The two-dollar hour has no denominator

The source is Roland Berger's study Humanoid Robots 2026 — The Convergence Moment for a New Market, published 15 April 2026. Its framing is careful: advances in AI and robotics hardware "could enable" humanoid systems to run at costs of around USD 2 per hour "in the future," and it speaks of "projected operating costs" making a compelling case where labor is scarce and wages high.

None of that is unreasonable. The problem is what happens to a ratio once it leaves the building.

Read the study page and its release closely and two absences stand out. Neither states what the USD 2 includes or excludes: energy, spares, maintenance, software, depreciation, the lease itself. And neither states a utilization assumption: no hours-per-day figure, no shift model, no break-even occupancy rate.

Secondary write-ups supply both. A formulation now in wide circulation holds that the USD 2 covers energy, software, network, maintenance and depreciation while excluding purchase or lease, and that it beats most US warehouse wages "once utilization clears 60 to 70 percent." Neither the breakdown nor the threshold appears in the primary documents. They may be sensible inferences; they are not citations, and they travel as though they were. The same aggregation layer reports the Figure 03 battery as a three-to-four hour pack with inductive coils in the feet, where Figure's own engineering note says 5 hours at peak performance from a 2.3 kWh structural battery on a 2 kW fast charge. Derived numbers are what get planned against.

A cost per hour has a numerator and a denominator. Almost all published analysis models the numerator, which is tractable: watts, wear parts, amortization. The denominator (hours the machine actually works) is the harder number, and today it is capped by chemistry rather than by demand.

Take the published figures on their own terms. Figure states 5 hours at peak performance for the F.03. Agility's Digit, after an efficiency upgrade, is reported at around four hours, with autonomous docking. UBTECH publishes no per-charge runtime for Walker S2 at all; it markets "24/7 continuous operation" and gets there via a dual-battery system and an autonomous hot swap it says completes within three minutes.

Against a ten-hour shift, a five-hour pack means one of three things: a mid-shift stop that is dead time, a second machine to cover the gap, or swap infrastructure. Note what none of those is: a property of the robot. The vendor sells a runtime; the denominator is assembled on your floor.

One industrial humanoid deployment has real published operating numbers: Figure's F.02 at BMW's Spartanburg plant, retired after eleven months. The reported figures are 1,250-plus hours of runtime, 90,000-plus sheet metal parts loaded into a welding fixture, a contribution to more than 30,000 X3 vehicles, accuracy above 99 percent, a 37-second load against a 5 mm tolerance, and an 84-second cycle. Figure deserves credit for publishing wear photographs alongside the wins.

Eleven months of ten-hour weekday shifts is roughly 239 working days, about 2,390 shift-hours. Against that, 1,250 hours is a little over half. But whether those hours are per robot or across the fleet matters enormously, and the record is inconsistent: Figure's own write-up reads as a single deployed robot, while trade coverage describes two. The arithmetic favors more than one. At an 84-second cycle a machine tops out near 43 parts an hour, yet 90,000 parts across 1,250 hours implies about 72, beyond one robot's reach.

So the honest answer is a range, not a rate: effective utilization in the best-documented humanoid deployment in industry sits somewhere between roughly a quarter and roughly half of the nominal shift window, and the public record will not narrow it further. That is no criticism of the pilot, which was a ramp as much as a production run and reached full line deployment only in its tenth month. The one variable that decides whether USD 2 per hour is a floor or a fantasy is the one nobody reports in auditable form.

This is the thesis in arithmetic form. The numerator is the robot vendor's business, and it is improving fast: Figure reported scaling BotQ from one robot a day to one an hour, with over 350 F.03 units delivered as of late April 2026. That competition is legible, crowded, and already priced.

The denominator is somebody else's business, and it is barely contested. Hours worked is a function of task routing, shift scheduling, charge and swap logistics, cell certification, changeover time, and the ability to prove afterward what ran and for how long. That is a framework problem, not a robot problem, and it is where the thesis says the durable value sits, the reason the box beats the body. A robot idle at 25 percent is expensive at any purchase price; the same machine at 80 percent is cheap at a bad one. The factory OS is where that scheduling and evidence layer is meant to live, and the tracker is where we watch the hardware move. Our dated, falsifiable bets sit in the predictions register; this note scores none of them.

Four limits. This is not a claim that USD 2 per hour is wrong. It is a projection, offered as one, and it may well arrive; the objection is to treating a projected ratio as a measured one. Battery swap can genuinely decouple charge time from shift time, so if autonomous swapping works at scale the runtime ceiling stops binding, though UBTECH publishes a capability rather than audited uptime. The utilization range above rests on inferring robot count from cycle-time arithmetic, not on a disclosure; if the count or the cycle figure differs from what was reported, the range moves. And one pilot is not a fleet. What survives is narrow: the headline number for this industry is a ratio whose denominator is unpublished, physically constrained, and set by the layer above the robot.

Load-bearing claims are checked against the primary document where one exists (company engineering notes and product pages for specifications, the study itself for the projection), with independent trade coverage as a second source. Where the record is ambiguous, this note says so rather than picking a number.

Roland Berger, Humanoid robots 2026: the convergence moment (study page and figures) ·Roland Berger, Labor for 2 dollars per hour: humanoid robots as the next trillion-dollar industry ·Consulting.us: humanoid robots could soon be as big as the auto industry ·Figure: F.02 contributed to the production of 30,000 cars at BMW ·Interesting Engineering: Figure humanoid robots retire bruised after 11 months at BMW ·ASSEMBLY: humanoid robots complete trial project at BMW assembly plant ·Figure: F.03 battery development (2.3 kWh, 5 hour run time, 2 kW fast charge) ·Figure: ramping Figure 03 production at BotQ ·UBTECH: Walker S2 autonomous battery swapping ·Robotics 24/7: UBTECH Walker S2 offers autonomous battery swap capabilities ·Humanoids Daily: Agility Robotics upgrades Digit with longer runtime and autonomous docking ·Agility Robotics: new innovations for Digit