Figure-Nscale $3.5B Nvidia GPU Deal 2026: 100K Vera Rubin

Last verified: October 9, 2026.

On September 3, 2026, Figure AI (the humanoid robotics company behind the Helix physical AI model) and Nscale (a UK-based full-stack AI cloud platform) signed a multi-year strategic partnership that represents an initial commitment of $3.5 billion in compute, scaling to over $6 billion for 100,000 NVIDIA Vera Rubin GPUs targeted for initial deployment in the second half of 2027 in Barstow, Texas. As part of the agreement, Nscale is making a strategic investment in Figure. Nscale provides the power, compute platform, and orchestration layer integrating NVIDIA's advanced AI infrastructure; Figure trains its Helix physical AI model on the cluster. The deal makes Figure one of the largest buyers of NVIDIA Vera Rubin compute in 2026-2027, alongside OpenAI, xAI, Oracle, and Anthropic, and is the largest single humanoid-robotics compute commitment announced to date. This walkthrough reads the deal the way a robotics company, an AI infrastructure investor, or a humanoid-robotics competitor encounters it: the deal mechanics, the Barstow Texas deployment, the NVIDIA Vera Rubin platform, the Helix model, the Nscale strategic investment, and the implications for the humanoid-robotics race.

The Deal Mechanics

The Figure-Nscale deal structure:

The $3.5 billion initial commitment makes this one of the largest single-customer compute commitments in 2026-2027. For comparison, OpenAI committed to $10+ billion in compute with Microsoft and Oracle in 2025-2026, Anthropic committed to $5+ billion with AWS and Google Cloud, and xAI committed to $5+ billion with NVIDIA directly. Figure's $3.5 billion commitment is in the same tier as the largest AI compute buyers, despite Figure being a much smaller company than OpenAI or Anthropic. The implied valuation of the compute: roughly $35,000-$40,000 per GPU-equivalent for the initial $3.5B (assuming an average of $35K/GPU including NVIDIA AI Enterprise software and Nscale's orchestration platform markup).

The strategic investment component is unusual for a compute deal. Nscale is taking an equity stake in Figure as part of the agreement, which aligns the two companies' incentives for the multi-year term. The structure is a hybrid of (a) traditional compute-as-a-service, where the customer pays per GPU-hour, and (b) a strategic partnership, where the supplier has equity upside. The structure is the highest-leverage deal type for both parties: Figure secures compute at lower effective cost (because Nscale's equity upside offsets part of the compute margin), and Nscale captures long-term upside if Figure becomes the leading humanoid-robotics company.

Barstow, California: The Deployment Site

The Barstow, California deployment is the right pick for several reasons:

The high desert location (Barstow is in the Mojave Desert at ~2,200 feet elevation) is the right pick for free-air cooling. The cool, dry desert air is ideal for the data center cooling system: the mechanical chillers can be supplemented (or replaced) by direct outside air for most of the year, reducing the energy cost of cooling by 30-40% relative to a traditional data center. The savings are meaningful for a 100,000-GPU cluster, which can consume 100-200 MW of power for compute and 30-60% additional for cooling; a 30% cooling reduction is roughly 10-15 MW of avoided power.

The Barstow location is also relatively close to Figure's engineering headquarters in San Jose (3.5-hour drive) and to Los Angeles (1.5-hour drive). The proximity to the engineering team matters for the operational support of the cluster; the cluster is not a fully lights-out facility, and the engineering team visits regularly for hardware swaps, software updates, and operational troubleshooting.

The cluster deployment is expected to be phased: the first phase (5,000-10,000 GPUs) is targeted for H2 2027, with subsequent phases scaling to 100,000 GPUs over the multi-year term. The phasing is driven by NVIDIA's Vera Rubin supply availability; the GPU supply ramp is a multi-quarter process.

NVIDIA Vera Rubin: The GPU Platform

NVIDIA Vera Rubin is the post-Blackwell GPU platform, announced at NVIDIA GTC 2025 and shipping in 2027:

Vera Rubin's 1.5x-2x performance improvement over Blackwell at the same power envelope is the meaningful generation-over-generation gain. The improvement is driven by the HBM4 memory bandwidth (the bottleneck for large-model training) and the NVLink 6 interconnect (the bottleneck for distributed training across GPUs). For Figure's Helix model — a large multimodal model that processes video, audio, and proprioception data — the HBM4 bandwidth is the most important improvement, because the data pipeline is the bottleneck for training.

The 100,000-GPU cluster is the largest single training cluster announced for Vera Rubin to date. The scale is comparable to OpenAI's largest training cluster (100,000+ H100s in 2024) and xAI's Colossus cluster (200,000 H100s/H200s in 2024-2025). The 100,000-GPU scale is the right size for training a foundation model with 10+ trillion parameters, which is the scale at which Figure's Helix model is targeted.

The cluster will run on NVIDIA AI Enterprise software (the enterprise version of NVIDIA's AI software stack), with Nscale providing the orchestration layer (Kubernetes, observability, multi-tenant management). The combination is the standard enterprise AI cluster stack in 2026-2027.

The Helix Model: Figure's Physical AI

Helix is Figure's foundation model for humanoid robotics. The model is the AI that controls the Figure 02 humanoid robot (the second-generation robot that Figure has been developing since 2024):

Helix is the second generation of Figure's foundation model (the first was the February 2024 Helix V1, which was a smaller-scale model trained on Figure's robot fleet data). The new Helix is the larger-scale version that Figure is targeting for commercial deployment in 2027-2028. The 100,000-GPU compute commitment is sized for the larger-scale Helix training, which is the most compute-intensive AI training run in the humanoid-robotics industry.

The "physical AI" framing is the strategic context. Physical AI is the AI that controls physical systems (humanoid robots, autonomous vehicles, robots in factories) rather than information systems (chatbots, search, content generation). The training requires massive video + sensor + teleoperation datasets, which Figure has been collecting since 2023. The compute requirement for physical AI training is comparable to (and in some cases higher than) the compute requirement for LLM training, because the multimodal data and the action-decoder architecture are more compute-intensive than pure-text training.

The commercial target — "humanoid in every home" — is the long-term thesis. The near-term commercial deployments are in manufacturing, logistics, healthcare, and retail, where the humanoid can perform repetitive physical tasks. The home deployment is the long-term horizon, requiring safety, reliability, and cost reductions that the 2026 generation of humanoids is not yet ready for.

The Strategic Investment Component

The Nscale strategic investment in Figure is the unusual part of the deal:

The strategic investment pattern is the same one that has been used in the major AI compute deals of 2024-2026: the compute provider takes an equity stake in the AI company, which aligns incentives and reduces the effective compute cost. The Microsoft-OpenAI deal was the first major example; the AWS-Anthropic and Google-Anthropic deals followed in 2024-2025. The Figure-Nscale deal is the first major humanoid-robotics example.

The strategic investment is also the most leveraged position for Nscale. If Figure becomes the leading humanoid-robotics company (a $1+ trillion market opportunity by 2030), the equity stake is worth tens of billions. If Figure does not, the equity stake is the lower-bound loss for Nscale. The expected value calculation is heavily skewed to the upside, which is the right shape for an equity-as-part-of-compute-deal.

The estimated investment amount of $50-200 million is based on Nscale's funding history (Nscale raised $1.4B in 2024-2025) and Figure's last round ($1.9B in February 2025 at a $10B+ valuation). A $100 million strategic investment would represent roughly 1% of Figure's post-money valuation, which is in the typical range for a strategic investment in a Series A-E company.

What This Means for the Humanoid Robotics Race

The deal positions Figure at the front of the humanoid-robotics compute race. The competitive context:

Figure's 100,000-GPU Vera Rubin commitment is the largest single humanoid-robotics compute commitment announced. Tesla's Optimus is the closest competitor, with 50,000+ H100s for the internal training, but Tesla's cluster is H100 generation (the previous generation) and is internal-only (no external customers). Figure's Vera Rubin is the next generation, and the cluster will serve Figure's external customers (manufacturing, logistics, healthcare).

The deal is also the largest GPU commitment by any humanoid-robotics company to NVIDIA. The previous largest was Tesla's 50,000+ H100s, which is a significant commitment but at the prior generation. Figure's 100,000 Vera Rubin is the next-generation commitment that positions Figure as the largest external humanoid-robotics customer of NVIDIA's AI infrastructure.

The deal is the first major Nscale partnership. Nscale is a UK-based full-stack AI cloud platform that launched in 2024. The Figure deal is the largest single customer commitment for Nscale and is the deal that establishes Nscale as a major player in the AI cloud market. The success of the Figure deal is the proof point for Nscale's ability to deliver large-scale AI infrastructure.

What to Watch in 2026-2028

Three signals settle the rest of 2026 and 2027-2028 for the humanoid-robotics race.

First, the Barstow deployment timeline. The first phase of the deployment is targeted for H2 2027. The deployment timeline depends on NVIDIA's Vera Rubin supply ramp; the GPU supply ramp is a multi-quarter process. The deployment-on-time signal: Barstow phase 1 goes live in Q3-Q4 2027.

Second, the Helix model release. Figure is targeting a 10T+ parameter Helix model. The model release depends on the training timeline; a multi-month training run on 50,000-100,000 GPUs typically takes 3-6 months. The expected Helix model release is Q4 2027 to Q1 2028.

Third, the commercial deployment. Figure has been targeting manufacturing and logistics as the first commercial deployments. The first commercial deployments are expected in 2027-2028. The deployment scale signal: Figure deploys the first 1,000+ humanoids in commercial settings in 2027-2028.

Read next

NVIDIA GTC Berlin October 2026 Preview is the conference preview for the Vera Rubin platform announcement, and Jensen Huang $1 Trillion: Blackwell + Rubin Order Book is the supplier-side perspective on the order book that the Figure deal is part of.