Robots could draw 363 TWh a year by 2035 — and almost none of it is humanoids

Wood Mackenzie put a number on something the robotics industry has not had to think about: what the machines cost to run, in electricity, at scale. Its report — “Embodied AI: How Robotics Are Accelerating Global Power Demand”, reported on August 11, 2026 — projects robotics electricity demand reaching up to 363 TWh a year by 2035.

The composition is the interesting part. Industrial robots account for 357 TWh of the upper-range figure. Humanoids account for roughly 6 TWh. Against roughly 5 million operational robots today drawing about 78 TWh a year, the analysis has the installed fleet reaching about 16 million units by 2035, growing at around 12% annually.

The humanoid line is small because the base is small

Wood Mackenzie is not modest about humanoid growth — the report projects the humanoid stock compounding at more than 90% a year between 2025 and 2035 to exceed 10 million units, with annual shipments above 4 million. That still produces about 6 TWh, roughly 1.6% of the total, because a humanoid is a small, battery-limited machine that spends much of its duty cycle idle or charging. A welding cell running three shifts is not.

The practical read for anyone sizing a site: the power question for a humanoid deployment is charging infrastructure and peak concurrency, not annual consumption. The power question for an automated factory is annual consumption.

Why it matters

Every grid conversation in AI for the last two years has been about data centres. This is the first serious attempt to model the physical layer as its own load, and the argument that comes with it is the sharper one: power availability is already a constraint on where robots get deployed, not a 2035 problem. Interconnection queues, substation capacity and industrial tariffs decide which greenfield automation projects clear their business case.

That cuts against a common assumption in physical AI, which is that the binding constraints are models, actuators and data. Those are the constraints on capability. Electricity is a constraint on siting — and siting is what determines whether the 16 million units get installed anywhere near the forecast.

One caveat on the numbers: these are Wood Mackenzie’s projections as reported in trade coverage of the report, not observed data, and long-dated fleet forecasts in this category have a poor track record in both directions.