What is the Figure-Nscale deal?

Figure, a humanoid robotics company, signed a $3.5 billion compute deal with Nscale, a UK-based GPU cloud provider, for up to 100,000 NVIDIA GPUs. The initial infrastructure is based on NVIDIA's next-generation Vera Rubin platform. The deployment is expected in H2 2027 at a facility in Barstow, Texas. The deal is one of the largest single-customer GPU commitments in history (Figure AI, Nscale, 2026).

Why does Figure need 100,000 GPUs?

Figure needs massive compute to train its humanoid robot foundation models. Humanoid robots require: (1) Vision-language-action (VLA) models trained on millions of hours of robot interaction data, (2) Reinforcement learning from human feedback (RLHF) on physical tasks, (3) World models that simulate the physical environment, (4) Multi-modal perception models. Training a foundation model for a humanoid robot is comparable in compute cost to training a large language model (Figure AI, 2026).

Nscale is a UK-based GPU cloud provider founded in 2024. Nscale specializes in large-scale AI compute deployments using NVIDIA hardware. Nscale has data centers in the UK, Norway, and the US, with a focus on sovereign AI infrastructure. Nscale raised $155M in a Series A round in 2025 (Nscale, 2026).

When will the Figure-Nscale infrastructure be operational?

The initial Figure-Nscale infrastructure is expected to be operational in H2 2027 at a facility in Barstow, Texas. The deal could scale beyond $3.5 billion to $6 billion if Figure exercises the full 100,000 GPU option. The phased buildout allows Figure to start training humanoid robot foundation models as soon as the first compute cluster is online (Figure AI, 2026).

What is Figure's humanoid robot?

Figure 02 is Figure AI's second-generation humanoid robot. Figure 02 stands 5'6' tall, weighs 130 pounds, can carry 44 pounds, has 16 degrees of freedom in the hands, runs on a custom AI model, and has a battery life of about 5 hours. Figure 02 is designed for general-purpose labor in manufacturing, warehousing, and elderly care. Figure AI is valued at $40B with investors including Microsoft, OpenAI Startup Fund, NVIDIA, Amazon, and BMW (Figure AI, 2026).

Figure Signs $3.5B Deal for 100,000 NVIDIA GPUs

Figure signs $3.5B compute deal with Nscale for up to 100,000 NVIDIA GPUs. Initial infrastructure on Vera Rubin platform, deployment H2 2027 at Barstow, Texas. Could scale to $6B. Figure 02 humanoid robot training requires massive compute — comparable to large LLM training (Figure AI, Nscale, 2026).

Data last verified September 2026 from Figure AI, Nscale, and NVIDIA press releases.

Figure-Nscale deal — key terms

Field	Detail
Initial value	$3.5 billion
Maximum value	$6 billion (if 100,000 GPU option exercised)
GPU count (initial)	~50,000 NVIDIA Vera Rubin GPUs
GPU count (max)	Up to 100,000 NVIDIA Vera Rubin GPUs
Location	Barstow, Texas
Operational date	H2 2027
Use case	Training humanoid robot foundation models
Partner	Nscale (UK-based GPU cloud provider)

Source: Figure AI, Nscale (September 2026).

Why humanoid robots need massive AI compute

Humanoid robots are one of the most compute-intensive AI applications: (1) Foundation models — training a single humanoid foundation model can take 1-3 months on 50,000+ GPUs, (2) Multi-modal data — robot interaction data is multi-modal (video, audio, haptics, force feedback), generating petabytes per day, (3) Reinforcement learning — robots learn by trial and error, requiring millions of simulated and real-world trials, (4) World models — neural network simulators that model physical interactions with the environment, (5) Safety verification — extensive testing of robot behaviors in simulation before real-world deployment (Figure AI, 2026).

Figure's humanoid robot roadmap

2024: Figure 01 — first prototype, demonstrated basic manipulation tasks.
2025: Figure 02 — second-generation robot, deployed at BMW manufacturing plant for real-world testing.
2026: Figure 03 — third-generation, planned for general-purpose use in manufacturing and warehousing.
2027: Initial commercial deployment at customer sites, large-scale training with the Nscale compute.
2028-2030: Mass production target, $20,000-$30,000 per unit cost, millions of humanoid robots deployed globally.

Competitive landscape for humanoid robots

The humanoid robot market is heating up with several major players: (1) Figure AI — Figure 02 robot, $40B valuation, BMW partnership, (2) Tesla Optimus — Tesla's humanoid robot, expected 2027 commercial release, (3) Boston Dynamics Atlas — Hyundai-owned, focuses on industrial applications, (4) 1X Technologies — backed by OpenAI, Neo humanoid robot for home use, (5) Agility Robotics Digit — focused on warehouse logistics, partnership with Amazon, (6) Apptronik Apollo — backed by Google, manufacturing focus, (7) Sanctuary Phoenix — Canadian, focuses on cognitive AI for humanoid robots. The market is projected to reach $38B by 2035 (Goldman Sachs Research, 2026).

Why Barstow, Texas

Barstow, Texas was chosen for the Figure-Nscale data center for several reasons: (1) Land availability — large tracts of land for power and cooling infrastructure, (2) Solar and wind resources — Texas has excellent renewable energy potential, (3) Low electricity costs — ERCOT (Texas grid) has lower commercial rates than coastal states, (4) Tax incentives — Texas offers 10-year tax abatements for large data center projects, (5) Workforce — Texas has a growing tech workforce, (6) Connectivity — high-speed fiber networks to the major cloud regions, (7) Climate — dry climate reduces cooling water needs (Nscale, 2026).

NVIDIA Vera Rubin platform

The Vera Rubin platform is NVIDIA's next-generation GPU architecture, expected in 2027. Key improvements over Blackwell: (1) 10x performance per watt improvement, (2) HBM4 memory with 50% more bandwidth, (3) NVLink 6 with 4x the bandwidth, (4) Confidential computing support, (5) Improved FP8 and FP4 performance, (6) 2x the AI training throughput per dollar. The Vera Rubin is named after astronomer Vera Rubin, who confirmed the existence of dark matter. NVIDIA's roadmap after Vera Rubin is the Feynman architecture (expected 2028) (NVIDIA, 2026).

Figure AI's competitive moat

Figure's competitive moat comes from: (1) Real-world deployment — Figure 02 is the only humanoid robot deployed in a real BMW manufacturing plant, (2) Foundation model approach — Figure is building a general-purpose humanoid foundation model, (3) Top-tier investors — Microsoft, OpenAI Startup Fund, NVIDIA, Amazon, BMW, (4) Top-tier talent — former Apple, Tesla, Google DeepMind, and Boston Dynamics engineers, (5) Strategic compute — the $3.5B Nscale deal secures long-term GPU access, (6) Safety-first design — extensive safety testing and simulation (Figure AI, 2026).

What this means for the AI industry

The Figure-Nscale deal is significant for several reasons: (1) Validates the humanoid robot market — the largest single-customer GPU commitment to date for robotics, (2) Sovereign AI trend — Nscale's UK origin and Barstow US deployment show sovereign AI demand, (3) NVIDIA's continued dominance — even humanoid robot companies need NVIDIA GPUs, (4) The rise of GPU clouds — Nscale, CoreWeave, Lambda, and others are competing with AWS/Azure/GCP for AI workloads, (5) Compute becomes a strategic asset — humanoid robot companies are willing to spend $3.5B on compute, treating it as a core strategic resource (Nscale, 2026).

Resources and next steps

Follow Figure AI at figure.ai and Nscale at nscale.com. For humanoid robot industry trends, follow Goldman Sachs Research, Morgan Stanley, and Bloomberg. For NVIDIA's Vera Rubin platform, see nvidia.com/data-center. For GPU cloud comparisons, see The Next Platform and ServeTheHome. For AI compute infrastructure investment trends, follow hyperscaler capex reports from Microsoft, Google, Amazon, and Meta.