Written by Martin Kardzhilov
Every financial model for humanoid robotics assumes the asset holds still. A robot whose policy changes with every update does not. That breaks depreciation schedules, lease structures and residual values in ways the sector has not priced.
I spend my week looking at models that forecast when a robot pays for itself. Almost all of them share an assumption nobody states out loud: that the asset is stable. Purchase a machine, depreciate it over a defined life, compare the annual cost to the loaded cost of the labour it displaces, and read off the payback period.
That model works for a forklift. It does not work for a learning machine.
The industrial forecasters are bullish, and the numbers are large enough to justify serious analysis. The global humanoid robot market is estimated at $5.41 billion in 2026 and projected to reach $50.27 billion by 2035 — a trajectory that implies capital will be committed long before the operating data exists to justify it. When capital arrives ahead of evidence, the financial structures get invented first and corrected later. That is the risk I want to name.
Consider what a policy update does to an asset register. A robot running a frozen policy is a known quantity: its throughput, intervention rate and downtime are properties of a fixed system, and its residual value can be modelled like any other machine. A robot whose model is retrained every quarter is a different asset each quarter. Its utility can improve — that is the entire commercial promise — but improvement is not monotonic and is not guaranteed to survive a change in site, task mix or staffing.
Four consequences follow.
First, depreciation. Accounting lives assume a decline in value from wear. A learning robot’s value can rise with data and then fall sharply if a regression ships, if a task mix shifts, or if a competing platform generalises faster. Straight-line depreciation over seven years is a fiction in both directions.
Second, leasing. Most deployment economics in this sector will run through as-a-service structures, because customers want the productivity without the balance-sheet exposure. But someone must hold the residual. If nobody can underwrite what the robot will be worth in year three, the lease either does not get written or gets written with a risk premium that quietly kills the business case.
Third, insurance. Premiums are priced against measurable failure modes. An operator that cannot produce an intervention rate, a handoff latency or a recovery success rate is asking an underwriter to price a distribution they have never observed.
Fourth, and most consequential for the sector: the evidence gap is now a financing constraint, not just an engineering one. The capital is available. What is missing is the measurement that converts capability into an underwritable asset. Deployment data — how often humans intervene, how quickly handoffs resolve, how recoveries perform across operators and shifts — is precisely what a lender, a lessor and an actuary need. It is also what most vendors do not publish.
At Brainlayer this is the argument I make to anyone who will listen to a finance person talk about robots. Human-state modelling is not a research indulgence. It is the measurement layer that turns a demonstration into a depreciable, insurable, financeable asset. Without it, physical AI remains a venture-funded science project with an unusually heavy capex line.
Three things would tell me the market is maturing. Vendors publishing intervention and recovery metrics alongside capability claims. Lease terms stretching beyond three years. And insurance products written against observed operating distributions rather than prototype risk.
Until then, the honest position is that we can price the robot and we cannot price the uncertainty around it. The second number is the one that decides whether this industry gets financed at scale.
