Four companies, four factories, four very different definitions of the word “production.” The demo era ended sometime last year. What replaced it is a bookkeeping argument — and the books do not agree.
For three years, a humanoid robot update meant the same thing: a choreographed demo, a funding round, and a founder posting slow-motion video of a machine picking up a box. Everyone nodded. Everyone moved on.
That format died quietly this year, and something less cinematic took its place. The question is no longer whether these robots can walk. It is how many of them exist, who is holding them, and whether anyone outside the manufacturer has ever taken delivery of one. Those turn out to be three separate questions, and in August 2026 the four leading programs answer them in four incompatible ways.
All four have a gap between announced capacity and verified output — but the gap sits in a different place each time. Figure builds robots and has few customers. Tesla built the line and has not started. 1X took the orders and has not shipped. Atlas ships, but only to its own parent company.
Announced capacity is a press release. Verified output is a number someone can check. Here is the distance between the two, by company.
This is a structural comparison, not a quality ranking. Capacity figures are design targets disclosed by each company and do not represent current output. Unit counts are self-reported; none of the four programs is subject to audited unit disclosure.
Figure spent 2026 doing the least glamorous thing in robotics: manufacturing. In late April the company said BotQ had gone from one Figure 03 per day to one per hour — a 24x throughput improvement inside 120 days — with more than 350 third-generation units delivered off the line.Figure That is not a rounding error against a 12,000-unit annual line capacity, but it is the only number in this article that comes with published first-pass yields attached.
The commercial story is thinner than the manufacturing one, which is the honest tension in Figure’s position. BMW remains the anchor: after a Figure 02 fleet spent roughly eleven months on the X3 body shop line in Spartanburg, Figure 03 arrived in 2026 for a parts-sequencing project in logistics.BMW In May, Figure signed Catalyst Brands — parent of JCPenney, Aéropostale and Brooks Brothers — to put humanoids into the Reno distribution centre’s sorting operation.Catalyst Retail logistics is a genuinely nasty environment: irregular packages, changing light, humans wandering through the workspace. Choosing it as customer number two is either confidence or a very expensive way to find out.
What the coverage tends to skip: Figure has never published a price for the 03 and has no pre-order channel of any kind. Every unit that exists is going to a partner, a pilot or the company’s own floor. For a manufacturer running at one robot an hour, that makes demand — not supply — the constraint worth watching.
Tesla did something no automaker has done: it killed its two oldest car lines to make room for robots. The Model S and Model X lines at Fremont were decommissioned in Q2 2026, ending a fourteen-year run, and the floor is being rebuilt for Optimus. The first-generation line is designed for one million robots a year; a second-generation line at Gigafactory Texas is being designed for ten million.Tesla IR
Then read the Q2 shareholder update carefully, because one word went missing. In Q1, Tesla wrote that the Optimus lines were being installed in anticipation of volume production. In Q2, the same sentence says production. Not volume production. Production, anticipated later this year.Tesla IR The April guidance had been late July or August. That window opened and closed without a confirmed start.
The second detail is where the first robots go. Not to customers, and not to a factory line — to what Tesla now calls the Optimus Academy, an internal programme where fleets practise tasks and generate training data. Revenue beat in the quarter; earnings missed; the stock fell on the session.Transcript
None of this means the bet is wrong. Tesla is the only one of the four that has ever industrialised anything at eight figures a year, and the self-driving data flywheel is a real structural asset. But as of today, the most valuable humanoid programme in the world has published fewer verifiable units than a startup founded in 2022. That is the shape of the thing, and it will either close fast or become the story.
While the others fight over warehouse floors, 1X went straight for the living room. NEO is roughly 167 cm and 30 kg, wrapped in soft knit over a lattice polymer body, and runs at about 22 dB — roughly a refrigerator. It is the only humanoid in this comparison with a published price and a checkout page: $20,000 outright, or $499 a month, against a refundable $200 deposit. Pre-orders opened in October 2025 and Forbes reported 10,000 of them inside five days.Forbes
The Hayward NEO Factory opened on 30 April 2026 — 58,000 square feet, 200-plus staff, vertically integrated down to the tendon-drive motors, ramping toward 10,000 units a year with a San Carlos site to follow.1X And yet: 1X’s own factory page says the NEOs coming off the line are being prioritised for internal testing, and first customer shipments are still written in the future tense.1X
The part that deserves more scrutiny than it gets is Expert Mode. When NEO cannot do a task, a 1X teleoperator can take control through VR — with the owner’s approval — and the robot learns from the session. It is a defensible engineering answer to the data problem, and 1X is open about it. It is also a person, in your house, through a robot’s eyes. Whatever the privacy controls, that is a product category the rest of the field has not had to explain yet.
Boston Dynamics unveiled the production Atlas at CES 2026 on 5 January and said, in the same breath, that every 2026 unit was already spoken for: Hyundai’s Robotics Metaplant Application Center and Google DeepMind, with outside customers from early 2027.Boston Dynamics Specs are the most industrial of the four — 1.9 m, 90 kg, 56 degrees of freedom, joints that rotate past human range, roughly 50 kg peak lift, and a self-swapping battery.
Here is the correction worth making, because a lot of coverage gets it backwards: Atlas is not yet doing production-line work at Hyundai’s Georgia plant. The 2026 fleet goes to RMAC, a training facility Boston Dynamics’ CEO has described as a data factory. Parts sequencing at Metaplant America is scheduled for 2028, with component assembly by 2030.Hyundai Everything before that is rehearsal.
The vertical integration is the real story. Hyundai owns Boston Dynamics, Hyundai Mobis makes the actuators — the single largest share of a humanoid’s bill of materials — and Hyundai Motor Group runs the factories where the robots will work. Atlas is simultaneously the product, the pilot programme and the case study. No competitor has that loop.
Which brings us to the thing nobody had on their 2026 card.
In July, unionised Hyundai workers in Ulsan staged rolling partial strikes — walking off four hours a shift — in what was reported as the auto industry’s first stoppage tied to humanoid deployment.Semafor On 21 August the union escalated to a full eight-hour strike across the Ulsan, Jeonju and Asan plants, the first in a decade.Automotive World
The demand is not compensation for displacement. It is procedural: no robot enters a Hyundai workplace without a labour-management agreement in place first. Hyundai’s position is that Korean robot deployment is not part of the current wage talks at all, and that the 2028 plan covers Metaplant America only.
Read the timing. Atlas has not been assigned a single task at any Korean plant. Not one bolt until 2028, no component assembly until 2030. The union is bargaining over a machine that has not arrived — which is not panic, it is leverage exercised while leverage still exists. Whatever terms settle this dispute will be cited by works councils in automotive plants well beyond Korea. That is a larger 2026 development than any spec sheet in this article.
| Figure 03 | Optimus | 1X NEO | Atlas | |
|---|---|---|---|---|
| Primary target | Factory & logistics | Tesla’s own plants first | Private homes | Heavy industry |
| Production status | 1/hr at BotQ | Lines being installed | Ramping, internal units | In production, allocated |
| Named customers | BMW, Catalyst Brands | None announced | Consumer pre-orders | Hyundai, Google DeepMind |
| Can you buy one? | No public price | No; target end-2027 | Yes — $20K / $499 mo | No; external orders 2027 |
| AI stack | Helix / System 0 | Self-driving stack derived | Redwood + World Model | Gemini Robotics (DeepMind) |
| Stated ceiling | 12K/yr → 100K total | 1M/yr → 10M (Texas) | 10K/yr → 100K by 2027 | 30K/yr by 2028 |
| Open question | Does it generalise to messy retail? | When does a unit count appear? | Does anyone outside 1X get one? | Do the 2028 terms survive labour talks? |
Compiled from company announcements, investor updates and press releases published between January and August 2026. Categories describe stated strategy and disclosed status; they are not performance scores.
Not predictions — the four checkpoints where the gap between announcement and output either closes or widens visibly.
Optimus has never appeared as a line item in a production report. The Q3 2026 update is the next place one could show up. Its presence or absence is the single cleanest signal in the sector.
1X has committed to shipping in 2026 and the calendar is short. An independently documented delivery to a paying household would be the first of its kind anywhere.
Two named customers against a 12,000-unit line is a demand question, not a supply one. A third logo — especially outside automotive and apparel — changes the shape of the ramp.
Whether consent rights over robot deployment end up written into a collective agreement. The clause, if it exists, becomes the template everyone else negotiates against.
Zoom out and the honest comparison is to smartphones circa 2006: expensive, awkward, obviously on a curve. The difference is that in 2006 nobody was striking over the iPhone. Humanoids arrived with their labour politics pre-installed, which is either a sign the technology is being taken seriously or a sign that everyone has learned something since 2006. Possibly both.
On publicly reported figures, Figure AI. The company said in April 2026 that it had built more than 350 Figure 03 units and reached a cycle time of one robot per hour at its BotQ facility. Tesla has never disclosed an Optimus unit count, 1X has not reported customer deliveries, and Boston Dynamics has not published an Atlas production number. All counts are self-reported and not independently audited.
Only the 1X NEO has a published consumer price and an order page: $20,000 for Early Access or $499 per month, against a refundable $200 deposit. As of August 2026 there is no verified delivery of a NEO to a paying customer. Figure has never published a price, Tesla targets consumer availability at the end of 2027, and Atlas is enterprise-only with external orders opening in 2027.
Not as of the Q2 2026 shareholder update. Tesla confirmed the Model S and Model X lines were decommissioned and that first-generation Optimus lines are being installed, with production anticipated later this year. The earlier guidance had been late July or August 2026. Tesla also said initial builds go to an internal programme called the Optimus Academy for training-data collection rather than to customers.
The union is seeking consent rights — a commitment that no humanoid robot enters a Hyundai workplace without a labour-management agreement first — alongside wage and retirement-age demands. Rolling partial strikes ran at Ulsan in July 2026, followed by a full eight-hour stoppage on 21 August. Hyundai states that robot deployment in Korean plants is not part of the current negotiations and that its plan covers Metaplant America from 2028.