Nobody Has Published a Humanoid Robot Production Count. That Is the Story.

Start With What Is on the Record

On Tesla’s fourth-quarter 2025 earnings call, held 28 January 2026, Elon Musk said Optimus was “not in usage in our factories in a material way.” That is the chief executive, on a scheduled investor call, describing his own company’s robot.

On the first-quarter call, Musk said Optimus production at Fremont would begin in late July or August. As of mid-July 2026, it had not started. Tesla’s second-quarter delivery report, published 2 July, contained no Optimus numbers at all.

And the load-bearing fact: Tesla has never published an Optimus production count. Not in a delivery report, not in a 10-Q, not on a call. The figures circulating online — including claims that cumulative production has passed 50,000 units — do not originate with Tesla or with any of the companies involved.

That last point is worth stating precisely, because it is unusual. This is not a case of a company publishing an optimistic number that later proves wrong. It is a case of a number entering circulation with no corporate source whatsoever, and then being cited as though it had one.

Who Does Have Documented Deployments

Two companies come out of this comparison better, and neither is the one with the largest audience.

Figure has a confirmed production pilot with BMW — a named customer, a named facility, work on an actual line. Agility Robotics has confirmed Amazon warehouse testing for its Digit robot, and something rarer: a factory. RoboFab in Salem, Oregon is built for 10,000 units of annual capacity.

Read that Agility figure carefully, though, because it is the kind of number that gets misquoted within a week. Ten thousand units of annual capacity is a statement about a building. It is not a statement about how many robots exist, how many shipped, or how many are working in a warehouse tonight. A factory sized for 10,000 units and producing 200 is a perfectly normal early-stage situation, and the public record does not distinguish between those cases.

So the honest summary is narrow: Figure and Agility have documented deployments and named customers. Tesla has stated intentions and an acknowledgement that the robot was not materially in use in its own plants as of January. That is a real difference in evidentiary quality, and it runs opposite to the difference in coverage volume.

Why This Sector Resists Counting

There is a structural reason humanoids are harder to verify than almost any comparable technology, and it is worth understanding rather than treating as evasion.

A car is registered. A phone is sold through channels that report sell-through. A data centre GPU shows up in someone’s capital expenditure. A humanoid robot deployed inside a private facility under a commercial agreement generates no public artefact at all. There is no registry, no regulator collecting deployment data, no unit that has to be titled or licensed. If neither the manufacturer nor the customer chooses to say, there is no independent way to know.

That vacuum is filled by demonstration videos, which are the least informative evidence available. A robot folding laundry in a lab tells you nothing about uptime, cycle time, intervention rate, or whether it can do the task for eight hours without a technician. Those are the numbers that determine whether a humanoid is a product or a prototype, and none of them are published by anyone.

The Question Worth Asking Instead

Unit counts are the wrong metric anyway. The right one is intervention rate — how often a human has to step in — because that is what separates a machine that replaces labour from one that relocates it.

A humanoid requiring a technician every forty minutes has not automated a workstation; it has converted a warehouse job into a robotics-maintenance job, probably at higher cost. A humanoid running an eight-hour shift unattended is a genuinely different economic object. Both look identical in a promotional video, and both would count as one unit in any production tally.

Nobody publishes intervention rates. Until somebody does, deployment counts — even accurate ones — will not tell you what you actually want to know.

What to Watch

• The first company to publish a unit number in a filing. A figure in a 10-Q or an annual report carries liability that a keynote slide does not. Whoever does it first is the one to take seriously.
• Named customers, named sites. Figure/BMW and Agility/Amazon are checkable because both ends are identified. “A Fortune 500 manufacturer” is not evidence.
• Whether Fremont production actually started. Musk named a window of late July or August. Tesla’s next quarterly disclosure either contains an Optimus figure or it does not, and that absence has now been informative three quarters running.

The practical rule for this sector: treat any humanoid deployment figure as unsourced until you can trace it to a company filing, a press release from the manufacturer, or a named customer confirming it. Right now, most of the widely repeated numbers fail all three tests — and the most-cited one fails them completely.