Figure has signed a multi-year partnership with AI cloud provider Nscale that could put up to 100,000 Nvidia Vera Rubin GPUs behind the humanoid-robot maker’s Helix models. The companies say the first systems are targeted for deployment in the second half of 2027 at Nscale infrastructure in Barstow, Texas.

 

The agreement gives Figure a long-term compute path as it tries to scale physical AI beyond controlled demonstrations. Nscale says the deal starts with a $3.5 billion compute commitment and is intended to grow beyond $6 billion. Nscale is also making a strategic investment in Figure and becoming its preferred compute provider.

 

Figure’s Nscale Deal Puts Compute Behind Helix

Figure announced the partnership on September 3, describing data and compute as the two constraints now shaping development of its Helix vision-language-action models. Unlike a conventional software assistant, Helix has to connect visual perception and language instructions to continuous physical control across a humanoid robot’s body.

 

Nscale’s role is to provide the cloud and data-center infrastructure for that training. The companies are targeting Nvidia’s Vera Rubin platform, which is entering broad production this year and is designed for large-scale training, post-training and agentic inference workloads.

 

The agreement sets out four concrete elements:

  • Potential deployment of up to 100,000 Nvidia Vera Rubin GPUs
  • Initial GPU deployment targeted for the second half of 2027 in Barstow, Texas
  • A $3.5 billion starting compute commitment with intent to exceed $6 billion
  • A strategic Nscale investment in Figure and preferred-provider status

 

Reuters independently reported the transaction on September 3, confirming the multi-year structure and the initial $3.5 billion commitment. The large headline number is important, but the more consequential point is that Figure is treating compute capacity as a core input to robotics development rather than an occasional cloud expense.

 

Humanoid Training Needs a Data-Compute Loop

Figure’s recent Index project explains why the company wants that capacity. When Index came out of stealth on August 25, Figure said more than 264,000 people across over 100 countries had downloaded its data-collection app, with creators uploading more than 16 million videos. At launch, the pipeline was processing about 30 minutes of video every second.

 

In the new Nscale announcement, Figure said Index had already reached 35 minutes of data every second. The purpose is to gather diverse examples of human interaction with real environments so Helix can learn manipulation, navigation and task sequences that do not exist in ordinary internet datasets.

 

That data has to be converted into useful robot behavior through repeated training and evaluation. Figure’s Helix 02 system already controls walking, balance and manipulation through a unified visuomotor model, and the company has demonstrated long-horizon tasks such as unloading and reloading a dishwasher and tidying a living room without step-by-step human control.

 

The challenge is scale. A general-purpose robot needs to handle unfamiliar objects, layouts and physical edge cases, which means the training loop must absorb new examples continuously. More real-world data raises the amount of compute required to train, test and refine the models that turn those examples into reliable actions.

 

Related Coverage

 

Nvidia Rubin Becomes the Infrastructure Layer

Nvidia launched Vera Rubin as a rack-scale AI platform spanning GPUs, Vera CPUs, NVLink networking, storage and Ethernet. The company said in May that Rubin was ramping into full production and claimed up to ten times the agent throughput at scale of its previous Grace Blackwell platform, a vendor benchmark that will still need to be tested across real robotics workloads.

 

For Figure, Rubin’s appeal is less about one benchmark than about building a standardized training stack that can expand with the data pipeline. The partnership does not mean 100,000 GPUs will arrive at once. It creates a ceiling the companies can scale toward, with initial deployment not expected until the second half of 2027.

 

The Texas location also matters. Training large physical-AI models requires more than accelerators: power delivery, cooling, networking, storage and orchestration have to scale together. Nscale is positioning itself as the layer that assembles those systems while Figure focuses on datasets, models and robot deployment.

 

Figure Is Building for Production, Not Just Demonstrations

The compute commitment arrives as Figure pushes its robots into commercial settings. Figure 03 returned to BMW’s Spartanburg plant this summer for logistics work, following Figure 02’s earlier assembly-line deployment. The company also signed a May agreement with Catalyst Brands to deploy humanoids in distribution and logistics operations.

 

Those deployments give Figure something model labs cannot get from simulation alone: recurring contact with messy, variable physical environments. They also create a harder standard for reliability. A model that succeeds in a polished demonstration has to behave consistently across long shifts, changing objects, human coworkers and safety constraints before customers can treat humanoids as normal industrial equipment.

 

The Nscale deal is therefore a bet on a specific scaling thesis. Figure believes that more diverse physical data, paired with much larger training infrastructure, can produce broader robot competence in the same way that bigger datasets and compute transformed language and vision systems.

 

That thesis is not proven. The hardware rollout is more than a year away, the full 100,000-GPU capacity is only a potential scale target, and the economics of training general-purpose humanoids remain uncertain. But the partnership makes Figure’s strategy unusually explicit: collect physical-world data at internet-like scale, match it with frontier computing infrastructure, and use the combined loop to push Helix toward robots that can generalize across jobs instead of being programmed for one task at a time.