An independent magazine about AI at work in construction. Every story starts on a real job site, with the people who used the technology and the results they will answer for.
Figure signs Nscale deal for up to 100,000 NVIDIA GPUs to train its humanoid robots | ConstructionMagazine.ai
Figure signs Nscale deal for up to 100,000 NVIDIA GPUs to train its humanoid robots
The humanoid maker and the AI infrastructure firm plan a Vera Rubin cluster in Barstow, Texas, with first capacity targeted for the second half of 2027. Here is what the deal says and what it leaves out, as of 3 September 2026.
Autoplay is off because reduced motion is preferred. Use the playback controls to play the video.
Figure's announcement video for its strategic partnership with Nscale. Source: Figure.
Figure and Nscale said on 3 September 2026 that they have signed a strategic partnership under which Nscale will supply Figure with up to 100,000 GPUs built on NVIDIA's Vera Rubin platform. The initial deployment is targeted for the second half of 2027 in Barstow, Texas. Figure puts its initial commitment at $3.5 billion of compute and says it intends to scale that to more than $6 billion. Those are vendor figures, and neither company has published audited detail behind them.
Autoplay is off because reduced motion is preferred. Use the playback controls to play the video.
Video published by Figure with the Nscale announcement. Source: Figure.
On its face this is a story about a robotics company buying computing power. It matters to construction for two narrower reasons. The first is the site. A cluster of that size does not appear in a Texas town without a large, power-hungry, schedule-driven build behind it, and the announcement names Barstow as the location. The second is a line near the end of the release, in which the two companies say they will explore scaling Nscale's supply chain with humanoid robots. That is exploratory language, not a deployed job, and this article treats it as such.
What the two companies actually agreed
Still image from Figure's announcement of the Nscale partnership. Source: Figure.
The core of the agreement is capacity. Nscale is an AI infrastructure company that builds and operates GPU data centers. Figure is a humanoid robot developer whose control model, called Helix, is trained on video and sensor data collected from its robots. Under the partnership, Nscale will provide Figure access to as many as 100,000 GPUs on NVIDIA's Vera Rubin architecture, the chip generation that follows Blackwell. The release describes the number as a ceiling rather than a delivered quantity, and the parties have not said how quickly they expect to reach it.
Primary source. GPU count, Vera Rubin platform, Barstow location, H2 2027 target, $3.5B and $6B compute figures, strategic investment, supply chain exploration, Helix data and compute constraint, 35 minutes per second figure, and quotes from Brett Adcock, Jensen Huang, and Josh Payne.
Figure has framed the deal in terms of a constraint. The company says it is bound by two things in training Helix: data and compute. The data side is its own fleet. Figure's number is that its robots produce 35 minutes of training data every second, and that is the figure the company uses to explain why it needs more processing capacity than it can get today. The compute side is what Nscale is meant to solve. Whether the constraint is as binding as Figure says is not something an outside reader can check, but it is at least a coherent explanation for why a robot company would sign a multi-billion-dollar data center agreement.
"Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute. Today we're excited to partner with Nscale to bring the compute required to make this vision a reality." Brett Adcock, Founder and CEO, Figure, 3 September 2026
The financial terms are partial. The $3.5 billion initial commitment and the intent to exceed $6 billion are Figure's own numbers, and they describe compute spend over an unstated period. The release also says Nscale is making a strategic investment in Figure. The amount is not disclosed, and this article does not estimate it. Readers should treat both the compute figures and the investment as unaudited statements by the companies involved.
Where the GPUs will live
The announcement names Barstow, Texas, as the site of the initial deployment, with a target of the second half of 2027. That is the only location the release gives, and it does not describe the facility, its size, its power source, or who will build it. Anyone reading a specific project into this should not. What can be said is that a deployment of this scale will need a data center campus and the utility, civil, and mechanical work that comes with one, and that the clock the companies have set is roughly a year from now.
Texas has become a center for large AI data center construction because of available land, a comparatively fast interconnection process, and a grid operator that has been willing to work with very large loads. Whether Barstow follows that pattern, or whether Nscale intends to lease capacity in a facility that already exists or is already under way, is not stated. That distinction matters for anyone trying to gauge what construction work, if any, the deal generates, and it is the first thing to look for in any follow-up filing.
The timeline is worth reading carefully. A deployment targeted for the second half of 2027 is a target, not a commitment to a completion date. Data center schedules of this class routinely slip on transformer lead times, switchgear, and chip delivery. Vera Rubin itself is a next-generation product, and its production ramp will affect when the racks can be filled regardless of when the building is ready. A builder reading this should assume the date is the earliest the parties hope for, not the latest they will accept.
The flywheel Figure is describing
The most useful thing in the release for understanding Figure's strategy is the loop it draws through NVIDIA's product line. Helix is trained on Vera Rubin GPUs supplied by Nscale. Trained behaviors are then validated in Isaac Sim, NVIDIA's simulation environment for robots, before being deployed onto the NVIDIA GPUs that run inside each Figure robot. The same vendor sits at every stage, which is presumably why NVIDIA's chief executive appears in the announcement alongside the two signing companies.
The NVIDIA loop as Figure describes it
1Train Helix on Vera Rubin GPUs supplied by Nscale
→
2Validate behaviors in NVIDIA Isaac Sim
→
3Deploy on NVIDIA GPUs inside Figure robots
"Humanoid robots extend physical AI into the world designed for people, opening a major new industry. Nscale and Figure have activated the robotics flywheel: training Figure’s models on NVIDIA Vera Rubin through Nscale’s AI cloud, validating them in NVIDIA Isaac Sim, and deploying them on NVIDIA GPUs in Figure’s robots. This is the physical AI flywheel that will accelerate the path from models to robots in the world." Jensen Huang, founder and CEO of NVIDIA, 3 September 2026
For a construction reader, the practical meaning of that loop is about iteration speed. A humanoid that can learn a new task from demonstration data, test it in simulation, and push the result to hardware is a different proposition from one that has to be reprogrammed by an integrator for each job. Figure's argument is that the bottleneck is now training capacity rather than mechanical design. Whether that is true is a question the Barstow cluster is meant to answer, and nobody outside Figure will know the answer for some time.
Money, in the companies' own terms
Compute commitment in billions of dollars, Figure's figures, unaudited
Two numbers anchor the deal. The first is the $3.5 billion Figure says it has committed initially. The second is the more than $6 billion it says it intends to reach. Both describe compute purchasing, not equity, and both come from the vendor. There is no third-party confirmation, no disclosed contract term, and no breakdown between hardware, hosting, and power. The gap between the two figures is an intent, not a signed option, and the release does not say what would trigger the step up.
Nscale's strategic investment in Figure runs the other way, from the infrastructure provider into the customer. The size is not stated. Arrangements of this kind, in which a supplier takes a stake in a customer that has just committed to buy from it, have become common in AI infrastructure. They are not improper, but they do mean that the headline compute figures partly describe money moving in a circle, and readers should weight them accordingly.
"We’re excited to partner with Figure as physical intelligence becomes AI's next frontier. We've seen incredible growth with inference and agentic AI and Figure is pushing the boundaries of AI even further. We're proud to be enabling the future of AI robotics together." Josh Payne, CEO and Founder of Nscale, 3 September 2026
The humanoid line, read literally
The sentence most likely to be over-read is the one about Nscale's supply chain. The companies say they will explore scaling that supply chain with humanoids. In plain terms, Nscale builds and fits out data centers, Figure makes robots that walk and lift, and the two have agreed to look at whether the robots can help. The release does not say where, when, or on what task. It does not say a robot will work on the Barstow site. It does not say any contractor has agreed to anything, and this article does not name one.
That caution is not pedantry. Humanoid deployments announced so far in logistics and manufacturing have been pilot-scale, inside controlled buildings, with the robot maker's own engineers on hand. A data center under construction is an outdoor, changing, mixed-trade environment. Moving a humanoid from a warehouse floor to a live build involves fall protection, exclusion zones, and a safety case that no company has yet published. If Figure and Nscale do put robots on a project, that will be its own announcement, and it will deserve its own scrutiny.
What the exploratory line does tell you is where Figure sees a plausible next customer. Data center construction is currently the largest, best-funded, most schedule-pressured category of building work in the United States. If a humanoid maker wants a proving ground with a patient client and a repeatable task list, an infrastructure partner that is about to build for it is an obvious place to start. That is inference from the wording, and the parties have not confirmed it.
What is still unknown
The release leaves a list of open items. The Barstow facility's ownership, size, and builder are not described. The power source and grid arrangements are not described. The investment amount is not given. The period over which $3.5 billion is spent is not given. The word explore attaches to the humanoid language and nothing firmer. None of these gaps is unusual for a partnership announcement, and none of them should be filled by guesswork.
What should a contractor do with this news before 2027?
Not much yet, and that is the honest answer. The deal is real in the sense that two companies have signed it and a third, NVIDIA, has endorsed it. It is not yet a project in any sense a builder would recognize. The useful moves are watching for the permits and utility filings that would show Barstow moving from target to plan, and reading any follow-on humanoid announcement against a simple standard: a site, a task, and a published safety case. Until those appear, the right response is attention rather than action.
The larger reason to pay attention is that the compute being bought here is the training budget for a machine intended to do physical work. Figure has said in as many words that what limits its robots is data and processing, not hardware. If that is right, then the Barstow cluster is closer to a factory for robot capability than to an ordinary server farm, and the trades that build it may eventually be the ones asked to work alongside what it produces. That is a possibility, not a forecast. As of 3 September 2026 it is the most that can be said.