Figure commits $3.5 billion to Nscale for up to 100,000 Nvidia Vera Rubin GPUs to train its Helix robot model

Figure AI has agreed to buy at least $3.5 billion of computing capacity from Nscale, the humanoid robot developer said on September 3, 2026, in one of the largest infrastructure commitments yet disclosed by a robotics company. The partnership covers up to 100,000 GPUs on Nvidia's Vera Rubin platform, with the initial deployment targeted for the second half of 2027 in Barstow, Texas, according to Figure's announcement.

Figure said the commitment is designed to scale to more than $6 billion. Nscale is also making a strategic investment in Figure, the company said, without disclosing the size, price or structure of that stake. The two companies said they would explore using Figure humanoids to expand Nscale's own supply chain.

The compute is earmarked for Helix, the vision, language and action model that controls Figure's robots. Brett Adcock, Figure's chief executive, said in the release that Helix becomes more capable the way every learned system does, with more data and more compute. Nvidia chief executive Jensen Huang and Nscale chief executive Josh Payne were also quoted endorsing the deal.

Where the data comes from

The GPU order follows two disclosures that explain why Figure wants so much capacity. On August 25, 2026 the company introduced Index, an app through which paid contributors film themselves doing everyday physical tasks. Figure reported 264,000 downloads across 108 countries, more than 44,000 weekly active users, 16 million uploaded videos and $15 million paid to contributors to date. It said that for every 1,000 hours collected the dataset spans 373 unique tasks, 1,146 unique objects and 116 unique environments, and it committed to spend more than $1 billion on data and compute over the following 12 months.

On September 17, 2026 Figure published results for Helix 2.5. The company said a Figure 03 robot running an Index-pretrained model completed household tasks such as tidying toys, folding towels and making beds in 30 Bay Area homes it had never seen, with a 56% zero-shot success rate against 9% for the same model trained from scratch. Figure said the new model matched its predecessor while using half the task-specific data, and that a scaling law fitted on smaller runs predicted the loss of its largest run with 0.54% error.

Company background and financing

Figure raised a $1 billion Series C at a $39 billion post-money valuation in September 2025, according to the company. The Nscale commitment is therefore 3.5 times the cash raised in that round, and the upper target of more than $6 billion is roughly six times it. Figure has not said how the commitment will be funded beyond Nscale's undisclosed investment.

Competitive context

ROBOTNESS has not found a compute contract of comparable size disclosed by any other humanoid developer. Skild AI raised $1.4 billion at a $14 billion valuation in January 2026 to build a general robot foundation model, Apptronik closed a $520 million Series A extension in February 2026, and Agility Robotics agreed in June 2026 to go public through a merger with Churchill Capital Corp XI at a $2.5 billion equity value. Each of those companies also trains learned controllers, but none has tied itself to a dedicated six-figure GPU cluster.

The technology in plain terms

Helix maps camera images and spoken or written instructions directly to motor commands for a whole humanoid body. Figure's bet is that, as with large language models, generalization comes mainly from pretraining on huge volumes of varied human experience and then fine-tuning on a small amount of robot data. Index supplies the first ingredient cheaply by paying people to record themselves, and the Vera Rubin cluster supplies the processing needed to turn tens of millions of videos into a model.

What it means globally

The deal moves the competitive frontier in humanoids from hardware toward data and compute budgets. If Figure's scaling results hold, companies without access to billions of dollars of training capacity may have to license models from those that do, much as many software firms build on a handful of frontier language models today.

Risks and open questions

The commitment runs ahead of revenue Figure has not disclosed. Delivery depends on a 2027 hardware generation and a Texas site that Figure named but did not describe. The 56% household success rate, while far above the 9% baseline, still means nearly half of attempts fail, and Figure's results are self-reported rather than independently benchmarked.

What to watch

Investors will look for the size and terms of Nscale's stake, any new Figure equity round to fund the commitment, and progress at Barstow ahead of the second-half 2027 start.