Nscale’s $3.5B Figure Deal Could Bring 100,000 Vera Rubin GPUs to Humanoid AI

Nscale and humanoid-robotics company Figure have signed a multi-year partnership centered on cloud capacity for physical AI. The headline figure is an initial $3.5 billion compute commitment, with an intent to expand the arrangement to more than $6 billion. Nscale will also take an undisclosed equity stake in Figure and become its preferred compute provider. 14

The agreement is substantial, but its largest figures should be read as targets rather than installed capacity: the companies describe the potential to deploy up to 100,000 Nvidia Vera Rubin GPUs, with initial deployment targeted for the second half of 2027 at Nscale’s Barstow, Texas site. 14

What Nscale and Figure agreed to

The commercial structure has three connected parts:

• Compute: Nscale has committed $3.5 billion of cloud-computing resources for Figure, with a stated intention to scale beyond $6 billion. This is a commitment for compute capacity, distinct from the undisclosed amount Nscale is investing in Figure itself. 14
• Equity: Nscale will become a shareholder in Figure. Neither company disclosed the size or valuation of that investment. 14
• Preferred-provider position: Figure has named Nscale its preferred compute provider, tying Nscale’s infrastructure business more closely to Figure’s future model-development and deployment needs. 14

That combination matters. A conventional cloud contract makes a provider a vendor; an equity stake and preferred-provider arrangement align Nscale more directly with the robotics company’s growth. It does not, however, establish the value of future orders beyond the announced commitment or guarantee that the maximum GPU scale will be reached.

Scale, hardware and timing

The partnership is planned around Nvidia’s Vera Rubin platform. Nscale and Figure said the agreement represents the potential to deploy up to 100,000 Nvidia GPUs, with the first GPUs targeted for Barstow beginning in H2 2027. 4

Two distinctions are essential when assessing the announcement:

1. Up to 100,000 GPUs is a potential ceiling, not a confirmed day-one installation. 4
2. More than $6 billion is an intended expansion, not the initial contracted amount. The announced starting commitment is $3.5 billion of compute. 14

Those caveats are especially important for infrastructure projects, where hardware deliveries, data-center readiness, power availability and customer demand all affect the pace of deployment.

What the compute is intended to power

Figure plans to use the infrastructure to support its Helix AI models and the development of autonomous, general-purpose humanoid robots. 24

Robotics workloads can extend well beyond training a single foundation model. Building and improving a humanoid system can require large-scale model training, simulation, synthetic-data generation, testing and continual iteration as robots encounter real environments. That mix creates a plausible case for persistent cloud demand if robot fleets reach meaningful scale—but commercial deployment remains the key variable.

The physical-AI “flywheel”

The deal fits Nvidia’s broader robotics-development approach: centralized AI infrastructure for training and simulation, paired with hardware capable of running models on machines in the field.

Nvidia describes Isaac Sim as a framework for robotics simulation, testing and synthetic-data generation in physically based virtual environments. Its wider Isaac platform includes simulation and robot-learning frameworks, accelerated libraries, AI models and development workflows for autonomous machines, including humanoids.

In practical terms, the proposed workflow looks like this:

1. Cloud infrastructure: Nscale’s Vera Rubin capacity would supply the compute for model training and large-scale simulation.
2. Simulation and synthetic data: Isaac Sim can create virtual environments for testing robotic behavior and generating training data.
3. On-robot inference: Deployed robots need low-latency perception and control, which requires compute on the machine as well as resources in the cloud.
4. Iteration: Results from real-world operation can inform subsequent testing, simulation and model updates.

This is a development model rather than a confirmation of a specific finished Figure deployment architecture. The companies have announced the infrastructure partnership; the operational details and eventual volume of deployed robots remain to be proven.

Could Figure robots be used by Nscale?

The partners also plan to explore using Figure robots in Nscale’s supply chain. 8 That is a prospective use case, not an announced purchase order, operational rollout or confirmed timetable for robots in Nscale facilities.

Still, the possibility is strategically notable: Nscale could become both the provider of AI infrastructure and, potentially, an enterprise user of the physical-AI systems it helps train.

How the Figure deal fits Nscale’s larger compute strategy

The Figure agreement broadens Nscale’s exposure from general AI-cloud demand into embodied AI. Separately, Nscale has been associated with a reported roughly $45 billion, six-year Anthropic compute arrangement at its Monarch campus in West Virginia, involving about 460 megawatts of capacity.

The two arrangements point to different demand patterns:

• Frontier-model cloud demand centers on large, centralized capacity for developing and serving advanced AI models.
• Physical-AI demand combines centralized training and simulation with an eventual need to support learning and inference across robot deployments.

The Figure deal is much smaller in stated dollar value than the reported Anthropic arrangement, but it tests a different long-term market: whether increasingly capable robots create repeatable demand for simulation, multimodal model training and fleet improvement. The answer will depend on execution—particularly the Texas build-out, access to power and hardware, and Figure’s ability to commercialize humanoid robots at scale.

The bottom line

Nscale’s partnership with Figure is more than a GPU supply announcement: it combines a $3.5 billion compute commitment, a possible expansion above $6 billion, an equity investment and preferred-provider status. 14 But its most eye-catching numbers—100,000 GPUs and the larger spending target—remain prospective. The first Vera Rubin deployment is targeted for the second half of 2027, making delivery and robot-market adoption the real tests of whether physical AI becomes a durable new source of cloud-compute demand. 14