The Biped Is the Endpoint: Figure Is Financing Humanoid Intelligence Like a Frontier Model

Most coverage still scores Figure like a hardware company. Which biped walks cleaner. Who ships more units this year. Whether the BMW pilot turns into a fleet. That is the default frame because it is the frame you can film.

Two announcements in ten days argue for a different one. On August 25, Figure took Index out of stealth, a paid app where people film themselves doing ordinary physical work and get paid for the footage (https://www.figure.ai/news/introducing-index). On September 3, it signed with Nscale for an initial $3.5 billion of compute, with stated intent to scale past $6 billion, on up to 100,000 NVIDIA GPUs running the Vera Rubin platform (https://www.figure.ai/news/figure-and-nscale-sign-strategic-partnership). Neither release is about a robot. One is about inputs. The other is about the machine that turns inputs into a policy.

My claim: Figure is building an intelligence factory. The inputs are Creator video and reserved GPUs. The output is Helix, and Helix runs on Figure robots only. The biped is the endpoint. The product is the loop. Read the company that way and the last two weeks stop being separate headlines and become the two halves of one system.

Index is the data half of the loop

Start with the numbers Figure published, because they are more specific than the usual data-moat talk. Index ran in stealth for about four months. In that time it hit 264,000 downloads across 108 countries, with more than 44,000 weekly active Creators uploading over 16 million videos. Figure says footage arrives at roughly 30 minutes of video per second, which it translates to about 4.9 years of human work captured per day. It has paid Creators $15 million so far (https://www.figure.ai/news/introducing-index).

The diversity figures matter more than the volume figures. Per 1,000 hours of accepted footage, Figure reports 373 unique tasks, 1,146 distinct objects, and 116 environments. That is the number an ML lead should care about. A million hours of one warehouse cell is a narrow distribution with a lot of samples. A thousand hours spread across 116 environments is a wide distribution, and wide distributions are what you need if the goal is a robot that can be dropped into a kitchen it has never seen.

And you cannot scrape this off the internet. YouTube is full of humans doing things, but the camera is rarely where the robot's eyes would be, the hands are rarely in frame the whole time, and the task boundaries are wherever the editor put the cut. Embodiment data needs consistent egocentric framing, continuous hands, labeled starts and stops. Figure says it tried outside vendors first and none of them could hit throughput, diversity, and quality at once. So it built the pipeline itself: filtering, fraud review, deduplication, rebalancing, then hierarchical annotation. Five stages between a phone upload and a training sample. That is a factory floor, just for video.

The ImageNet comparison is tempting and half right. ImageNet did for vision what Index is trying to do for manipulation: a corpus large and varied enough that scale starts working. The difference is that ImageNet was public. Index feeds Helix and nothing else. Figure is explicit that the data is exclusive. So this is less a public good and more a private mine, and the mine is expensive to operate. Figure says it will spend over $1 billion on data and compute in the next twelve months, with a stated path to 100x the current pipeline.

One more detail deserves a flag. The Index site already floats a "Book a Creator" concept. That reads like a bridge to robots-as-a-service: today you pay a human to do a task on camera, later you book the robot that learned from the footage. The Creator network is a stopgap for fleet hours Figure does not have yet. I think that is the honest way to describe it, and I think Figure would agree.

Nscale is the compute half, and the structure is the story

The compute deal got the bigger headline, and the headline number is the least interesting part.

Here is the inventory of what was actually signed, as far as the public documents go (https://www.nscale.com/press-releases/nscale-and-figure; https://kelo.com/2026/09/03/ai-cloud-firm-nscale-commits-compute-worth-3-5-billion-for-figures-robotics-ambitions/):

1. A capacity reservation denominated in dollars of compute. $3.5 billion initially, with intent to exceed $6 billion. Neither side disclosed GPU-hours or a price per GPU, so the real unit of the deal is unknown from the outside.
2. Preferred-provider status for Nscale. Figure becomes an anchor customer, which is a volume guarantee for a cloud company that is itself building out.
3. Equity flowing into Figure from Nscale. Nscale is making a strategic investment of undisclosed size and becomes a Figure shareholder. The vendor is co-investing in the customer.
4. A 2H 2027 timeline. Initial deployment is targeted for the second half of 2027 in Barstow, Texas, on the Vera Rubin platform. That means the training Figure does over the next year happens somewhere else. This deal is runway, not today's cluster.
5. An option on the reverse trade. The parties say they will explore putting Figure humanoids into Nscale's own supply chain. The compute customer could become a product the infrastructure company buys.

Jensen Huang described the chain in the announcement: train on Vera Rubin through Nscale, validate in Isaac Sim, deploy on NVIDIA GPUs inside the Figure robot. He called it a physical AI flywheel. Brett Adcock's framing was simpler. Helix scales with data and compute, and Figure is constrained by both.

That last sentence is the whole strategy in twelve words. A company that believes its model is compute-constrained behaves like an AI lab. It reserves capacity years ahead, signs multi-year commitments, and accepts that the chips arrive after the contract does. A company that believes its product is a robot behaves like an OEM. It tools a factory and negotiates actuator pricing. Figure is doing some of the second. But it just spent a lot of contractual energy on the first.

The $1.9 billion raise against the $3.5 billion compute line

This is where the naive read goes wrong, and the reason deserves precision.

Figure's Series C closed September 16, 2025, with more than $1 billion committed at a $39 billion post-money valuation (https://www.figure.ai/news/series-c). Add the $675 million Series B from February 2024 at $2.6 billion, plus earlier rounds, and total equity raised sits somewhere around $1.9 billion, a figure that secondary-market summaries treat as approximate (https://valueaddvc.com/blog/figure-ai-valuation-2026-39b-humanoid-robotics-investors). NVIDIA, Microsoft, and Bezos were among prior backers.

So a company with roughly $1.9 billion raised has committed to $3.5 billion of compute. Run that as household math and it looks like insolvency on a schedule.

Household math is the wrong tool. $3.5 billion of compute is a multi-year obligation provisioned by the vendor, with the vendor also buying stock. Figure did not wire $3.5 billion out of the Series C account on September 3. In AI infrastructure, commitments of this shape are capacity purchases: the cloud provider builds and finances the racks, the customer commits to usage or payment over years, and the equity piece aligns both sides against one of them walking away. Figure still owes money or usage over the life of the deal. That is real, and it is large relative to the cash it has raised. I want to concede that plainly rather than wave it off. A company at this stage taking on an obligation nearly twice its lifetime equity is making a bet with the whole balance sheet.

But the shape of the bet is the tell. Traditional robotics capex is a factory you own and depreciate. This is reserved training capacity for a model that does not yet exist at the target scale, financed partly by the party selling the capacity. That is how hyperscaler-adjacent AI labs finance frontier runs, and not how Boston Dynamics or Fanuc historically financed anything.

One more thing not to do: do not add Index's "over $1 billion in the next twelve months" to the Nscale $3.5 billion and call it $4.5 billion spent. They are different horizons. The Index figure is a near-term operating plan covering data and whatever compute Figure buys before Barstow lights up. The Nscale figure is a multi-year reservation that does not deliver GPUs until late 2027. Some of the Index spend will land on other clouds. Some may eventually count against the Nscale line. Nobody outside Figure knows the overlap.

Where the flywheel closes, and where it does not

Huang used the word flywheel. He sells the bearings, so he is allowed to. The word is earned only if you can name the mechanism, so here it is.

A flywheel stores energy from many small pushes and gives it back as steady rotation. The pushes here are three: more diverse human video into the pipeline, more training FLOPs on that video, more deployed robots running the resulting policy. Each push should make the next one cheaper. Better Helix means robots that succeed at more tasks in more places. More deployed robots mean more on-robot data that does not need a Creator or a phone. More on-robot data means Index becomes a supplement instead of the primary source. At that point the Creator network was scaffolding, and Figure kicks it away or turns it into the booking layer for a service business.

That is the closed-loop version. Figure has not closed it yet, and the open joints deserve naming one by one.

First, the video-to-policy transfer is unproven at the scale Figure is buying for. A phone camera captures what the hand did. It does not capture how hard the hand gripped, what the fingertip felt, or how much the wrist compensated when the object was heavier than it looked. Force and tactile signals are absent from the Index corpus by construction. Huang's chain puts Isaac Sim in the middle for a reason: simulation is where you try to recover the physics the video threw away. Whether that recovery works well enough for contact-rich home tasks is the open research question, and no amount of reserved Rubin capacity settles it.

Second, the compute arrives in 2027. Every Helix improvement between now and then is running on whatever Figure can rent today. The Nscale deal changes the ceiling, not the floor.

Third, the loop is body-tied. Helix ships on Figure robots and nowhere else. Other physical AI efforts are pursuing models that move across bodies, which lets them collect data from many hardware partners. Figure has chosen the opposite: one body, one model, one exclusive corpus. That is a bigger moat if it works and a bigger cliff if the body is wrong.

Which brings up the competitive read. Tesla has factory and fleet data from its own vehicles and its own lines. Chinese humanoid OEMs have unit volume and a supply chain that gets cheaper every quarter. Figure has neither advantage at scale, so it is buying substitutes: global Creator diversity in place of fleet hours, and reserved GPU capacity in place of a mature factory. That is the strategy. Whether it is a good one depends on whether physical generalization actually follows a scaling law the way language did. Figure is betting a great deal that it does. I lean toward believing it, although the tactile gap makes me less sure than the people writing the checks.

Objection docket

The strongest version of the skeptical case, in the reader's voice:

"This is press-release theater. A $3.5 billion 'commitment' with GPUs that show up in late 2027. An equity stake nobody will size. No published Helix benchmark that isolates what Index data added. Figure raised at $39 billion on demos and a BMW pilot, and now it wants credit for a loop that exists only in a diagram. Call me when success rates move on unstructured tasks."

Most of that is fair, and I would rather concede it than argue around it. The compute is late. The equity is opaque. There is no public ablation showing Helix with Index data against Helix without it. Figure's public evidence for its model is still mostly footage, and footage is the least trustworthy evidence in robotics.

But the objection reads the announcements as claims about results. I am reading them as evidence about priorities. A company does not build a five-stage annotation pipeline, pay strangers $15 million to film chores, and sign a multi-year GPU reservation with a vendor taking equity unless it has decided its bottleneck is training, not fabrication. You can learn a company's actual model of its own problem by watching where it spends contractual energy. Figure is spending it on data and compute. That is true whether or not the bet pays.

Scope conditions, so this does not become dogma. The thesis fails if Index footage does not move Helix success rates on tasks the robot has not seen. It fails if Rubin capacity slips past 2027 or gets reallocated to a customer with a clearer revenue line. It fails if the tactile gap turns out to be a wall rather than a sim-bridgeable ditch. Any of those would mean Figure built an expensive intake for a model that cannot use it. I do not think that is the likely outcome. I think it is a live one.

Three numbers, and a door left open

If you want to track this without watching another walking video, watch three things.

Accepted hours in Index, not uploads. The 16 million video count is a top-of-funnel number. What matters is how much survives filtering, fraud review, and deduplication, and how the 373-tasks-per-1,000-hours diversity figure holds as volume grows. Diversity tends to collapse as a paid network scales, because people learn what pays and repeat it. If Figure keeps publishing the per-1,000-hour breakdown and it stays wide, the data half is working.

Helix success rates on unstructured home tasks. Not warehouse totes, not BMW body panels. Kitchens and laundry rooms nobody at Figure has mapped. That is what Index footage is supposed to buy, and it is the only benchmark that tests the claim.

First racks in Barstow. Second half of 2027 is the stated target. Infrastructure timelines slip. When those GPUs come online and Figure starts talking about training runs at a scale it could not previously afford, the compute half moves from contract to cluster.

The harder question I am parking for a later piece: if the loop closes, who owns the Creators. Forty-four thousand people a week are currently teaching a private model how to do their own jobs, at a piece rate, with the footage exclusive to one company. That is a labor arrangement as much as a data pipeline, and the terms of it will matter more once the robots those people trained start booking the same work.

In five years nobody will remember which biped walked best at which conference. They will remember who owned the loop.