Figure's $3.5 Billion Compute Bet: Humanoids Become an Infrastructure Business

Two numbers from the past week tell you where the humanoid industry’s center of gravity now sits. The first is $3.5 billion: the initial compute commitment Figure AI announced in a multi-year partnership with AI cloud provider Nscale, with provisions to scale beyond $6 billion, for access to up to 100,000 GPUs built on NVIDIA’s upcoming Vera Rubin architecture. The second is 95 percent: the success rate OpenAI’s GPT-6 Astra achieved on a gross pick-and-place task in independent hardware testing by RoboCurve, zero-shot, on robot arms it had never controlled before.

A robotics company now spends like a frontier AI lab. A frontier LLM now manipulates like a mid-tier robotics stack. The convergence of those two facts is the real story of the week, and it deserves a harder look than either headline got on its own.

The Deal, in Facts

The structure of the Figure-Nscale agreement matters as much as its size. Figure plans to deploy up to 100,000 Vera Rubin-generation GPUs at Nscale’s facility in Barstow, Texas, with hardware rollout targeted to begin in the second half of 2027. The initial financial commitment is $3.5 billion in compute capacity, with contractual room to grow past $6 billion over time. Alongside the cloud allocation, Nscale has made an undisclosed strategic investment in Figure, and the two companies are exploring operational deployments of Figure’s humanoids inside Nscale’s own infrastructure supply chains.

Read that last clause carefully, because it is the most futuristic sentence in the announcement. A compute provider part-owns a robot company whose robots may build and maintain the compute provider’s data centers. The flywheel is not a metaphor here; it is a business model with balance sheets attached.

The public framing leaned on scale. “Figure is entering a phase where we are largely bound by data and compute needed to train Helix,” the company stated, referring to the end-to-end AI architecture that powers its machines across three hardware generations. “To ship a robot into every home, we need a massive amount of compute.” CEO Brett Adcock went further: “Helix becomes more capable the same way every learned system does: with more data and compute. Looking back, this will be a key inflection point to putting a robot into every home.”

Nscale CEO Josh Payne positioned the deal as the logical evolution of cloud demand. “We’ve seen incredible growth with inference and agentic AI, and Figure is pushing the boundaries of AI even further,” he said.

Hardware Is Commoditizing, Says the Company That Just Built Its 1,000th Robot

There is a deliberate irony in Figure’s messaging. This is the company that scaled past its 1,000th Figure 03 build, that spent years refining actuators, modular assembly lines, and whole-body control. Its leadership now argues that hardware manufacturing is rapidly commoditizing, and that the limiting factor in autonomous physical labor has shifted to onboard intelligence and general-purpose reasoning.

Coming from most companies, that claim would be self-serving spin. Coming from Figure, it is a strategy document. The company launched Index, its consumer data engine that crowdsources real-world video from tens of thousands of contributors and currently generates roughly 35 minutes of human interaction footage every second. That firehose of unstructured video is raw material for Helix training, but converting video into robust motor policies requires processing power commensurate with the data volume. The Nscale deal is the industrial response to a data asset the company already owns.

The sequence matters: body first, then data, then compute. Figure spent its first years proving it could manufacture at volume. Then it built the data pipeline. Now it is buying the furnaces. Compare that to OpenAI, which confirmed last week it will “definitely” build a humanoid but is starting from the brain and hiring actuator engineers to work backward. Same destination, opposite order of operations, and both companies have concluded they need to own the entire stack.

The NVIDIA Triangle

Jensen Huang characterized the agreement as the full activation of the “physical AI flywheel,” and the loop he described is worth spelling out because it positions NVIDIA at every station. Figure trains Helix on Vera Rubin clusters operated by Nscale. Physical loco-manipulation policies are validated in NVIDIA Isaac Sim. Optimized policies deploy onto onboard NVIDIA GPUs inside Figure’s robots.

Train on NVIDIA, validate on NVIDIA, deploy on NVIDIA. A week after NVIDIA closed its $12.93 billion acquisition of Hugging Face and its LeRobot ecosystem, the Figure-Nscale deal confirms that the humanoid industry’s compute layer has consolidated to a single vendor relationship, with cloud operators like Nscale acting as the financing layer that converts multi-billion capex into subscription compute. Robot companies no longer need to own data centers; they need to sign deals with the people who do, ideally with equity attached.

This is the same structure that financed the LLM boom, now transplanted into robotics. Microsoft-OpenAI, Amazon-Anthropic, Google’s TPUs for Gemini: compute commitments as the deepest form of strategic alignment. Figure’s version adds a twist the LLM world never had. The customer’s product can eventually work inside the provider’s facilities, robots maintaining the buildings that train the robots.

The 2027 Problem

For all the headline scale, the timeline reveals a long runway. Vera Rubin silicon does not begin rolling out until the second half of 2027, which means Figure’s near-term training runs continue to depend on existing Hopper and Blackwell-class clusters. Procuring 100,000 GPUs in Barstow requires power generation, cooling capacity, and capital allocation on a scale that has strained data center buildouts across the entire tech sector. The deal is a down payment on a capability that is two years from existing.

There is also an open research question baked into the bet: whether scaling compute and passive video ingestion can overcome the deep nuances of contact-rich physical dynamics. Raw parameter count does not always substitute for physical compliance and sensor fidelity. Which brings us to the second number of the week.

95 Percent and the Last Millimeter

RoboCurve, a public benefit corporation building standardized evaluation for physical AI, tested GPT-6 Astra against Anthropic’s Claude Fable 5.1 and Fable 5 on dual I2RT YAM robot arms using its open-source Inspect Robots framework. The models operated as high-level agent policies: three camera feeds plus proprioceptive state in, 6-DoF end-effector targets out, handed to an inverse kinematics solver.

On the gross task, placing a red block in a bowl, Astra succeeded in 19 of 20 trials. Fable 5.1 managed 40 percent. Fable 5 managed 5 percent. The efficiency story was even starker: Astra averaged 2,100 output tokens per trial against Fable 5.1’s 12,900, completed runs in 2.5 minutes against 6.8, and cost $0.94 per completed run at list pricing against $2.12. RoboCurve’s Jay Chooi noted that if current performance curves hold, LLMs could control robotic arms in real time within two to three years.

Then came the precision task: grasp a circular puzzle piece by a slender center knob and seat it into a tight circular recess. Astra completed 2 of 20 trials. Ten percent, identical to Fable 5.1, while Fable 5 failed entirely. Under RoboCurve’s five-stage human-scored rubric, Astra consistently reached stage 3, hovering the piece directly over the groove, then hesitating or misaligning the final press. Sub-millimeter insertion tolerance collapsed three models to the same failure wall.

The benchmark has honest caveats: Astra’s bowl trials ran on a different rig than the Fable runs, the trials were executed two days apart rather than interleaved, and human observers scored stage progression with knowledge of which model was active. The direction of the result survives the caveats. Gross manipulation is becoming a commodity capability of frontier reasoning models, available zero-shot through an API. Precision insertion is not.

Unitree CEO Wang Xingxing has offered the cleanest explanation of why, in his assessment of why humanoids still are not ready for factory scale: digital models operate in lossless vector spaces, while physical manipulation accumulates tiny errors with every movement. Without high-frequency tactile feedback and mechanical compliance, models cannot correct for micro-deviations in the final millimeters of contact. No number of GPUs changes the geometry of that problem by itself.

What Compute Buys, and What It Does Not

Put the two stories side by side and the strategic logic of the Nscale deal comes into focus, along with its blind spot.

Compute buys the first 95 percent. It turns crowd sourced video into training data, scales Helix’s reasoning, and drives the per-task cost of gross manipulation toward the sub-dollar range that consumer robotics economics require. Astra’s $0.94 per completed run is an early glimpse of per-task inference pricing, and it is falling fast. When a reasoning model can pick and place out of the box, the differentiated value migrates to whoever has the fleet, the data, and the training clusters to fine-tune for specific domains.

Compute does not buy the last millimeter. The 10 percent precision result is a hardware and sensing problem: tactile arrays, compliant end effectors, force control loops that run faster than any language model’s decision cycle. Figure’s own engineering debates over tendon-driven versus direct-drive hands acknowledge this. Adcock says hardware is commoditizing; Wang says the physical last mile is unsolved. Both are right, because they are talking about different layers. Gross-motion hardware is commoditizing precisely because compute solved it. Tactile hardware is not, precisely because compute has not.

The uncomfortable implication for the industry’s capex plans: you need both lines on the balance sheet now. A hundred-thousand-GPU cluster and a dexterous hand program. The companies pretending otherwise are either underfunded in brains or underfunded in fingers.

The Market Is Repricing Around Compute Access

The funding headlines this week show the market repricing accordingly. Alibaba-backed Dexmal is seeking a $3 billion valuation in a new round after previous rounds led by Chinese internet and EV capital. Paris-based Wandercraft is targeting €100 million at a €750 million valuation, advised by Goldman Sachs, with Renault planning 350 Calvin robots deployed by 2027. LimX Dynamics has filed confidentially for a Hong Kong IPO targeting up to $300 million, following the Unitree listing.

These are hardware valuations, large by robotics historical standards and modest by AI ones. Figure’s $3.5 billion compute commitment exceeds the entire raise of nearly every humanoid startup in existence. That is the asymmetry now forming: the cost of competing at the frontier of robot intelligence has detached from the cost of building robot bodies. Chinese manufacturers dominate unit shipments, but the training compute question will decide whose robots generalize. American labs dominate model capability, but the tactile bottleneck and manufacturing scale live in Shenzhen and Shanghai’s supply chains.

The humanoid P&L has a new line item, and it is the biggest one. Figure just became the first robot company to sign a frontier-lab-style compute contract, with a cloud provider that took equity and wants robots in its data centers. It will not be the last. The question worth tracking over the next eighteen months is not who signs the next hundred-thousand-GPU deal. It is whether anyone solves the millimeter problem before the compute arrives in Barstow, because the company that solves both stops competing in this market and starts owning it.

Sources: Humanoids Daily (Figure-Nscale partnership), Humanoids Daily (RoboCurve GPT-6 Astra evaluation), RoboCurve / Inspect Robots, RobotsBeat (Wandercraft), RobotsBeat (Dexmal).