Nscale Figure AI Compute Deal Turns Humanoid Robotics Into an Infrastructure Test

Nscale and Figure signed a $3.5 billion compute agreement, tying Figure’s humanoid ambitions to as many as 100,000 next-generation NVIDIA GPUs.

The Nscale Figure AI compute deal is not simply a large cloud purchase. Nscale will also invest in Figure and become its preferred compute provider. Initial deployment is targeted for the second half of 2027 at Nscale’s planned infrastructure in Barstow, Texas.

That structure creates the central tension. Figure says more data and compute will make its Helix robotics models more capable. Yet the announced scale arrives well before the companies have disclosed payment schedules, deployment milestones, or independent evidence connecting larger training runs to commercial robot performance.

The deal also places NVIDIA throughout the technical chain. Figure plans to train models on NVIDIA Vera Rubin systems, validate them with NVIDIA Isaac Sim, and deploy NVIDIA processors inside its robots.

For Figure, the agreement reserves a path toward frontier-model levels of infrastructure. For Nscale, it creates a major robotics customer while adding an equity position in that customer.

The announcement therefore turns humanoid robotics into an infrastructure test. The question is no longer whether a robot can complete a carefully selected demonstration. It is whether data centers, simulation, real-world data, financing, and robot deployments can improve together.

What the Nscale Figure AI Compute Deal Actually Includes

The agreement gives Figure a route to enormous computing capacity, but it does not mean 100,000 GPUs are operating today.

Figure and Nscale announced the multiyear partnership on September 3, 2026. The companies described an initial $3.5 billion compute commitment, with an intention to expand the relationship beyond $6 billion.

The infrastructure could eventually include up to 100,000 GPUs based on NVIDIA’s Vera Rubin platform. Initial systems are targeted to begin deployment during the second half of 2027 in Barstow, Texas.

Those qualifications matter. “Up to” describes a potential ceiling, while “targeted” describes a schedule rather than completed delivery. The announcement does not provide a year-by-year GPU count or a date for reaching the maximum scale.

The companies also did not disclose the agreement’s payment schedule. It remains unclear how much capacity Figure must purchase, when those obligations begin, or which milestones govern later phases.

According to the partnership announcement, Nscale will make a strategic investment in Figure. The size and terms of that investment were not disclosed.

Nscale will also become Figure’s preferred compute provider. That arrangement links the cloud supplier’s commercial returns to Figure’s demand, while its equity stake provides exposure to Figure’s valuation.

The agreement covers more than rented processors. Nscale says its model combines data-center capacity, power, GPU infrastructure, and cloud orchestration. Orchestration is the software layer that schedules and manages computing workloads across many systems.

Figure intends to use that infrastructure for future generations of Helix, its family of models for controlling humanoid robots. Helix connects visual input, language instructions, and physical actions.

The planned workflow spans three stages. Training would run through Nscale’s NVIDIA infrastructure. Simulation and validation would use NVIDIA Isaac Sim, a virtual environment for developing and testing robotic systems.

Trained models would then run on NVIDIA computing hardware inside Figure’s robots. Real-world activity could produce more training data, which would feed later model development.

NVIDIA CEO Jensen Huang called this process a robotics flywheel. The description captures the intended loop, but it does not establish how quickly each cycle improves performance.

The transaction also includes a possible operational link. Figure and Nscale said they will explore using humanoid robots in Nscale’s supply chain.

That proposal remains exploratory. The companies did not announce a deployment count, location, task list, or timetable for robots inside Nscale-related operations.

The most reliable reading is therefore narrower than the headline number suggests. Figure has secured a planned path to large-scale compute, starting no earlier than 2027.

Nscale has secured a significant prospective customer and an ownership interest. NVIDIA has positioned its hardware and software across training, simulation, and robot deployment.

What changed is the scale of Figure’s infrastructure commitment. What has not changed is the need to deliver, finance, and productively use that infrastructure.

Why Figure Says Its Robots Need Frontier-Scale Compute

Figure is betting that humanoid progress now depends as much on model training and diverse data as it does on motors, hands, and manufacturing.

Traditional industrial robots usually operate inside tightly controlled environments. Engineers define their workspaces, movements, tooling, and safety boundaries.

General-purpose humanoids face a broader challenge. They must recognize varied objects, interpret instructions, adjust their balance, and complete long sequences of physical actions.

That complexity produces a large training problem. A useful system needs examples covering different rooms, object shapes, lighting conditions, viewpoints, human preferences, and unexpected interruptions.

Figure says Helix improves through both data and compute. Brett Adcock, Figure’s founder and CEO, described those resources as the primary constraints on further model development.

The compute agreement followed Figure’s launch of Index, a system designed to collect real-world video for robot training. Figure’s Index update said the platform had received more than 16 million uploaded videos by August 25.

Figure also reported more than 264,000 downloads and over 44,000 weekly active users. The company said its pipeline was processing approximately 30 minutes of uploaded video every second.

These figures come from Figure and have not been independently audited. Still, they explain the sequence behind the Nscale robotics compute strategy.

Index attempts to expand the supply of physical-world data. Nscale supplies a future path for processing that data into new model generations.

This differs from training a language model mainly on text and code. Physical data contains motion, timing, depth, contact, force, object state, and environmental variation.

A robot must also translate model output into safe actions. An incorrect text response can frustrate a user, while an incorrect physical movement can damage property or injure someone.

Simulation helps developers test more conditions without placing a robot into every possible situation. Isaac Sim can model scenes, sensors, and robot behavior before software reaches physical hardware.

Simulation still has limits. Real objects deform, slip, break, and behave differently from their virtual representations. Human environments also contain irregular arrangements that are difficult to model fully.

Figure’s proposed flywheel tries to bridge that gap. Human-generated data supports training, simulated environments support validation, and deployed robots generate evidence from real operations.

More compute can expand experiments, model size, and training volume. It cannot automatically repair biased data, weak evaluation, unreliable hardware, or gaps between simulation and reality.

This distinction is central to any Figure AI compute explanation. The agreement provides capacity, not guaranteed intelligence.

The infrastructure matters because Figure wants one learning system to support many tasks. That goal demands more generalization than programming a robot for a single fixed station.

Figure has shown robots performing logistics and household-style tasks. It has also announced commercial work involving BMW and Catalyst Brands.

These examples give Figure real environments for testing. However, public demonstrations and announced partnerships do not reveal fleet-wide reliability, intervention rates, or customer economics.

The Nscale Figure AI compute deal addresses one bottleneck before those commercial metrics are public. It assumes that training capacity will remain a major constraint as Figure collects more data.

That assumption is plausible, but it is still an assumption. The decisive evidence will come from measurable improvements in unfamiliar environments and extended deployments.

The Real Contest Is the Infrastructure Promise Versus Deployment Reality

Figure has reserved the ingredients for a large physical AI system, while the difficult work of converting them into dependable robots remains ahead.

The primary opponent in this story is not another humanoid company. It is the gap between an infrastructure commitment and an operating robotics business.

The announced figures describe potential capacity. They do not describe delivered GPU hours, completed training runs, production robot volume, or customer renewal rates.

Initial deployment is scheduled for the second half of 2027. That leaves the project exposed to construction schedules, power availability, networking installation, cooling systems, chip supply, and software integration.

A cluster approaching 100,000 GPUs is not a simple collection of accelerator cards. It requires high-speed interconnects, storage, power distribution, cooling, monitoring, and failure management.

Training also needs software that can keep those processors productive. Poor data pipelines or inefficient model code can leave expensive infrastructure underused.

Vera Rubin is NVIDIA’s next-generation computing platform following Blackwell. NVIDIA has presented Rubin as a rack-scale system rather than a standalone chip.

That design combines accelerators with CPUs, networking, and other components. It also increases the dependence on coordinated delivery across an entire system.

Nscale must therefore execute on both construction and operations. Figure must develop training workloads valuable enough to justify the reserved capacity.

The data-center plan places the planned Figure deployment at Nscale’s Barstow location. Nscale is also developing infrastructure there for Microsoft.

Multiple large customers can improve a facility’s economic foundation. They can also create questions about capacity allocation, delivery sequencing, and shared infrastructure.

The public announcement does not explain how Figure’s prospective GPUs fit beside other commitments. It also does not specify whether every system would be dedicated exclusively to Figure.

Figure faces a different execution challenge. It must show that larger training runs produce improvements that customers can observe and value.

Model benchmarks alone will not settle that question. Buyers need robots that complete useful work for long periods without frequent human intervention.

Relevant measurements include task completion, operating hours between failures, safety incidents, remote assistance, maintenance requirements, and performance in unfamiliar settings.

Figure has not published a complete set of those measures across a large commercial fleet. That absence is common in an early robotics market, but it limits outside evaluation.

The timeline makes staged validation essential. Figure can improve Helix before 2027, yet the largest promised infrastructure arrives later.

That sequencing gives Figure time to collect more data and refine its models. It also means the announced capacity cannot explain current robot performance.

Nscale carries its own commercial risk. Building infrastructure for a fast-growing customer can generate durable cloud revenue when demand materializes.

However, future commitments become less valuable if deployment slips or customers consume less capacity than expected. Contract structure determines how much risk remains with the supplier.

The companies have not disclosed enough commercial detail to resolve that issue. Outsiders cannot determine how firmly Figure must purchase the entire initial commitment.

This is why the equity investment deserves attention. Nscale is not only selling compute to Figure. It is also becoming a shareholder in the company buying that compute.

Such an arrangement can align incentives. Nscale benefits if Figure grows, while Figure gains a supplier motivated to support its expansion.

It can also complicate demand analysis. Some economic value flows from the supplier to the customer, even as the customer commits to purchasing supplier capacity.

That does not make the transaction artificial. It means readers should separate independently funded demand from strategically supported demand.

The agreement becomes more convincing as physical infrastructure comes online and Figure consumes it for productive training. Until then, the scale remains a promise under construction.

NVIDIA Sits Across the Entire Physical AI Loop

NVIDIA stands to supply the training systems, simulation environment, and onboard computing that connect Figure’s data center to each robot.

The Nscale robotics compute agreement shows how NVIDIA is extending its position beyond training large language models. Humanoid robotics creates another potential market for accelerated computing.

The opportunity includes several layers. Large GPU clusters train robotics models, simulation platforms test them, and smaller onboard systems execute them in physical machines.

NVIDIA participates in each layer. Vera Rubin provides the planned training platform, Isaac Sim supports virtual testing, and NVIDIA processors run inside Figure robots.

This structure gives developers a more integrated path from cloud training to physical deployment. It also concentrates substantial technical dependence within one vendor’s platform.

Figure could benefit from shared software and tooling. Models developed within an aligned environment can move more easily between training, simulation, and deployment.

Yet tighter integration can raise switching costs. Moving workloads to different accelerators may require changes to software, networking, model optimization, and operational procedures.

That tradeoff matters because compute contracts span years. Hardware roadmaps and competitive pricing can change during that period.

NVIDIA’s role also places the deal within a wider infrastructure race. Cloud providers and specialized GPU operators are building capacity for AI labs, enterprises, and model developers.

Nscale competes in that market with established hyperscalers and specialized providers. CoreWeave and Lambda are among the companies pursuing large accelerated-computing workloads.

Amazon, Google, and Microsoft combine cloud infrastructure with internally designed processors. Those chips offer large customers another route to diversify beyond NVIDIA hardware.

Robotics adds a different demand profile. Training may require concentrated clusters, while deployed robots need efficient local inference and continuous data handling.

This could favor providers that coordinate several layers. It could also encourage robotics companies to avoid dependence on one supplier as their fleets expand.

Nscale’s approach is vertically integrated. It combines data centers, power access, computing systems, and a software layer for delivering AI workloads.

The company previously announced major NVIDIA infrastructure agreements involving Microsoft. Its Texas deployment included approximately 104,000 NVIDIA GB300 GPUs for a separate site and customer plan.

That background shows Nscale is pursuing projects at a similar numerical scale. It does not prove that every announced project will arrive on its original schedule.

Data-center construction has become a central constraint across the AI sector. Access to chips alone is insufficient without grid connections, cooling, land, financing, and network capacity.

Figure’s demand adds robotics to this competition for infrastructure. It also raises the bar for other humanoid developers that follow a similar scaling thesis.

Tesla, Apptronik, Agility Robotics, and other companies are pursuing different combinations of hardware, data, simulation, and deployment partnerships.

Their public strategies vary, and disclosed numbers do not support a clean comparison of model-training capacity. The relevant contest is therefore not raw GPU commitments alone.

Competitors can respond through better data efficiency, specialized models, alternative accelerators, or larger real-world deployments. A smaller system can outperform a larger one when its data and training process are better aligned.

Figure’s decision puts one clear marker in the market. It believes access to very large centralized compute will become necessary for general-purpose humanoids.

If that thesis succeeds, robotics developers will increasingly resemble frontier AI laboratories. Their capital requirements will extend well beyond factories and mechanical engineering.

If the thesis falls short, the sector may discover that embodiment, reliability, and deployment data remain harder bottlenecks than training capacity.

The Nscale Figure AI compute deal is therefore also a bet on where robotics progress comes from. NVIDIA is supplying tools for that bet across almost every stage.

What the $3.5 Billion Commitment Does Not Yet Prove

The headline confirms strategic intent, but it does not establish contract firmness, technical returns, or commercial demand for Figure’s robots.

The first uncertainty concerns the agreement itself. The companies described an initial commitment, yet they did not publish contract terms or minimum annual usage.

A multiyear compute commitment is not necessarily an upfront payment. Spending can occur as capacity becomes available and workloads begin running.

The distinction matters because the headline amount exceeds what most robotics companies can support from current operations. Figure’s financing and revenue details remain private.

The reported agreement confirms the multiyear structure. However, the public record does not show whether the full amount is unconditional.

The second uncertainty concerns Nscale’s equity investment. Neither company disclosed its size, valuation, governance rights, or connection to compute payments.

An equity component can support a young customer while creating long-term alignment. It can also shift part of the supplier’s risk into an investment position.

Without the terms, readers cannot calculate the transaction’s net economics. They also cannot determine how much of the commitment depends on future fundraising or robot revenue.

The third uncertainty is technical. Figure says Helix becomes more capable with additional data and compute, which matches a common pattern in machine learning.

Robotics does not always scale as cleanly as language modeling. Physical systems introduce sensor noise, mechanical wear, latency, safety constraints, and unpredictable environments.

A model can improve on training distributions while still failing on rare physical events. These failures matter because robots act in spaces shared with people and valuable equipment.

Figure’s Index project could broaden its data. However, uploaded human video is not automatically equivalent to robot experience.

Human bodies, cameras, reach, strength, and movement differ from a humanoid machine. Training pipelines must translate that material into actions the robot can execute.

Figure says its system can use physical-world data for general robotics. Independent evaluations will need to test how well that transfer works across new objects and environments.

The fourth uncertainty concerns commercial adoption. Announced pilots show customer interest, but pilots do not guarantee larger purchases.

Enterprise buyers will compare humanoids with fixed automation, mobile robots, redesigned workflows, and human labor. Reliability and operating cost will matter more than visual novelty.

Customers also need support, maintenance, safety procedures, and integration with existing systems. These requirements can slow deployment even when the underlying model improves.

The fifth uncertainty is timing. The first Nscale systems for Figure are targeted for the second half of 2027.

Construction, grid access, hardware production, and system qualification must align. A delay in any layer can postpone the training capacity Figure expects.

Neither company has promised that the entire 100,000-GPU ceiling will arrive during that initial period. The phrase describes potential scale, not a single delivery date.

These limitations do not invalidate the agreement. They define the evidence required to judge it.

Investors and customers should avoid treating booked compute as a substitute for robot deployment. Developers should avoid treating processor count as a substitute for measured model progress.

The strongest future validation would connect specific training increases with lower intervention rates and better task performance. That evidence should span multiple settings rather than selected demonstrations.

Until such results appear, the deal remains a high-conviction infrastructure commitment. It is not independent proof that general-purpose humanoids have reached commercial scale.

Three Signals Will Show Whether the Bet Is Working

The next phase should be judged through infrastructure delivery, measurable robot improvement, and commercial usage rather than another larger headline.

The first signal is physical deployment in Barstow. Nscale and Figure are targeting the second half of 2027 for initial systems.

The most useful updates will specify delivered GPU counts, available power, cluster commissioning, and the workloads running on that capacity.

A completed first phase would strengthen confidence in Nscale’s execution. Delays or vague revisions would weaken the announced schedule, especially if no replacement capacity appears.

Readers should distinguish site announcements from energized systems. A data center can be financed or under construction without serving production workloads.

The second signal is measurable Helix performance. Figure needs to connect new data and training resources with results that matter outside controlled demonstrations.

Useful indicators include success rates across unseen tasks, intervention frequency, continuous operating time, recovery from errors, and behavior around people.

Independent testing would carry more weight than company-selected videos. Results from customers would also reveal whether improvements survive unfamiliar workplaces.

Better performance before the Barstow deployment would support Figure’s broader scaling thesis. It would show that current data and compute investments already produce repeatable gains.

Little measurable progress would weaken the assumption that a much larger cluster will solve the remaining problems. The bottleneck might instead lie in embodiment or deployment.

The third signal is repeat commercial use. Figure’s partnerships with BMW and Catalyst Brands provide settings where robots can encounter real operating demands.

The key evidence will be expansion beyond pilots. Fleet size, paid usage, contract renewals, and reduced human supervision would indicate genuine customer value.

These commercial signals also affect the financing story. Recurring robot revenue would make a multiyear infrastructure commitment easier to support.

If deployments remain small, Figure may depend more heavily on external capital. That would make funding conditions an important constraint on compute consumption.

Contract disclosures could clarify the picture sooner. Either company might eventually explain minimum purchases, investment terms, or phased capacity.

Those details would help separate the binding portion of the agreement from its expansion target. They would also reveal how risk is divided between customer and supplier.

For developers, the story highlights a broader shift. Building advanced robotics models increasingly involves managing data, evaluations, simulation, and infrastructure evidence across many systems.

Teams tracking this market need a reliable way to connect announcements with later delivery. A searchable AI knowledge base can help preserve claims, milestones, and source material as the timeline develops.

The same discipline applies to enterprise buyers. Record the promised capability, deployment date, evaluation method, and commercial requirement before comparing vendors.

The Nscale Figure AI compute deal deserves attention because it places an unusually large infrastructure commitment behind humanoid robotics. Its importance now depends on what arrives after the announcement.

Watch Barstow for operating systems, Helix for independently meaningful gains, and Figure’s customers for repeat usage. Those signals will show whether physical AI is becoming a scalable business or accumulating capacity ahead of demand.

The next announcement matters less than the next verified milestone. Which of those three signals would change your view of the humanoid robotics market first?