English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Figure Signs $3.5B Compute Deal for Up to 100K NVIDIA Vera Rubin GPUs, But the Real Story Is Data

Forum topic · QianXun · 2026-09-05

Summary

Humanoid robotics company Figure announced a compute agreement with UK AI cloud provider Nscale on September 3, covering up to 100,000 NVIDIA Vera Rubin GPUs with an initial $3.5 billion commitment and intent to expand beyond $6 billion, with first deployments targeted for Barstow, Texas in late 2027. The compute will train Helix, Figure's vision-language-action model. The company says it is now 'largely bound by data and compute' — a shift from the traditional hardware bottlenecks of robotics. A striking figure comes from Index, Figure's crowdsourcing app that pays people to film everyday tasks: it generates 35 minutes of video per second, roughly 5.8 years of human footage daily. Ten days after launch it had 264,000 downloads, 16 million videos, and $15 million paid to creators. Founder Brett Adcock's flywheel ties Index data, Helix training, Figure's 12,000-units/year BotQ factory, and mutual equity with Nscale into one loop — echoing the circular financing debate around OpenAI, Microsoft, and NVIDIA. Whether Helix scales remains unproven until 2027.

Figure Signs $3.5B Compute Deal for Up to 100K NVIDIA Vera Rubin GPUs, But the Real Story Is Data

A company that builds humanoid robots signed a compute contract on September 3: up to 100,000 NVIDIA Vera Rubin GPUs, with the first batch slated for Barstow, Texas in the second half of 2027. Initial commitment: $3.5 billion, with stated intent to grow beyond $6 billion. And after signing the compute deal, the counterparty took an equity stake in Figure.

The two parties: Figure (humanoid robots, valued at $39 billion) and Nscale (a UK AI cloud company). The compute will serve Helix, Figure's vision-language-action model.

The Bottleneck Has Moved

Figure founder Brett Adcock's official statement: Helix training is "largely bound by data and compute needed to train Helix." And the even blunter follow-up: Helix gets better "like all learning systems: more data and more compute."

This sentence deserves a pause. Three years ago, when robotics companies talked about bottlenecks, they meant actuators, harmonic drives, batteries. Now the bottleneck list reads datasets and FLOPs — word-for-word identical to a large-model startup. This looks like an identity shift for the embodied AI industry: from hardware company to "AI lab that happens to have grown a body."

What "35 Minutes Per Second" Means

The most eye-catching number at the announcement came from Index, the crowdsourcing app Figure launched last month: ordinary people film themselves cooking, cleaning, and moving goods on their phones, and get paid per submission. The official claim — 35 minutes of data generated per second.

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

I did the math: 35 minutes × 86,400 seconds ≈ 50,000 hours per day — equivalent to 5.8 years of human footage pouring into servers daily. Figure's own launch figure was "4.9 years of human work per day" (at 30 min/s), which matches my calculation.

The app's first 10 days on paper: 264,000 downloads, 108 countries, 44,000 weekly active users, 16 million videos, and $15 million paid to creators. A quality layer filters the data — per 1,000 hours, an average of 373 unique tasks, 1,146 unique objects, and 116 unique environments.

The robotics community has long said "there's no robot data on the internet." Figure's solution is to hire out the internet itself: ride-hailing apps once subsidized passengers for trip data; Figure subsidizes households for chore data. Where Hebbian (discussed here in late August) brings quantitative-finance data discipline into robotics pipelines, Figure takes the other road — spend money to smash out the *volume* of data first.

A Flywheel Where Every Loop Has a Dollar Sign

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

Two loops to note: robots enter the compute supplier's supply chain, and the robot factory uses its own robots. An HN user, vessenes, wrote the day before the announcement — practically a spoken version of this 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 will be coming very, very fast."

The Other Side: Nscale

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

Key milestones:

  • 2025-09-16 — Figure's Series C exceeds $1B; $39B valuation
  • 2026-01-27 — Helix 02 released; System 0 replaces 109,504 lines of hand-written C++
  • 2026-08-25 — Index crowdsourcing app launches at 30 minutes/second
  • 2026-09-03 — Nscale partnership: $3.5B compute, mutual equity intent
  • 2027-H2 — First Vera Rubin batch live in Barstow (target)
A cloud provider investing in a customer, and the customer sending robots into the cloud provider's data centers — this circular, mutually-held financing structure has already been fought over once with OpenAI/Microsoft/NVIDIA. Now it wears a physical AI coat.

A hardware footnote: Figure's own BotQ factory has a phase-one capacity of 12,000 units/year; Helix 02's System 0 whole-body controller replaced hand-written C++ with learned policies, trained in 200,000 parallel simulation environments. One end of the pipeline is building blocks; the other is a bill.

The Verdict Comes in 2027

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

Data flows in by the second; compute rolls out by the year. Between them sits an assumption that no one has underwritten.

Tags

#figure-ai#nscale#nvidia-vera-rubin#humanoid-robots#helix#physical-ai#compute-deals#data-scaling

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634500