Figure's $3.5B Compute Deal Starts Behind Its Data • The Weighted Average

Nscale and Figure’s September 3 partnership commits an initial $3.5 billion of compute, but its first GPUs are targeted for the second half of 2027. Figure’s own disclosures already show a 16.7% increase in its stated video-ingestion rate between its August Index launch and the partnership announcement—a reminder that collecting more training material and delivering the infrastructure to use it run on different clocks.

Reconstructed on September 7, 2026, from records available by September 6; this weekend edition draws on September 3–6 developments.

The data arrives before the hardware

The agreement is consequential without being immediate. Nscale describes the potential to deploy the NVIDIA Vera Rubin platform with up to 100,000 GPUs, an initial $3.5 billion compute commitment, and an intention to scale beyond $6 billion. Initial deployment is targeted for Barstow, Texas, in the second half of 2027. Those qualifiers belong in the same sentence as the large numbers: potential capacity is not installed capacity, and an intention to expand is not the same obligation as the initial commitment.

The buyer is pursuing a specific constraint. In Figure’s own announcement of the partnership, the company says it is increasingly bound by the data and compute needed to train Helix. It reports its Index pipeline generating 35 minutes of data every second. The earlier Index launch described 30 minutes of uploaded video every second. Using those disclosed rates, (35 ÷ 30 − 1) × 100 produces 16.7%. This is a change between two company-reported observations, not an independently measured growth rate or a forecast of training demand.

That qualification makes the calculation more useful, not less. Incoming video is not automatically usable training data, and usable data is not automatically improved robot performance. Figure describes an intake process that filters material, reviews it for fraud, removes duplicates, rebalances the collection, and adds annotations. A larger flow at the entrance creates work throughout that pipeline before it reaches a training run. Operators planning physical-AI infrastructure should budget for that intermediate work rather than jumping directly from collected footage to GPU count.

The decision this quarter is therefore about sequencing. A robotics company with a growing proprietary data pipeline may need to reserve future training capacity before it can prove every resulting capability. But it should contract in stages that distinguish data preparation, available training resources, and acceptance of the eventual infrastructure. A company with a small or poorly characterized dataset should not copy Figure’s reservation merely because the headline makes scale look inevitable.

The distinction echoes our analysis of Gimlet’s larger financing round and its delivery obligations. A funded infrastructure strategy is a reason to begin diligence, not a substitute for it. Here the diligence must establish how Figure bridges the interval between the data it says it can collect now and the new hardware targeted for next year. The public release does not disclose that bridge in enough detail to price it.

Follow the obligations, not the circle

The supplier is also becoming an investor. Nscale says it will make a strategic investment in Figure and become its preferred compute provider. The amount of that equity investment is not disclosed in the partnership release. The Next Web’s September 6 examination of the deal highlights the resulting relationship: a cloud provider receives a major customer while acquiring a financial interest in that customer’s success. That structure creates alignment and a question about how independently the demand should be assessed.

The public financing record adds context without answering the question. Nscale’s March 9 Series C announcement disclosed a $2 billion round at a $14.6 billion valuation. The initial Figure compute commitment is therefore 1.75 times that financing round: $3.5 billion divided by $2 billion. This compares a multiyear customer commitment with one supplier financing event. It is not a leverage ratio, a measure of cash coverage, or evidence that Nscale must fund the contract entirely from that round.

Nscale has also arranged project financing. Its August 31 release describes approximately $3 billion in delayed-draw loan commitments for Ward County, Texas, and Madison, North Carolina. The facilities support GPU infrastructure and related networking, storage, and cooling. This establishes that delivery is being financed at the project level as well as through corporate equity. It does not establish which dollars fund the Figure agreement, nor does the geographic overlap authorize treating the announcements as a single financing package.

Figure’s own Series C announcement disclosed more than $1 billion in committed capital and named NVIDIA among its investors. Nscale’s March release also names NVIDIA among its funding participants. Those first-party records support the observation that the chip supplier has interests on both sides of the compute relationship. They do not, by themselves, prove artificial demand or improper accounting. A shared investor can help coordinate a technically demanding deployment; the economic test remains whether the customer receives useful service and can support its obligations.

The agreement extends beyond training. The partnership release describes training Figure’s models on Vera Rubin through Nscale, validation in NVIDIA Isaac Sim, and deployment on NVIDIA GPUs inside the robots. It also says the parties will explore using humanoids in Nscale’s supply chain. The last point is an exploration, not a disclosed robot order. Counting it as revenue would convert a possible future use into a completed sale.

For a procurement team, the practical response is to separate the relationships on paper. Ask for the compute delivery schedule, the payment milestones, the service-acceptance criteria, and the remedies if delivery slips. Ask separately how the investment affects governance or commercial terms. The public announcement does not provide those provisions, so this article cannot say whether they are strong or weak. It can say that the headline contract value does not reveal them.

More footage does not settle the robot economics

Figure has a technical case for investing, but it remains a case to test. Its original Helix explanation describes a vision-language-action system that connects perception, language understanding, and learned control. The later Helix 02 announcement extends the stated control scope to the full body, including walking, manipulation, and balance. That progression explains why the company wants broader training material and more compute. It does not quantify the revenue produced by the next increment of either.

Earlier operating results show what useful evidence looks like. In its logistics update, Figure reported 4.05 seconds per package, alongside roughly 95% barcode-orientation success, and attributed improvements to both more demonstration data and changes in the model. Those are company-reported results in a defined task, not a general humanoid productivity rate. They are nevertheless closer to a purchasing decision than a count of uploaded videos, because they describe a repeatable piece of work and a visible acceptance condition.

That is the evidentiary ladder the new investment must climb. First establish the quality and diversity of the accepted dataset. Then measure whether additional training improves performance on held-out tasks. Finally determine whether those improvements survive deployment conditions at a cost a customer will pay. Skipping any rung makes the argument circular: more compute is justified by more data, and more data is justified by the promise of more capable robots.

The 16.7% intake-rate change should not be mistaken for progress through that ladder. The two releases may describe slightly different reporting boundaries—uploaded video in one, generated data in the other—and the underlying measurements are not audited. We use the rates as Figure’s stated picture of its intake, not as proof that useful training examples grew proportionately. A disclosed acceptance rate or a measured improvement per unit of training would be stronger evidence.

There is a credible case for reserving early despite those uncertainties. Infrastructure with a long delivery schedule cannot always be bought only after the workload is fully proven. Figure may be paying for the option to keep training rather than risk a later shortage. Our earlier examination of Anthropic’s premium for scheduled compute capacity explored that trade-off. The principle carries over, but its unit prices do not: Figure’s release does not disclose enough capacity and contract-duration detail to calculate a comparable rate.

The skeptical case is equally concrete. If the useful data pipeline grows more slowly than the uploaded footage, the reserved infrastructure could outrun the training work that justifies it. If deployment slips, the contract could fail to relieve the bottleneck when needed. If smaller models or more efficient training make the workload cheaper, the economic value of the reserved capacity could change. None of these outcomes is established by the announcement; each belongs in a scenario review before making a similar commitment.

Today’s brief on K2 Horizon’s small-model benchmark trade-off is a useful counterweight. Its local coding results are not evidence about humanoid training. They illustrate the discipline of asking what an additional increment of model or compute actually buys. Physical-AI operators need their own version of that measurement, grounded in task performance rather than borrowed from a language-model leaderboard.

Reserve capacity only against a testable milestone

The operating recommendation is selective. Robotics teams with a validated data pipeline and a credible training schedule should start capacity negotiations early, but should not let a long-dated reservation become an unexamined strategic commitment. Teams still discovering which demonstrations improve performance should prioritize that measurement first. Their bottleneck may be data quality, evaluation, or deployment design rather than the amount of compute they can reserve.

The cost should be described with the same restraint. The public agreement supports an initial $3.5 billion commitment and a possible larger expansion. It does not support a cost per robot, a training price per video hour, or a GPU-hour tariff. Producing any of those would require missing inputs about utilization, service duration, workload mix, and payment terms. The right next step is to request those inputs from a supplier, not to divide two impressive headline numbers and call the result a price.

Evaluation and oversight belong in the purchase as well. Today’s account of OpenAI’s wiki acknowledgment and disclosure interval concerns a different class of agent, but its procurement lesson is relevant: a supplier’s promise becomes operational only when someone owns the evidence and the response. A robotics deployment needs an equally clear record of what was tested, what failed, and what changes before it reaches an external customer.

The closing checklist follows the actual bottlenecks:

For robotics infrastructure leads: negotiate future capacity when the training schedule can justify it, and require dated acceptance milestones. Budget for the intervening data-processing and evaluation work; the new reservation does not perform those jobs by itself.
For finance and procurement teams: keep equity investment, project financing, compute purchases, and possible robot deployments in separate analyses. The same companies appearing in several roles does not make the obligations interchangeable.
For model and data teams: report accepted training material and held-out task improvements alongside intake volume. Treat the disclosed 30-to-35-minute change as a pipeline observation, not a claim of equivalent capability growth.
For prospective robot customers: ask for performance on your task and the conditions under which it holds. A larger training contract is not evidence that a particular installation is ready.

Evidence that would change the verdict is specific: delivered capacity against the stated schedule, disclosed utilization, and independently credible task results showing that the new resources improve useful output. Those would turn a coordinated financing and infrastructure story into a demonstrated operating advantage. Until then, Figure’s expanding data stream explains the urgency of the reservation, while the delivery date explains why buyers should resist calling the bottleneck solved.