There is a number you have seen. In 2026, China will produce 100,000 humanoid robots. It appears in market notes, in conference decks, in the framing of nearly every story about who is winning. It reads like a fact about the world — a count of machines coming off lines this year.
It is a forecast. And the clearest evidence of that is where the number lives.
One sentence, two numbers
The 100,000 figure travels best when it travels alone. In the Deutsche Bank research note where it is quoted, it does not travel alone. The bank's analyst, Iris Zheng, writes it into the same sentence as a second, smaller number — her own:
"China is leading in production, with government officials estimating 100,000 units of humanoid robot production in 2026, significantly higher than our expectation of ~40,000 units."
Read it again. One sentence contains two production forecasts for the same country, the same product, the same year — and they are sixty thousand units apart. The 100,000 is what officials estimate. The ~40,000 is what the analyst quoting them actually expects. She reports the larger number and, in the same breath, disagrees with it by a factor of two and a half.
That gap is the story. Not which number is right — that when the people forecasting a single year's output disagree by 60,000 units inside one sentence, the number has stopped being an observation and become a position.
What a forecast is not
Neither figure is a shipment. Both are estimates of production that has not yet been counted. The 100,000 is an official projection; the ~40,000 is a bank's projection. A projection is a claim about the future, and the Robot Ledger does not price the future. We price machines that have shipped, at prices we can observe and archive.
Our index has one entry condition, stated in our first print and unchanged since: a machine enters when it ships at an observable price — not when it is demoed, announced, or forecast. A production estimate, however authoritative its source, is none of those things. It is a number about units that may exist. The index counts units that do.
So both forecasts sit in the same place: the forecast column of the announced-vs-shipping watch, where we keep the dated record of what the industry has claimed but not yet shipped at a price. The 100,000 is there. The 40,000 is there. What is not there — for either number — is a single humanoid with an observed unit price, produced and sold in 2026, that we could archive and add to the count.
Why the spread matters to a buyer
A 2.5x disagreement is not a rounding difference between analysts. It is the range of the unknown. If a board is sizing a market, planning a supply chain, or pricing a competitor's cost curve off "100,000 units this year," it is building on the top of a range whose own author expects the bottom. The honest input is not either endpoint. It is the spread itself: nobody producing these machines has yet published a number a buyer could underwrite, so the people forecasting the volume differ by more than the smaller forecast's entire size.
This is the same gap the vaporware watch documents one company at a time — a Series A with no price, a demo with no ship date — expressed here as a headline count instead of a single product. The mechanism is identical. The industry publishes the momentum and withholds the number that would let you check it.
What we are counting
We take no position on whether 2026 humanoid production will land nearer 40,000 or 100,000. We have no forecast of our own to offer, and we will not manufacture one to fill the gap. Both numbers are forecasts made by others; we report that they were made, that they disagree, and that neither is a shipment.
The number we are counting is the one both forecasts skip: how many humanoids have actually shipped in 2026 at a price we can observe. That count is not 100,000. It is not 40,000. Today it is still blank — and when the first unit clears the bar, we will publish it with the receipt, and not a machine before it.
The Robot Ledger is published by Pexara AI LLC. Every figure above traces to a dated, checksummed archive entry; figures may be cited with attribution. This piece reports two published forecasts and takes no position on either; it makes no forecast of its own.