Helix as a vision-language-action model that links perception, language and learned control. Its newer Helix 02 system extends control across walking, manipulation and balance.

That approach makes humanoid development resemble frontier AI more than traditional industrial automation. Better hardware still matters, but capability increasingly depends on collecting physical interaction data and repeatedly training models against it. Business 2.0’s coverage of NVIDIA’s physical AI stack shows how simulation, world models and accelerated computing are converging around the same constraint.

Vera Rubin Raises Both Capacity and Delivery Risk

The proposed infrastructure centres on NVIDIA Vera Rubin, a rack-scale platform designed for training and inference workloads. Nscale has separately announced plans to deploy more than 100,000 Vera Rubin GPUs in Europe, indicating that the Figure commitment belongs to a broader infrastructure expansion rather than a single bespoke cluster.

However, the headline value is an initial compute commitment, not cash already transferred or GPUs already operating. Deployment starts in the second half of 2027, and the phrase “up to 100,000” defines potential scale rather than a guaranteed installed count. Power delivery, construction, networking, cooling and NVIDIA’s product schedule all affect execution. The same infrastructure economics appear in Business 2.0’s analysis of AMD data-centre consolidation.

Commercial Proof Must Follow Technical Progress

Figure has moved beyond laboratory demonstrations. Its Figure 03 deployment at BMW targets more complex factory sequencing after the previous generation contributed to vehicle production. Yet demonstrations and limited deployments do not establish fleet-level economics. Buyers still need evidence on uptime, intervention rates, safety, maintenance and useful work completed per shift.

Competitive pressure is growing. Business 2.0 has tracked China’s humanoid market tests, Chery-backed AiMOGA’s expansion and physical AI in autonomous logistics. Figure’s compute commitment gives it training capacity, but rivals can compete through lower hardware cost, manufacturing scale, specialised datasets or tighter customer integration.

What This Means for Robotics Investors

The agreement makes compute procurement a balance-sheet issue for robotics companies. Investors should separate the $3.5 billion commitment from Nscale’s undisclosed equity investment and from the non-binding intention to exceed $6 billion. They should also test whether model improvements translate into fewer human interventions and more productive operating hours.

The strongest outcome would be a repeatable loop: deployed robots generate useful data, that data improves Helix, and improved models make additional deployments economical. Without that loop, vast GPU capacity becomes an expensive input rather than a competitive moat. Bloomberg’s reporting confirms the equity investment remains undisclosed, reinforcing the need to separate verified commitments from implied valuation signals.