Four Companies Say Humanoid Robots Are “in Production.” They're Measuring Four Different Things.
BMW's 30,000 cars, Tesla's 1,000-plus units, Agility's 100,000 totes and Unitree's 20,000-unit target get cited as one trend. They don't belong on one chart.
Humanoid Robots: The Same Word, Four Different Meters
In September 2026, four separate corporate announcements get cited as evidence for the same claim: humanoid robots have moved out of demo reels and onto factory floors. BMW says Figure's robot helped build more than 30,000 vehicles at its Spartanburg plant. Tesla says over 1,000 Optimus units are running at Fremont. Agility Robotics says its Digit robot has moved past 100,000 totes for the logistics firm GXO. Unitree says it will ship 20,000 humanoid units in 2026, on top of roughly 5,500 the year before. Put the four numbers in one sentence, which is how most coverage handles them, and they read like four points on the same curve.
They are not. Line the claims up and the first thing that stands out is that each one measures a different kind of activity, backed by a different kind of business relationship. One counts finished vehicles a robot helped assemble on a jointly engineered pilot line. One counts robots present on a factory floor doing work its own maker has described, on an earnings call, as training rather than production. One counts a signed service contract with an invoiced monthly fee. One counts units that rolled off an assembly line in China, with no public claim at all about what most of them do next. Humanoid robots being “in production” in 2026 turns out to mean four different things, wearing one marketing word, and anyone budgeting against this market in 2027 is comparing numbers that were never meant to sit on the same chart.
BMW and Figure: A Pilot Wearing Production's Clothes
Figure's deployment at BMW's Spartanburg plant is the most substantial of the four claims, and the easiest to overstate. Starting in 2025, Figure's 02 model spent roughly eleven months on a single job: retrieving sheet-metal parts from bins and positioning them for welding on the X3 line, five days a week, ten-hour shifts, working alongside human colleagues rather than behind a safety cage. Over that period it moved more than 90,000 components, logged about 1,250 operating hours, and touched over 30,000 finished vehicles. That is real, repeated, largely unsupervised work on a live line, not a video clip.
It is also narrower than the “BMW deploys humanoid robots” headline suggests. BMW and Figure engineers co-developed the task specifically for that one station, and in 2026 Figure's newer 03 model took over a different job at the same plant: sequencing parts for just-in-time delivery to assembly workers, with tactile-sensor hands and palm cameras its predecessor lacked. BMW opened a second pilot at its Leipzig plant under a new internal group it calls the Center of Competence for Physical AI in Production. Two pilots, at two plants, after roughly two years of dedicated joint engineering, is a different claim from a robot you could order off a catalogue and drop into a competitor's line next quarter. The 30,000-vehicle figure is genuine and worth taking seriously. It describes one carefully built integration, not a shipping product.
Tesla's Own Word for What Optimus Is Doing
Tesla's number sounds like the most dramatic of the four and is the least verifiable. An estimated 1,000 to 1,200 Optimus Gen 3 units are reportedly operating across Fremont and Giga Texas as of mid-2026, a figure Elon Musk confirmed in January. Tesla has never filed an audited production or deployment count for Optimus in an SEC filing; the 10-Q and 8-K filings that reference the programme describe it as a risk factor and a research and development cost line, without a per-unit output or reliability figure attached.
“Primarily for learning and data collection rather than performing productive tasks.”
That is Musk's own description of what those thousand-plus robots are actually doing. Tesla has sold zero units externally and has no paying customer for Optimus anywhere. The target Musk has stated, up to a million units a year eventually from Fremont, scaling to ten million at Giga Texas, is a production plan, not a deployment figure. Counting an internal training fleet alongside a warehouse that invoices a customer for finished work treats a research programme as though it were a running business, and it is the comparison that does the most damage to any honest reading of “humanoid robots in production.”
Agility and GXO: The One Number With a Contract Behind It
In late 2025, the logistics company GXO signed a multi-year commercial agreement with Agility Robotics for its Digit robot: the industry's first formal commercial deployment, and first Robots-as-a-Service contract, for a humanoid robot. The arrangement followed an earlier pilot phase at the same site. At a GXO-run warehouse in Flowery Branch, Georgia, handling operations for the apparel brand Spanx, Digit picks totes off conveyors and autonomous mobile robots and stacks them for onward handling. By November 2025 it had moved more than 100,000 totes at that single site.
This is the only one of the four numbers with a paying customer, an invoiced fee, and a renewal decision behind it, which makes it the most commercially meaningful of the set and also the narrowest. It is one task, at one site, run by one integrator. GXO has not disclosed how many Digit units are on the floor, and Agility has not said whether the contract will extend to other GXO facilities. A hundred thousand totes is a real, audit-adjacent number describing a single working relationship. It is not yet a market.
Unitree's Number Is a Different Kind of Count Altogether
China's figures work differently again. Unitree says it shipped around 5,500 humanoid units in 2025 (independent estimates from the analyst firm Omdia put the number closer to 4,200) and is targeting 20,000 for 2026. TrendForce forecasts China's entire humanoid sector will ship roughly 62,500 units in 2026, a 94 percent increase on 2025, with Unitree and rival AgiBot together accounting for close to 80 percent of that volume. Morgan Stanley has separately doubled its own 2026 China sales forecast.
These are manufacturing output figures: robots that left a factory, counted the way you would count phones or cars off a line. They are not robots confirmed to be doing a specific job at a customer site. A meaningful share of China's shipped units go to research institutes, universities, entertainment venues and trade-show demonstrations, none of which disclose what task, if any, the robot performs afterwards. Comparing a shipment count to Agility's tote count, or to BMW's finished-vehicle count, puts a supply-side number next to two different demand-side ones and calls the result a trend line.
Four Categories, Not One Trend Line
Laid out next to each other, the four numbers stop contradicting one another, because they were never answering the same question.
| Robot / company | What's counted | Business relationship | Task scope |
|---|---|---|---|
| Figure 02/03 at BMW | Vehicles and parts a robot helped assemble | Co-developed pilot; no disclosed commercial fee | One task per plant, across two plants |
| Tesla Optimus | Units present on the factory floor | Internal R&D fleet; zero external sales | “Learning and data collection,” per Tesla's own description |
| Agility Digit at GXO | Totes moved under contract | Paying customer; multi-year RaaS agreement | One task, one warehouse site |
| Unitree (China) | Units shipped from the factory | Manufacturing output across many buyers | Unknown or mixed: research, demo, unspecified use |
Once separated this way, the “hype versus reality” framing that most 2026 coverage reaches for stops making sense. None of these four numbers actually disproves another. A training fleet with a stated non-production purpose is not evidence against a real commercial contract running elsewhere, and a rising shipment count in China does not confirm that any particular shipped unit is doing useful work. They are four separate facts about four separate businesses, and the confusion is what happens when a market brief, or a headline, collapses them into one.
What's Actually New in 2026, Compared With 2025
The genuine news in 2026 is not that humanoid robots got smarter across the board, though Figure 03's tactile-sensor hands, palm cameras and speech-to-speech audio are real capability gains over its predecessor. The genuine news is that one of the four categories crossed a threshold for the first time: a signed, renewable service contract with a named customer and an invoice attached. That is a different kind of milestone from a larger fleet or a higher shipment target, because it is the first of the four data points that lets anyone outside a vendor's own communications team judge return on investment rather than take a demonstration on faith. Everything else in this set is still either a jointly engineered pilot or a factory that has not shipped a single unit to an outside, paying buyer.
What to Track Before the Next Budget Cycle
For anyone weighing a humanoid robot pilot against a 2027 budget, the useful question is not whether the technology has “arrived.” It is which of the four categories a vendor's headline number actually belongs to. Ask whether the figure describes a co-developed pilot, an internal R&D fleet, a signed service contract, or units that left an assembly line. Ask for the contract structure behind a deployment claim, not just the unit count. Ask what happens to the task if the vendor's own integration engineers, who did most of the early work at both BMW and GXO, are no longer on site. And treat a shipment number, wherever it is reported, as a supply-chain signal rather than proof that any of those units are doing a paying job.
The next year will not settle the argument by producing one clean answer. It will more likely produce a fifth or sixth category, as more vendors move from pilot to contract, and the same collapsing-together of numbers will happen again until buyers start asking which meter they are actually reading.
Frequently asked questions
Related reading
OpenAI Reports 948 Linux Privilege-Escalation Findings. The Number Is a Review Queue, Not Your Exposure.
OpenAI says an unreleased cyber model produced 948 Linux kernel privilege-escalation results for about $1.2M. Here is what that number measures, and what to check on your own fleet.
Android Developer Verification Is Live in Four Countries. It Checks Who You Are, Not What Your App Does.
Google began enforcing Android developer verification on 30 September in Brazil, Indonesia, Singapore and Thailand. The check covers publisher identity for seven stores, not app safety, and ADB bypasses it.
Gartner Says $3.5 Trillion of AI Infrastructure Could Be Impaired. Read Your GPU Contract Before You Read the Headline.
Gartner's forecast of $3.5 trillion in AI infrastructure impairment is a write-down warning for owners. For engineers buying GPU capacity, the real exposure sits in contract terms.