Figure AI Nscale Partnership Secures 100000 Vera Rubin GPUs

Figure AI Commits Massive Compute to Scale Helix Training

On September 3, 2026, Figure AI disclosed a strategic partnership with UK-based AI cloud provider Nscale that commits up to 100,000 NVIDIA Vera Rubin GPUs for training its Helix vision-language-action models. The initial agreement carries a $3.5 billion value with an explicit intent to expand beyond $6 billion over multiple years. Initial hardware deployments are scheduled for the second half of 2027 at Nscale’s leased facility in Barstow, Texas, a 234 MW site previously tied to Bitcoin mining operations and already hosting Microsoft infrastructure. Nscale will also take an equity stake in Figure and become its preferred compute provider while exploring humanoid deployments in its own supply chain.

This scale of dedicated training infrastructure marks one of the largest single commitments yet seen for any humanoid robotics developer. Figure’s CEO Brett Adcock emphasized that progress on Helix is now limited primarily by data volume and compute throughput rather than algorithmic breakthroughs alone. The company recently launched its Index dataset initiative, which generates 35 minutes of diverse humanoid trajectory data every second, yet Adcock noted that data alone cannot close the gap without corresponding increases in training capacity. The Nscale arrangement directly addresses that bottleneck by providing rack-scale Vera Rubin systems optimized for mixture-of-experts training and large-scale embodied AI workloads.

NVIDIA Vera Rubin Platform Advantages for Robotics Workloads

The Vera Rubin architecture, unveiled by NVIDIA at CES 2026, pairs a new Vera CPU with the Rubin GPU to deliver claimed 10x reductions in inference token cost versus Blackwell and a 4x reduction in GPUs required for mixture-of-experts model training. Each Rubin GPU targets 50 petaFLOPS of NVFP4 inference performance, a substantial uplift that enables faster iteration over the high-dimensional trajectory datasets required for humanoid control policies. Figure plans to leverage the full robotics flywheel described by NVIDIA CEO Jensen Huang: training on Nscale’s cloud, validation inside NVIDIA Isaac Sim, and eventual edge deployment on NVIDIA Jetson AGX Thor chips embedded in the robots themselves.

For a company targeting general-purpose humanoids, these efficiency gains translate directly into accelerated model release cycles. Earlier generations of Helix relied on comparatively modest clusters; scaling to 100,000 GPUs represents a step-function increase in simultaneous experiment throughput. Training runs that previously required weeks could compress into days, allowing Figure to ingest larger volumes of real-world and simulated interaction data before each iteration. This cadence matters because humanoid policy learning remains sample-inefficient compared with language or vision models.

Competitive Context Among Humanoid Developers

The timing of the Nscale commitment places Figure ahead of several peers in securing next-generation training capacity. Tesla’s Optimus program continues to iterate on in-house Dojo clusters and Blackwell systems, yet has not disclosed a comparable multi-year external cloud reservation of this magnitude. Unitree Robotics and Boston Dynamics maintain active development pipelines but operate at smaller reported training footprints. Figure’s valuation reached $39 billion after its September 2025 Series C round that raised over $1 billion, giving it the balance sheet flexibility to underwrite such an infrastructure bet. The equity investment from Nscale further aligns incentives, turning the cloud provider into a stakeholder with skin in the physical-AI outcome.

Nscale itself is a 2024-founded neocloud that has already closed multi-billion-dollar rounds and secured hyperscale contracts with Microsoft. Its willingness to take Figure equity and explore humanoid integration in data-center logistics reflects a broader thesis that physical AI will become both a major compute consumer and a labor solution for the infrastructure layer itself. The Barstow site’s 234 MW capacity provides headroom for phased rollout, though power delivery, cooling, and networking at 100,000-GPU scale will require additional capital beyond the stated compute commitment.

Fleet Economics and Long-Term Implications

From a fleet-economics perspective, the compute reservation signals Figure’s expectation that training costs will remain the dominant line item even after initial robot deployments begin. Each additional GPU-hour improves policy robustness, which in turn reduces the mean time between failures (MTBF) and service interventions once robots reach customer sites. Lower inference token costs promised by Vera Rubin also improve the economics of on-robot edge inference, potentially allowing longer autonomous operation windows between cloud syncs.

The partnership does not disclose exact per-GPU pricing or utilization guarantees, leaving open questions about effective cost per trained token once the cluster reaches steady state. Nscale’s status as preferred provider suggests Figure will route the majority of future training workloads through this channel rather than building or leasing its own dedicated clusters. That decision trades capital expenditure for operating expense predictability but introduces dependency on a single infrastructure partner during a period of rapid hardware generational turnover.

Path to Deployment and Remaining Uncertainties

Initial Rubin GPUs are not expected online until H2 2027, creating a roughly 18-month window during which Figure must continue advancing Helix on existing Blackwell or earlier hardware. The company has already placed Figure 02 humanoids in limited commercial pilots, and the additional compute runway is intended to accelerate the transition to more capable successors. Nscale’s exploration of humanoid use in its own supply chain adds a potential early-adopter channel that could generate valuable operation