One per hour, ten thousand parts

In late April, Figure reported that its BotQ facility had gone from building one Figure 03 per day to one per hour (a 24x throughput improvement in under 120 days), with more than 350 units delivered at the time of writing. The detail matters more than the headline: end-of-line first-pass yield above 80 percent and improving weekly, a battery line at 99.3 percent first-pass yield, more than 9,000 actuators produced across ten-plus SKUs. These are a manufacturing engineer's numbers, disclosed the way an automaker discloses them.

In late June, Shanghai's AgiBot rolled its 15,000th robot off the line, an industrial G2, from a company founded in February 2023. The cadence is the story: the climb from 1,000 to 5,000 cumulative units took about a year; 5,000 to 10,000 took three months. Omdia ranked AgiBot first globally in humanoid shipments for 2025, at 5,168 units and a 39 percent share. Whatever one thinks of any single Chinese unit count, the shape of the curve is consistent across sources.

Tesla ended Model S and X production at Fremont in early May (14- and 11-year runs, respectively) to convert the lines to Optimus. On the April earnings call, Musk said production would begin in late July or August, warned that initial output would be quite slow, and compressed the whole problem into one sentence: the line "will move as fast as the least lucky, slowest, dumbest part in the entire 10,000" parts in the robot.

By the Q2 shareholder update in late July, the July-or-August window was gone, replaced by production "in 2026," with the first builds earmarked for an internal Optimus Academy to generate training data rather than for customers. As of late July, no verified production unit had been disclosed. None of this is a scandal; it is the point. The company with more experience mass-producing complex electromechanical machines than any humanoid startup is finding this the hardest ramp it has attempted, because the supply chain does not exist yet, and every one of those ten thousand parts needs a qualified source.

Read Figure's own account of how its ramp happened and notice what it is not about. It is not about the actuators, and it is not about the AI. It is about qualifying hundreds of suppliers against incoming inspection criteria; custom manufacturing execution software running across more than 150 networked workstations; over 50 in-process inspection points; more than 80 functional verification tests per robot before sign-off, including burn-in runs of squats, presses, and jogging by the thousand.

That list is a framework: supplier qualification, routing, inspection, data capture, traceability. Every item on it transfers to the next product; almost none of it is visible in the product. AgiBot's milestone tells the same story from the other side: the celebrated 15,000th unit was filmed doing quality inspection on someone else's electronics line, but its own production line was the actual news.

BMW ran Figure 02 at Spartanburg for roughly 1,250 operating hours and 90,000-plus handled components, called the pilot a success. Then it selected a different vendor entirely, Hexagon's AEON, for its first German deployment at Leipzig, folding both programs into a new Center of Competence for Physical AI in Production. That is what a buyer does when it believes the capability lives in the integration layer rather than in the machine: it multi-sources the machine and keeps the layer. The competence being centralized (how to qualify a robot for a cell, train it, schedule it, and prove afterward what it did) is precisely the part no robot vendor can ship in a crate.

There is a recursion here worth stating plainly. The humanoid is marketed as flexible production capacity: labor you can buy by the unit. But the industry making humanoids is currently being sorted by its command of conventional production capacity: supplier qualification, first-pass yield, takt time, execution software. The thesis usually argues that the durable layer in this industry is the framework around the robot (standards, data, certification, training, routing) rather than the robot itself. This summer's numbers extend that argument from deployment to manufacture: on both sides of the transaction, whoever owns the framework sets the rate. The tracker follows the hardware; our dated, falsifiable bets sit in the predictions register, and this note scores none of them.

Four limits. Every ramp figure above is company-reported; no third party has audited Figure's cycle time or yields, and a rate demonstrated is not a rate sustained. AgiBot's 15,000 counts robots across its whole product line, not humanoids alone; the humanoid-specific figure is Omdia's 5,168 for 2025. Tesla's status is inferred from shareholder-letter language and an absence of disclosure, and an absence is not a measured zero. And a ramp curve says nothing about demand: robots per hour is a supply fact, robots at work is the number that decides the industry, and nobody publishes that one yet; last week's note is about why.

Load-bearing claims are checked against the primary document where one exists (company production notes and press releases for the ramp figures, the shareholder update for Tesla's guidance), with independent trade coverage as a second source. Where a figure is company-reported and unaudited, the text says so.

Figure: ramping Figure 03 production at BotQ ·Interesting Engineering: Figure claims BotQ can make one humanoid robot per hour ·eWeek: Figure's humanoid robot factory hits a major production milestone ·AGIBOT: 15,000th robot rolls off the production line ·Robotics & Automation News: Agibot reaches new milestone as its 15,000th robot rolls off the line ·eWeek: AGIBOT rolls out 15,000th robot ·Electrek, Tesla Q1 2026 call: Optimus production at Fremont, Model S/X line conversion ·TrendForce: Musk shares Optimus production team photo, says initial output will be slow ·Tech Times: Tesla Optimus production count remains zero as Q2 earnings call looms ·Tesla: Q2 2026 shareholder update ·BMW Group: humanoid robots in production in Germany for the first time; Center of Competence for Physical AI