Two documents landed within ninety-six hours of each other that, read together, say more about the state of humanoid robotics than any demo video this year. The first is a Form S-4 registration statement, filed September 4 by Agility Robotics and its SPAC partner Churchill Capital Corp XI, containing the first audited financial statements any American pure-play humanoid developer has ever handed to public markets. The second is a production line in Guangzhou, where on September 7 XPENG officially launched mass production of its IRON humanoid and let a completed robot walk off the line under its own power.
One is a confession of how expensive the frontier still is. The other is a claim that the frontier has already been industrialized. Both cannot be the whole truth, and the gap between them is where the industry’s next two years will be decided.
What the Auditors Signed
The topline of Agility’s S-4 filing, as reported by Humanoids Daily, is a study in how far venture optimism has to fall before it meets GAAP:
| Metric | FY 2024 | FY 2025 |
|---|---|---|
| Net sales, trade | $276,393 | $650,107 |
| Net sales, related parties | $33,908 | $1,131,860 |
| Total net sales | $310,301 | $1,781,967 |
| Cost of goods sold | $462,555 | $4,473,234 |
| Gross profit | ($152,254) | ($2,691,267) |
| R&D expense | $53,367,536 | $91,643,691 |
| SG&A expense | $16,199,517 | $45,841,052 |
| Net loss | ($70,549,619) | ($138,086,332) |
Revenue grew 5.7x year over year, which sounds impressive until you note the base was $310,301, roughly the price of two and a half Digits. The company spent $4.47 million to deliver $1.78 million of product, a gross margin of approximately negative 151 percent. Every robot sold in 2025 cost Agility more than twice what the customer paid for it. R&D nearly doubled to $91.6 million, SG&A almost tripled to $45.8 million, and the net loss landed at $138.1 million against $103.0 million of year-end cash.
That cash position explains the going-concern language, which is not boilerplate here. Management stated that existing cash will not fund operations for a year from issuance and that substantial doubt exists about the company’s ability to continue as a going concern. The bridge is the SPAC itself: approximately $87.9 million already raised through SAFEs in July 2026, plus roughly $620 million in expected gross proceeds at closing, built from $414 million in Churchill’s trust and a $201 million PIPE at $10 per share. The merger carries a $200 million minimum cash condition. The public listing is not a victory lap; it is the funding event that keeps the lights on.
The $300 Million Backlog, Dissected
The number Agility has marketed hardest is its $300 million in committed multi-year orders for the upcoming Digit v5. The footnotes reframe it substantially.
First, the entire figure traces to a single three-year RaaS contract for 1,000 robots, subject to milestones, product features, and specifications. Second, the counterparty is a related party: someone already invested in Agility’s success. Third, the deal includes an equity sweetener. The customer receives 453 common stock purchase warrants per robot sourced, an aggregate of 453,000 warrants vesting as robots deploy. Fourth, the filing concedes the accounting treatment is unresolved; warrants issued to customers in sales arrangements are frequently deducted from the transaction price, meaning booked revenue against the marquee order could come in below $300 million.
The pattern extends through the rest of the revenue base. Of 2025’s $1.78 million in sales, $1.13 million, roughly 64 percent, came from related parties. Agility’s commercial architecture, at nearly every load-bearing point, rests on counterparties who benefit from Agility itself succeeding. That is not fraud, and it is not even unusual for a company at this stage. But it is the single most important fact for anyone pricing the equity after the merger closes, because it converts a “$300 million backlog” headline into “one customer with an equity stake, buying at terms still being negotiated with the auditors.”
Two Margins for Two Audiences
The most instructive section of the S-4 is the unit economics blueprint, because it shows how a humanoid company presents profitability when the actual P&L shows negative 151 percent gross margin.
Under the illustrative RaaS model, a Digit rents for about $8,500 per month plus a one-time deployment fee of roughly $25,000, producing on the order of $535,000 in cumulative revenue over a five-year useful life. Under outright purchase, a customer pays about $200,000 upfront plus a $20,000 deployment fee and $36,000 a year in software and maintenance, totaling roughly $400,000 over five years.
Against that, Agility cites “in excess of 70 percent” margins and a sub-one-year payback. Both numbers are real, and both require their definitions attached. The 70 percent figure is a five-year unit contribution calculation, cumulative revenue minus cumulative cost divided by cumulative revenue, excluding corporate SG&A and R&D entirely. The sub-year payback measures only the time for subscription fees to exceed the bill of materials, and it excludes roughly $15,000 in deployment costs and $15,000 a year in delivery costs per unit. The GAAP gross margin the company actually reported in 2025 was negative 151 percent. Both numbers describe the same robot. One describes what a scaled, efficient Agility could look like in 2031; the other describes what Agility looked like in 2025. An investor who confuses them is buying the wrong company.
The cost trajectory behind the model is genuine progress, whatever the framing: engineering iterations and supplier sourcing have cut the Digit v4 bill of materials to roughly $125,000 per unit, with management describing a path toward $30,000 over time. RoboFab in Salem, Oregon is designed for up to 10,000 units annually, and the S-4 projects roughly 800 Digit v5 deployments in 2027, 7,000 by 2030, and up to 25,000 by 2035.
“Development Stage” Versus “Deployed at Nine Facilities”
The operational record is the strongest part of the document, and the tension between it and the financial statements is stark. Digit v4 is deployed or committed across nine customer facilities, has accumulated more than 65,000 operating hours in real environments, and counts Schaeffler, GXO, Toyota Motor Manufacturing Canada, Amazon, and Mercado Libre among its customers. At GXO alone, Digit units have moved more than 100,000 totes. The deployments have passed OSHA-recognized Nationally Recognized Testing Laboratory field evaluations, which matters more than any benchmark score for the right to operate near human workers.
Then Note 1 of the audited financials states that Digit sales, deployment arrangements, and customer acceleration engagements are “still largely considered to be research and development,” and that Agility “is currently considered to be development stage company.” The company’s accountants and its marketing department are describing two different businesses. Both are accurate. That is precisely the problem the S-4 exists to force into the open, and it is why the document matters beyond Agility.
The bridge between the two descriptions is the Customer Acceleration Program, established in late 2025, which charges participants roughly $500,000 to progress from proof of technology through on-site proof of concept toward commercial RaaS deployment. As of May 2026, four new customers had signed and more than 30 prospects were engaged, though the filing introduces CAP under a risk-factor warning that it may not result in commercial deployments or revenue. Notably, existing deployments predate the program, meaning CAP is still an unproven funnel rather than the engine of the current base.
The Other Road: Guangzhou
While Agility was converting hype into footnotes, XPENG was doing the opposite. On September 7 the automaker launched mass production of IRON on what it calls the world’s first commercial humanoid assembly line, with completed units walking off the line autonomously, no teleoperation, no human assistance.
The production details are the substance, not the walk-off stunt. XPENG built the line to vehicle-grade tolerances with over 80 percent automation across key manufacturing stages, including dedicated automated equipment for dexterous hand assembly, the component every Western rival still assembles largely by hand. More than 85 percent of IRON’s supply chain overlaps with XPENG’s EV business, buying motors, structural alloys, harnesses, and battery integration at automotive volumes that no pure-play robotics startup can access. The program sits behind a $900 million private round at a $6.3 billion valuation, and inside a 110,000-square-meter humanoid manufacturing hub in Guangzhou.
“The robot production lines were created from scratch with no precedent to follow,” CEO He Xiaopeng said at the commissioning. “Today’s step is small, but XPENG is building the production lines for an entirely new product category.”
The production IRON also carries a serious onboard stack: three in-house Turing AI chips delivering an aggregate 2,250 TOPS running physical foundation models locally, a VLA pipeline shared with XPENG’s autonomous driving program after the two data teams merged in Q2 2026, autonomous charging dock location and plug-in, and a multi-modal LLM with persistent memory across a dozen faces. The rollout is deliberately conservative: initial units go to XPENG’s own showrooms and campuses through Q4 2026 as data collection hubs, with external deliveries targeted for 2027 and heavy industrial tasks deferred after early trials exposed durability limits in dexterous hands.
Tesla, Between the Two
The XPENG milestone lands hardest in Fremont. Supply chain reports indicate Tesla has placed initial orders for roughly 5,000 component sets across Chinese actuator and precision roller screw suppliers, working toward an internal target of 10,000 to 20,000 units, after summer debugging of the Optimus production line. But Tesla missed its self-imposed Q1 2026 window to unveil the production-intent Gen 3 design, which remains in final adjustment. XPENG applying the identical automotive playbook, in-house silicon, shared supply chain, end-to-end models, has now moved a current design onto an automated line while Tesla’s next platform stays in the laboratory.
One line in Agility’s S-4 sharpens the same picture from the American side. The competition section names eight rivals: Boston Dynamics, Tesla, Apptronik, Figure, 1X, Sanctuary AI, Hexagon, and NEURA Robotics. Every one is American, Canadian, Swedish, or German. No Unitree, no UBTECH, no AGIBOT, no XPENG appears anywhere in it, a disclosure that tells you more about how Western humanoid companies wish the market were structured than about who is actually building assembly lines.
Two Roads, One Test
Set the two models side by side and the industry’s real dividing line is not East versus West or hardware versus software. It is disclosure versus integration. Agility is submitting to the full discipline of US securities law, with all the uncomfortable footnotes that entails, and betting that public capital markets will fund a company whose own auditors classify its sales as R&D. XPENG is submitting to the discipline of the factory floor, betting that automotive manufacturing economics will produce cheap, reliable humanoids faster than capital markets can price them.
Each model has an obvious failure mode. The SPAC route fails if public investors read Note 1 and the related-party disclosures as reasons to starve a company burning $138 million a year of its capital. The automotive route fails if showroom deployments turn out to generate data too thin to bridge into the unstructured work, homes, construction sites, hospitals, where the actual market lives, a gap XPENG’s own head of robotics has acknowledged on the hardware side.
What to watch is now concrete. The S-4 going effective and the Churchill merger closing, with its $200 million minimum cash condition, will tell us whether public markets will hold humanoid equity at all. Digit v5’s initial release, slated for late 2026, against the 800-unit 2027 deployment target will test whether the RaaS math survives contact with a bill of materials still five times its long-run target. And XPENG’s 2027 external deliveries will test whether an 80 percent automated line producing a walk-off humanoid can convert showroom pilots into contracts.
The first company to publish audited numbers has done the industry a favor, whatever happens next. Every humanoid funding round, every “multi-billion dollar backlog” press release, and every consumer pre-order from here forward will be read against the standard Agility just set: show the footnotes, or expect the market to assume they are worse. Meanwhile the first mass production line is running in Guangzhou, roughly on schedule, at automotive tolerances. The industry now has one company that must tell the truth quarterly and one that must ship robots annually. By the end of 2027 we will know which discipline matters more.