Economics and Reality Check

12.1 What is actually deployed

Very little, but no longer zero.

Figure is the clearest Western datapoint: Figure 02 at BMW Spartanburg, an eleven-month deployment, 1,250+ hours of runtime on 10-hour weekday shifts, 90,000+ parts loaded, contributing to 30,000+ vehicles, at 84-second cycle time and 5 mm tolerance. Note what that is: one repetitive part-loading station, and 1,250 hours is about seven months of one human's working time.

Agility has moved from three to ten Digits at Toyota Motor Manufacturing Canada, plus GXO, Schaeffler and Mercado Libre, with 65,000+ cumulative operating hours.

UBTech has the largest disclosed order book — Walker S2 orders exceeding ¥800M, roughly 500 units delivered in 2025.

Galbot runs autonomous stores in 30+ Chinese cities and warehouses 24/7.

In applied manipulation the numbers are much larger: Chef's 100M meal servings, Dexterity's 100M autonomous production actions, Ambi's 100M+ packages, Diligent's 1M picks, RightHand's cells at 3-second cycle times.

Tesla is the demo-ware pole: Musk acknowledged on the Q4 2025 call (28 January 2026) that no Optimus robots were doing useful work in Tesla factories.

12.2 The one audited P&L

Unitree's STAR Market listing gave the field its first real humanoid financials: ~$248M 2025 revenue, 59.8% gross margin, 3,701 humanoids produced and 3,551 sold in the first nine months of 2025, raising roughly $610M.

Two numbers from that filing matter more than the revenue.

The cost curve: average selling price fell from $85,000 in 2023 to $25,000 in 2025 — a 71% decline in two years, driven by vertically integrated actuators. Actuators are 40–60% of humanoid bill-of-materials cost by McKinsey's estimate. This is the strongest evidence in the book that the hardware problem is genuinely being solved.

The demand mix: 73.6% research and education, 17.4% demonstrations and displays, 9.01% industrial. The market that exists today is overwhelmingly labs and spectacle. The market everyone is valued on does not yet exist.

12.3 Forecasts, handled carefully

Goldman Sachs, in work led by Jacqueline Du, raised its 2035 humanoid TAM from $6B to $38B and shipments 4× to 1.4M units, with unit costs falling from $50k–$250k to $30k–$150k — a 40% decline versus an expected 15–20%. Important caveat: that report is dated February 2024, not 2026. It is widely cited as if current; it is not.

Morgan Stanley and Bank of America humanoid forecasts are frequently quoted in trade press but could not be verified against primary sources during this book's verification pass. Do not cite specific MS or BofA numbers without independent checking.

For calibration against reality: IEEE Spectrum's Evan Ackerman notes industry projections of 18,000 humanoid units in 2025 against "a small handful of robots in carefully controlled pilot projects." And the IFR puts the entire industrial robot installation market at $16.7B — smaller than several humanoid startups' combined paper valuations.

12.4 China, and the July 2026 rupture

China controls roughly 80–90% of global humanoid shipments, more than half of industrial robot installations, and over 90% of rare-earth magnet refining.

On 28 July 2026 the FCC added mobile robots over 2 kg to its Covered List, barring new foreign-made humanoids, quadrupeds and unmanned ground vehicles from US import (existing certifications unaffected). Evan Beard of Standard Bots called it "one of the strongest technology-security actions in modern U.S. history." Georg Stieler of STM offered the counterpoint: "restrictions can reduce security exposure, but they do not by themselves create a competitive domestic ecosystem."

The practical effect on research is significant and underappreciated: a very large fraction of academic robot-learning work runs on Unitree hardware, and 73.6% of Unitree's humanoid revenue is research and education. A US import restriction on the field's default research platform is a research-capacity story as much as a security one.

12.5 Reliability, and the number nobody reports

Industrial customers expect 99.99% uptime. Digit runs roughly 90 minutes on 9-minute recharges. The best-documented long-horizon autonomy in the field is a four-minute, 61-action sequence and Gemini Robotics 2 tasks "lasting several minutes and involving hundreds of decisions."

Almost nobody in humanoid robotics reports MTBF. Figure's BMW disclosure ("minimal hardware failures" over 1,250 hours) is a rare exception, and it is qualitative. The gap between "minutes without a reset" and "eight hours without an intervention" is the single most consequential unreported number in the field, and until companies publish it, deployment claims cannot be evaluated.

Plus One's Shaun Edwards frames the commercial version: "the conversation has moved from 'Can they do the work?' to 'Can we justify the cost?'" — with humanoids at roughly 1,000 picks/hour in controlled demonstrations against specialised parcel robots at 1,300–3,000.

12.6 Safety standards are not ready

ISO 10218-1:2025 (Edition 3, published February 2025) is the industrial robot safety standard, and it explicitly excludes service robots accessible to the public and consumer products.

ISO/TS 15066:2016 — the power-and-force-limiting specification underpinning collaborative operation — remains a 2016 technical specification, confirmed in 2022 and flagged for revision.

Neither was written for a 1.7 m bipedal machine that becomes a falling mass when you cut power. Boston Dynamics' Matt Powers describes the industry approach: "We're going to start with relatively low-risk deployments, and then expand as we build confidence in our safety systems." The IFR now lists safety and security among its top five 2026 trends, citing the need for "clear liability frameworks."

Home deployment — 1X's NEO — will force this issue first, and it will be forced by an incident rather than by a committee.

12.7 Is it a bubble?

The evidence for: Figure at $39B pre-revenue. Skild at $14B on roughly $30M of company-stated revenue (a ~450× multiple). Walden at $1.1B six months from founding, pre-product. Travis Kalanick's ATOMS at $1.7B. Agility exiting via SPAC — historically a reliable late-cycle signal. Two of the highest-profile 2025–26 startups acquired within twelve months of founding.

The evidence against: the cost curve is real and steep (Unitree's 71% ASP decline in two years). Deployment hours are genuinely accumulating (Agility's 65,000, Figure's 1,250, Dexterity's 100M actions). Model capability is genuinely improving on measured benchmarks, even if smoothly rather than discontinuously. And the labour-shortage demand is structurally real in logistics, food production, welding and eldercare.

This book's reading: the technology is not a bubble; the valuations substantially are. The gap will close by valuations compressing toward the applied-manipulation companies in Chapter 6, which are already doing paid work, rather than by the humanoid companies growing into their prices on the current timeline. Watch for the first humanoid company to publish MTBF and revenue per robot — whoever does that first is either very confident or very desperate, and either way it will reprice the sector.

---