Figure Committed $3.5 Billion for Compute After Raising $1.9 Billion

Figure announced on September 3 that it has signed a strategic partnership with Nscale to deploy the NVIDIA Vera Rubin platform across up to 100,000 GPUs, with an initial compute commitment of $3.5 billion and stated intent to scale the arrangement past $6 billion. Deployment targets Barstow, Texas, beginning in the second half of 2027. Nscale is also making an undisclosed strategic investment in Figure, and the two say they will explore scaling Nscale's supply chain using humanoids.

Forbes put the obvious question in its headline: Figure has raised roughly $1.9 billion in its history. It just committed $3.5 billion.

What the structure actually implies

A commitment that exceeds lifetime fundraising by 84% is not a purchase order. It is a forward contract that presumes future capital, and the deal's own architecture tells you how everyone involved expects that to work.

Nscale investing in Figure is the tell. Compute providers taking equity in their largest customers has become the standard shape of AI infrastructure financing over the past eighteen months — Nvidia into MediaTek's bond offering, Nvidia's guarantee structures around large buildouts, hyperscalers taking stakes in the labs that rent their capacity. The vendor's return depends on the customer's survival, so the vendor underwrites the customer. It reduces near-term cash requirements for the buyer and converts a receivable into an option for the seller.

It also means the $3.5 billion number is doing less work than it appears to. Multi-year compute commitments of this type are typically staged against delivery, and delivery starts in H2 2027 — nearly two years out, on a platform that has not shipped at volume. Between now and then Figure has to raise, and this announcement is part of how it raises.

That is not a criticism. It is the mechanism. Announcing secured access to 100,000 next-generation GPUs is a materially different fundraising position than announcing a robot demo, and Figure is competing for capital against companies that already have compute locked.

Why a robot company needs a hyperscaler's worth of GPUs

Brett Adcock's quote in the release is the actual thesis: "Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute."

That sentence is a bet, and it is contested. The scaling hypothesis is well-evidenced for language and reasonably evidenced for vision. For robot manipulation policies it is a hypothesis being tested in real time with other people's money.

The case for it: manipulation policies trained end-to-end have improved roughly in line with data and compute, robot data collection has industrialized, and simulation plus real-world fleet data now produces training corpora at scales that were unavailable three years ago. If embodied control follows the same curve language did, whoever holds the most compute in 2028 holds the best robot.

The case against it: robot data is not text. It is expensive, physically bounded, and does not benefit from an internet-sized pre-existing corpus. Manipulation failures are also not graceful — a language model that is 90% right is useful, while a robot that grips correctly 90% of the time drops one item in ten, which for most commercial tasks is worse than useless. Whether more compute fixes that or whether it requires a different approach entirely is genuinely unresolved.

Figure is spending $3.5 billion on the first answer.

The Vera Rubin detail matters

The platform choice is not incidental. Vera Rubin is Nvidia's next-generation rack-scale system, and Nvidia claims up to a 10× reduction in inference token cost and a 4× reduction in GPU count for training mixture-of-experts models versus Blackwell.

For a robotics company those two numbers point in different directions and both are load-bearing. Training efficiency determines how fast Helix improves. Inference cost determines whether a fleet of humanoids is economically viable at all — every robot in the field is an inference workload running continuously, and unit economics for a home or warehouse humanoid collapse if per-robot compute cost does not fall hard.

Committing to a platform two years before deployment locks Figure to Nvidia's roadmap and to Nvidia's delivery schedule. Given that memory and advanced-packaging capacity are the binding constraints across the industry, securing allocation early is defensible. It is also a bet that Vera Rubin ships close to spec and close to schedule, which is not a bet Nvidia has always won.

What this says about the humanoid sector

The robotics funding environment has gone somewhere unusual. Robotics startups have raised over $23 billion in 2026, approaching the whole of 2025. XPeng just closed $900 million at a $6.3 billion valuation for its humanoid unit. Unitree listed publicly. And now a company with $1.9 billion raised is signing a $3.5 billion compute contract targeting delivery in 2027.

The sector has moved from proving robots can move to proving robots can be afforded. Every one of those numbers is a claim about a market that does not yet generate meaningful revenue — humanoids in commercial deployment today number in the low thousands globally, mostly in pilots.

That gap between capital committed and revenue realized is the thing to watch. It is not automatically a bubble; the AI infrastructure buildout ran the same pattern and the demand arrived. But it means the sector's financing now depends on a continuous stream of capability announcements to justify the next round, and Figure has just put a two-year clock on itself.

What to watch

Whether Figure raises. A round at a valuation that supports a $3.5 billion commitment, announced in the next two to three quarters, converts this from a promise into a plan.

Whether Helix improvements track compute. Figure should publish capability curves against training scale. If the scaling story holds for manipulation, this deal is prescient. If it plateaus, $3.5 billion of Vera Rubin is a very expensive way to learn that robot policies need something other than more GPUs.

Whether Nscale's humanoid supply-chain exploration produces anything. That clause is the most interesting sentence in the release and the easiest one to quietly drop.