Here is the thing most people get wrong about the humanoid capital race. They think the story is hype versus reality — the flashy demos against the disappointing shipments. It isn't. The story is disclosed reality versus undisclosed narrative, and the dividing line is whether the SEC forces you to open your books.
Two numbers from the same stretch of early September make the point.
On September 3, Figure committed $3.5 billion to Nscale for compute, with plans to "scale beyond $6 billion," framing this, per Figure, as bound up with the demand for compute at an unprecedented scale to bring a robot into every home. The next day, Agility Robotics' S-4 landed with the SEC and reported $1,781,967 in total net sales for 2025.
The gap between those figures is not a gap in ambition. Both companies are chasing the same trillion-dollar promise. The difference is that one of them had to file, and the other one didn't.
The gap between those figures is not a gap in ambition. Both companies are chasing the same trillion-dollar promise. The difference is that one had to file, the other didn't.
Last week, this newsletter argued that Agility's SPAC deal called for sobriety. This week the S-4 supplies the actual ledger — and it is worse than the call for sobriety implied. Set against Figure's opposite bet, a capex-first strategy that spends billions before shipping meaningful revenue, the pairing exposes what the sector's valuations are really made of.
Start with the audited side, because it is the only side that exists.
Agility booked $1.78 million in net sales for 2025 against a net loss of $138.1 million. Of that revenue, $1,131,860 — roughly 64 percent — came from related parties, meaning entities already invested in Agility's success.
Cost of goods sold was $4.47 million, two and a half times revenue. That produces a gross loss of $2.69 million and a gross margin of about negative 151 percent — which arguably means the company loses more than a dollar and a half in production cost for every dollar it sells.
And the auditors do not hedge. Management concluded that "substantial doubt exists" about the company's ability to continue as a going concern, with existing cash insufficient to fund operations for even one year from issuance.
Then there is the backlog — the number Agility uses to gesture at a future that justifies a $2.5 billion valuation.
The headline is more than $300 million in multi-year Digit v5 orders. The S-4 reveals it is a single contract: 1,000 units on a three-year RaaS term, placed by a customer the filing itself identifies as a related party.
The deal was sweetened with 453 warrants per robot sourced — 453,000 in aggregate — vesting proportionately as robots deploy. In other words, the flagship customer is, per the filing, being paid, partly in equity, to be the flagship customer.
And Agility, per the filing, "is still evaluating the accounting" for the arrangement. The centerpiece of the commercial story is a transaction the company cannot yet decide how to record.
Now hold that structure in your mind and look at Figure, because it is the same financial engineering viewed from the other side of the disclosure wall.
Nscale is committing at least $3.5 billion in compute to Figure — and simultaneously taking an undisclosed equity stake in the customer it is selling to. The infrastructure provider becomes a shareholder in the demand it is booking as revenue.
Wall Street has a word for this. As TechFundingNews reported, some flag these circular arrangements — where a supplier invests in the buyer of its compute — as "a way to make demand look bigger than it is."
It is the same shape as Agility's warrant-sweetened backlog. A related party is paid, in equity, to generate the demand signal that anchors the valuation. On Agility's side, an S-4 forces the words "related party" and "still evaluating the accounting" onto the page. On Figure's side, the terms are simply "undisclosed."
That is the whole trick. Not that one company is honest and the other lying, but that only one of them is required to translate optimism into standardized accounting.
Jensen Huang, per TechFundingNews, described the Figure–Nscale loop as: "training Figure's models on NVIDIA Vera Rubin through Nscale's AI cloud, validating them in NVIDIA Isaac Sim, and deploying them on NVIDIA GPUs." NVIDIA supplies the chips, invests in Nscale, and sits underneath a deal in which Nscale invests in Figure to buy those chips. He calls it a flywheel. A flywheel is also a closed loop that funds its own inputs.
IEEE Spectrum put it more dryly: "Figure is scaling compute so that its robots can... Uh... Have their compute scaled, I guess?"
The circularity is the point, because the input is doing the work the output cannot yet do. The sector raised $8.6 billion in 2026, already 1.8 times all of 2025. That capital funds the inlet — the compute, the GPU campuses, the training runs — while the outlet, robots that pay their own bills, remains theoretical.
Here is why the disclosure asymmetry matters so much. Figure is chasing a $39 billion valuation. Skild AI raised $1.4 billion in January at a valuation above $14 billion.
Neither has filed an S-4. Neither has had to print a gross margin of negative 151 percent, or disclose that its anchor customer is an insider paid in warrants, or write the phrase "substantial doubt exists."
Agility's valuation, at $2.5 billion against $1.78 million in sales, works out to roughly 1,400 times revenue — and that number is only calculable because the SPAC forced the denominator into public view. Figure's $39 billion rests on a denominator no one outside the cap table has seen.
That is the moat. Not data, not scale, not compute. The defensible thing being built is the balance sheet's privacy.
You can see the same gap mirrored in China's A-share market, where a different disclosure regime tells the same tale. Four leading listed body manufacturers carry a combined market value above 60 billion yuan against a first-half attributable net profit of negative 96 million yuan. Wherever the books are open, the narrative and the industry diverge by orders of magnitude. The only question is whether anyone can see the divergence.
So the uncomfortable question the private players never have to answer is simple. If Figure filed an S-4 tomorrow, what would its gross margin be? What fraction of its bookings would be labeled "related party"? Would its auditors reach for "going concern"?
We know what Agility's numbers looked like the moment daylight hit them. The $6 billion compute commitment isn't evidence that Figure has solved the economics Agility hasn't. It is evidence that Figure has not yet been forced to show whether it has.
The data-and-scale moat is a financing structure dressed as an economic one. And the surest sign it's a story rather than a business is that it only holds up as long as nobody has to file.