Inference Costs Are Collapsing, Humanoids Aren't: The Bill Comes at Decommissioning

The price of a robot-hour is set at decommissioning

The falling cost of AI inference per million tokens has convinced the market that robotic labor will fall along the same slope. The consensus has the wrong frame. A humanoid's brain costs a fraction each year of what it cost the year before; the body obeys different laws.

The thesis is blunt: the real price of a robot-hour is set at decommissioning, and today that line item sits outside every cost model in the sector.

The figures circulating around industrial pilots, roughly $25 per robot-hour, measure usage: energy, maintenance, amortization share of the BOM. Retiring the machine remains an empty row on the spreadsheet.

This is an accounting regime change, not an operational detail. Once end of life enters the numerator, cost parity with manual labor slides forward. Whoever signs now is buying that difference sight unseen.

Anatomy of a liability: 10,000 to 15,000 parts

A standard humanoid contains 10,000 to 15,000 individual components, organized into 200-500 major sub-assemblies: the census comes from The Robot Report^[1].

Density is the real problem. Every component family carries a different separation procedure, and every procedure carries an hourly specialist cost.

• 20-40 electric motors, each with a precision reducer and gearbox
• 30-50 structural elements in aluminum, carbon fiber or titanium
• 1,000-3,000 specialized fasteners holding the skeleton together
• 40-80 position encoders and 50-200 distinct sensors, lidar and cameras included
• up to 80 memory and storage semiconductor devices

A car breaks down into reasonably homogeneous material streams. A humanoid comes apart piece by piece. Recent actuation modules are sealed, and a sealed module is designed to resist being opened.

Every fastener is worth a minute of human labor or robotic cell time. Multiply by three thousand, then add the kilometers of internal wiring. The bill shows up immediately, before you even touch the battery and the memory.

This profile looks more like surgery than scrapping. The cost that matters is that of a qualified technician, multiplied by the hours of a machine designed to resist disassembly.

The four line items left off the books

The same analysis isolates four areas of end-of-life liability. Each has a cost, and each has a legal owner the typical contract leaves vague.

1. Kinetic data theft. Proprietary navigation maps, biometric logs, facial recognition records, behavioral patterns: all of it lives on the machine's memory media.
2. Stored energy. Lithium packs must be reduced to black mass to recover the elements, or tested one by one for a second life.
3. Trapped pressure. Hydraulic and pneumatic components hold pressure at high velocity and must be discharged by specialists.
4. Material fatigue. Recovering high-performance servomotors pays off for the manufacturer, and carries heavy risk on structural reuse.

The first item is the one the market most underestimates. A retired machine is a mine for industrial espionage: hardware resold with intact media amounts to an open door onto the data of the company that ran that fleet.

The second item burns, literally. Punctured or crushed cells risk thermal runaway, with toxic gas release. Safe treatment requires licensed facilities, and licensed facilities set the price.

Add the four items together and you get a cost structure resembling civil nuclear in miniature: making it safe weighs as much as a substantial share of operating it.

The comparison is strong, and it holds on one precise point. The exit cost is technical, regulated and hard to compress with volume.

Two diverging curves: intelligence falls, the body holds

The inference curve is documented and brutal. Serving a frontier-class model today costs a tiny fraction of what it cost three years ago, with order-of-magnitude annual declines on per-million-token price lists.

The physical disposal curve follows different physics. Specialized labor, hazardous materials logistics, certification: these are items that fall slowly with volume, and in some cases rise with regulation.

Capital confirms it. On 19 September 2026 Tesla's $10.1 billion solar gigafactory in Texas cleared a key tax hurdle, according to pv magazine^[2]. The money is racing toward mass production.

The same holds for compute infrastructure: the generative AI boom is accelerating and transforming the Ethernet switching market, as The Next Platform^[3] documents. The reverse chain, the recovery one, attracts a fraction of that capital.

Two diverging curves produce an accounting surprise. The cost of thinking trends toward zero; the cost of the ending stays where it is.

Cliff event: retirement enters the contract, then the price

The first industrial humanoid fleet reaches end of life all at once, by construction: units bought in the same quarter age in the same quarter.

The jump comes when an insurer demands proof of memory destruction before covering the risk. At that point the take-back clause becomes mandatory for insurance reasons, and the price of retirement enters the leasing rate card.

I place that transition between 2027 and 2029, with the bulk of European and American pilot fleets coming out of their first cycle. The causal mechanism is linear: data liability plus battery pack liability equals insurance requirement.

This shifts the cost parity many take as reached in 2028. With end of life on the books, real parity slides by twelve to eighteen months, and the difference lands on the balance sheet of whoever signed first.

Three categories that change shape by 2030

The first to change is generalist electronics recycling. Operators like Re-Teck, cited in the same analysis, work today on flat, repetitive device streams; a humanoid requires dedicated lines, certified staff and a chain of custody for data.

The second is the sealed actuation module supplier. The integrated module wins on the assembly line and loses at teardown: whoever designs for the seal today will end up competing with whoever designs for rapid opening.

The third is fleet leasing. Whoever can price retirement with a real margin will take market share from competitors who sell the machine and run.

Names and trajectories matter less than the direction: value migrates from the body's manufacturer to whoever controls the full cycle, use and exit included.

What changes for whoever signs now

For a CTO there is one useful question: who owns the memory media at the end of the contract, and who pays for certified destruction.

For procurement the leverage sits in the structure, not the discount. A lease with mandatory take-back shifts the liability onto the manufacturer; an outright purchase leaves it with you, along with 1,000-3,000 fasteners to open.

For venture capital the non-obvious bet sits downstream: robotic teardown, cell diagnostics, certified media destruction. These are small markets today, hitched to an installed base growing fast.

For the chief strategy officer the point is hard. A three-year plan assuming a robot-hour cost with no end of life in it describes a world set to vanish before the mid-course review.

Prediction, horizon, kill signal

Prediction: by 31 December 2027 at least one manufacturer among the top five by humanoid units delivered will publish a standard commercial offering with mandatory end-of-life take-back and certified memory destruction.

Confidence: 65 out of 100. Horizon: 466 days from today.

Kill signal: as of 31 December 2027 the public leasing contracts of the top five manufacturers remain without a mandatory take-back clause and without certified destruction of memory media.

This thesis is inevitable before it is imminent: the liability already exists, and it is waiting for the first fleet to age.