# Figure AI Commits $3.5B for 100K NVIDIA Vera Rubin GPUs

On September 3, 2026, autonomous humanoid robotics manufacturer Figure AI announced a multi-billion-dollar strategic infrastructure partnership with specialized AI cloud provider Nscale to deploy up to 100,000 NVIDIA Vera Rubin GPUs. The commercial agreement establishes an initial $3.5 billion capital commitment, with formal structural options to expand beyond $6 billion over the contract term. Scheduled to begin production deployments in the second half of 2027 within Nscale's custom-engineered high-density computing facility in Barstow, Texas, the deal represents the largest dedicated compute commitment ever executed by an embodied robotics enterprise.

The scale of the transaction signals an inflection point in machine learning economics: the frontier compute race is no longer confined to digital chatbots, synthetic voice generators, and web reasoning models. Physical artificial intelligence—spanning bipedal humanoid robots, autonomous mobile industrial platforms, and embodied manipulators—demands foundation model pre-training at parametric scales rivaling the largest frontier text models like OpenAI's GPT-6 and Anthropic's Claude flagships. While large language models train on discrete, static text tokens mined from the open web, Figure's proprietary Helix foundation models must continuously ingest, model, and predict continuous real-world physics across high-bandwidth spatial video, joint torque teleoperation trajectories, and complex multi-finger tactile sensor arrays across dozens of mechanical degrees of freedom.

The partnership structure also underscores the accelerating rise of specialized "neocloud" infrastructure operators. Rather than undertaking the capital-intensive, multi-year engineering challenge of acquiring sub-stations, securing high-voltage power purchase agreements (PPAs), and designing gigawatt-scale liquid cooling facilities directly, Figure partnered with Nscale to deliver turnkey high-density capacity. This separation of operational responsibilities allows Figure's engineering team to focus entirely on neural motor policy synthesis, synthetic simulation environments, and physical robot manufacturing, while offloading data center civil engineering, electrical substation construction, and thermal management to specialized infrastructure partners.

 <a aria-hidden="true" id="executive-fast-facts"></a>  Fast Facts 

- **Announcement Date:** September 3, 2026
- **Principal Commercial Partners:** Figure AI (embodied robotics) and Nscale (specialized GPU neocloud infrastructure provider)
- **Hardware Allocation Target:** Up to 100,000 NVIDIA Vera Rubin graphics processing units
- **Capital Expenditure Structure:** $3.5 billion initial compute commitment, with expansion capacity exceeding $6 billion
- **Facility Geography:** Nscale high-density data center campus located in Barstow, Texas
- **Deployment Horizon:** Commissioning and phased cluster activation slated for the second half of 2027
- **Primary Workload Target:** Large-scale pre-training and reinforcement learning of Figure's Helix embodied foundation models
- **Industrial Significance:** Establishes the largest dedicated compute cluster ever contracted exclusively for physical embodied AI
 

## Embodied AI Scaling and Neocloud Infrastructure Economics

The transition of humanoid robotics from classical control theory toward end-to-end neural network architectures represents a fundamental transformation in physical automation. For decades, robotic systems developed by pioneering research labs relied on analytical physics engines, explicit trajectory optimization, and hand-tuned inverse kinematics. These systems operated with high precision in structured assembly line settings but failed catastrophically when confronted with novel, unstructured environments.

Modern embodied AI replaces handcrafted kinematic heuristics with monolithic foundation models. Figure's Helix architecture processes multi-modal inputs—comprising 60-frame-per-second stereoscopic RGB-D video streams, high-frequency joint encoder telemetry, and tactile fingertip pressure maps—through a unified visual-spatial-action transformer. The network outputs direct continuous joint torque commands at low latencies, enabling the robot to manipulate fragile objects, navigate cluttered industrial warehouses, and adjust dynamically to physical disturbances without written programming rules.

Training an embodied foundation model to generalize across real-world physics requires unprecedented computational scale:

First, the training pipeline demands vast ingestion of multimodal simulation data. Because collecting millions of real-world physical teleoperation hours on physical hardware is mechanically constrained by hardware wear and human operator labor, Figure relies on high-fidelity simulation engines (such as NVIDIA Isaac Sim and customized Omniverse physics environments). These platforms simulate millions of synthetic robots interacting with complex physical objects in parallel, generating trillions of synthetic sensor frames that require massive GPU clusters to process through reinforcement learning loops.

Second, NVIDIA's Vera Rubin architecture represents a critical generational leap beyond current Blackwell (B200/GB200) silicon. Manufactured on advanced process nodes, the Vera Rubin platform introduces sixth-generation Tensor Cores, next-generation High Bandwidth Memory (HBM4) offering multi-terabyte-per-second memory bandwidth, and integrated optical NVLink interconnect fabrics. In embodied AI training, where attention layers must process high-resolution video tokens across extended temporal horizons, memory bandwidth and inter-node communication latency serve as the primary limits on training throughput.

Third, the sheer power density required to operate 100,000 next-generation GPUs has rendered traditional enterprise colocation obsolete. The Barstow, Texas deployment engineered by Nscale utilizes 100% direct-to-chip liquid cooling systems capable of dissipating over 100 kilowatts of thermal energy per server rack. By situating the facility adjacent to utility-scale renewable energy infrastructure and dedicated natural gas generation in West Texas, Nscale secures the hundreds of megawatts of continuous electrical baseline capacity required to run uninterrupted multi-month model pre-training runs.

## AI Compute Deal Comparison (2026)

The table below contrasts Figure AI's historic compute contract against other landmark AI compute superclusters announced in 2026 across provider models, scale, and strategic focus:

| Entity | Infrastructure Provider | GPU Commitment | Silicon Generation | Financial Scale | Primary Workload Scope |
|---|---|---|---|---|---|
| **Figure AI** | Nscale (Neocloud) | Up to 100,000 | NVIDIA Vera Rubin | $3.5B initial, &gt;$6B total | Humanoid embodied intelligence and Helix foundation models |
| **Safe Superintelligence (SSI)** | NVIDIA Direct / Specialized | Undisclosed (Est. 50K+) | NVIDIA Vera Rubin | Undisclosed Enterprise Tier | Pure artificial general intelligence safety research |
| **xAI (Grok)** | NVIDIA / Oracle Cloud | 100,000+ Active | NVIDIA H100 / H200 | $10B+ (Memphis Supercluster) | Frontier multimodal language and reasoning models |
| **OpenAI** | Microsoft Azure | Undisclosed (Est. 100K+) | NVIDIA Blackwell / Custom | $10B+ (Stargate Initiative) | Next-generation GPT frontier reasoning and agentic models |
| **Anthropic** | Amazon Web Services (AWS) | Undisclosed (Est. 50K+) | AWS Trainium2 / Inferentia | $4B+ Capital Partnership | Claude model series pre-training and alignment research |
| **DeepSeek** | Huawei Technologies | 160,000 Provisioned | Huawei Ascend 950DT | Undisclosed Sovereign | Domestic Chinese foundation models and inference clusters |

## Real-World Utility &amp; Policy Implementation

For enterprise industrial leaders, automotive manufacturers, and logistics operators, Figure AI's aggressive compute investment provides clear visibility into the deployment timeline of commercial humanoid labor. Organizations must begin establishing facilities and network readiness for physical AI integration.

### The 4-Step Robotics Compute Planning Playbook

1. **Model Physical AI Training and Inference Asymmetries:** Distinguish strictly between the massive compute required to train embodied foundation models and the lightweight silicon required to execute inference on physical robots. While Figure commits 100,000 Vera Rubin GPUs in Texas to train its Helix neural models, the physical humanoid robots operating on factory floors execute optimized, quantized inference across low-power on-board edge chips consuming less than 150 watts.
2. **Evaluate Specialized Neocloud Infrastructure Partners:** Organizations training proprietary physical AI or high-density simulation models should assess specialized neocloud operators (including Nscale, CoreWeave, and Lambda) alongside traditional hyperscalers. Neoclouds frequently offer superior direct-to-chip liquid cooling densities, dedicated non-blocking InfiniBand fabrics, and faster cluster provisioning timelines without the overhead of enterprise cloud software tiers.
3. **Secure Long-Range Silicon Allocation Commitments:** Because global frontier GPU manufacturing remains constrained by advanced packaging capacity (CoWoS) and high-bandwidth memory supplies, capital allocation commitments must be executed twelve to twenty-four months in advance. Enterprise AI programs targeting 2027 and 2028 operational rollouts must engage hardware vendors early to secure next-generation silicon allocations.
4. **Prepare Industrial Facilities for Embodied Agent Telemetry:** Enterprise logistics centers and manufacturing facilities planning to pilot humanoid robots must upgrade edge network infrastructure. Humanoid robots operating on factory floors require high-density Wi-Fi 7 or private 5G network coverage to stream diagnostic telemetry, operational video buffers, and human supervisor overrides back to central fleet coordination servers.

  Next Steps 

1. **Assess Operational Feasibility for Humanoid Robotics Pilots:** Audit corporate manufacturing and distribution logistics operations to identify structured, repetitive material handling tasks that can be automated by initial commercial deployments of humanoid systems like Figure 02 and Figure 03.
2. **Benchmark High-Density Simulation and Training Requirements:** Evaluate your organization's internal computer vision, physics simulation, and autonomous robotics pipelines. Determine whether existing on-premises or hyperscaler GPU clusters provide sufficient memory bandwidth to support modern visual-spatial transformer architectures.
3. **Formulate Long-Term AI Infrastructure Capital Projections:** Coordinate with corporate treasury and technology procurement teams to develop five-year compute capital models. Analyze the comparative cost advantages of multi-year neocloud reservation contracts versus hyperscaler cloud instances to support sustained enterprise artificial intelligence development.