35 Minutes of Data Per Second, $3.5B in Compute: Figure Is Running a Robotics Company Like a Frontier AI Lab

A humanoid robotics company just signed a compute contract. On September 3, Figure agreed to a deal with Nscale, a UK AI cloud provider: up to 100,000 NVIDIA Vera Rubin GPUs, first batch landing in Barstow, Texas in H2 2027, with an initial commitment of $3.5 billion and stated intent to expand beyond $6 billion. And after signing the compute deal, the counterparty also took (or signaled) an equity stake in Figure.

The parties: Figure (humanoid robots, valued at $39 billion) and Nscale (UK AI cloud). The compute is dedicated to Helix—Figure's vision-language-action model.

The Bottleneck Has Moved

Figure founder Brett Adcock's official statement: "largely bound by data and compute needed to train Helix." He added, even more bluntly, that Helix gets better "the same as all learning systems: more data and more compute."

That's worth pausing on. Three years ago, when robotics companies talked about bottlenecks, they meant actuators, harmonic drives, and batteries. Now the bottleneck list reads datasets and FLOPs—word for word what an LLM startup would say. Judgment: this is an identity shift for embodied AI: from hardware company to "an AI lab that happens to have grown a body."

What 35 Minutes Per Second Actually Means

The most striking number at the announcement came from Index, Figure's crowdsource data app released last month: ordinary people film themselves cooking, cleaning, and moving goods on their phones, and get paid per clip. Official figure: 35 minutes of data generated per second.

It sounds like a translation error. I checked the original—it's verbatim. At launch on August 25 it was 30 minutes/second; nine days later it hit 35.

I did the math: 35 minutes × 86,400 seconds = 50,000 hours per day—roughly 5.8 years of human footage flowing into servers daily. Inference: Figure's own launch-time comparison was "4.9 years of human effort per day" (at 30 min/sec), which matches my calculation.

The app's 10-day ledger: 264,000 downloads, 108 countries, 44,000 weekly active users, 16 million videos, $15 million paid out to creators. The data quality layer has been cleaned—per thousand hours of data, there are on average 373 unique tasks, 1,146 unique objects, and 116 unique environments.

The robotics world has long said "there's no robot data on the internet." Figure's answer is to hire out the internet itself: like ride-hailing apps subsidizing passengers for mobility data, it subsidizes homemakers for housework data. Back in late August we discussed Hebbian importing quant-trading data discipline into robotics pipelines—Figure is taking the other route: buying data volume with cash first.

A Flywheel Where Every Link Has a Price Tag

Jensen Huang's framing for the deal: "physical AI flywheel." Roughly:

1. Index crowdsource app — hundreds of thousands of people filming video 2. Data cleaning and labeling 3. Helix training — on Nscale's 100K Vera Rubin GPUs 4. Figure humanoid robots deployed in the real world 5. Nscale's supply chain — robots working inside data centers 6. Loop back into training; plus BotQ factory (12,000 units/year, robots building robots) looping into deployment

Note the two loops: robots entering the compute vendor's supply chain, and robot factories using robots. HN user vessenes wrote this the day before the announcement, and it reads like the verbal version of the diagram: "the first thousand robots will only be used to build the next robot factory... by the time you actually see humanoids in retail stores, they'll be arriving very, very fast."

The Other Side: Nscale

Nscale's 12-month staircase is worth listing: December 2024, $155M Series A → September 2025, $1.1B Series B → October 2025, Microsoft contract for ~200,000 GB300s → March 2026, $2B Series C at a $14.6B valuation → now, an equity tie-up with Figure.

Key timeline:

• 2025-09-16 — Figure raises Series C above $1B at a $39B valuation
• 2026-01-27 — Helix 02 released; System 0 replaces 109,504 lines of hand-written C++
• 2026-08-25 — Index crowdsource app launches at 30 min/sec
• 2026-09-03 — Nscale deal: $3.5B compute, mutual equity intent
• 2027-H2 — first Vera Rubin batch in Barstow, targeted

Cloud vendor investing in its customer, customer sending robots into the cloud vendor's data centers—this circular, cross-holding financing already caused one controversy with OpenAI/Microsoft/NVIDIA. Now it's wearing embodied AI clothing. Judgment.

One hardware footnote: Figure's own BotQ factory has phase-one capacity of 12,000 units/year; Helix 02's System 0 whole-body controller replaced 109,504 lines of hand-written C++ with learned policies, trained across 200,000 parallel simulation environments. At both ends of the pipeline: building blocks on one side, invoices on the other.

The Verdict Comes in 2027

The most important date in the contract is actually H2 2027—the Vera Rubin platform only ships in H2 2026, and Nscale claims to be among the first deployers outside Microsoft. In other words: money signed now, GPUs arriving the year after next, and whether Helix can actually absorb this scaling law won't be known until the flywheel truly spins.

Data pours in by the second; compute rolls out by the year. In between sits an assumption nobody is underwriting.