# Robotics is a factory problem

*September 2026*

On Tesla's Q4 2025 earnings call, on January 28, Elon Musk said: ["we are replacing the S, X line in Fremont with a 1 million unit per year line of Optimus."](https://earningscalls.dev/transcripts/tesla-inc_tsla_earnings_call_transcript_2026-01-28) Later in the same call he added a caveat. Asked how many Optimus units were working in Tesla's factories, he said the robot was "still in the R&D phase" and "We wouldn't expect to have any kind of significant Optimus production volume until probably the end of this year."

Put that million next to what humanoid makers actually shipped. Unitree, the largest, [shipped over 5,500 units in 2025](https://www.reuters.com/world/asia-pacific/unitree-plans-shanghai-ipo-testing-interest-humanoid-robots-2026-03-20/), 32.4% of the global humanoid market according to its IPO prospectus. Omdia counted [about 13,000 humanoids shipped worldwide](https://longbridge.com/news/272030450) that year. The pace has picked up: a market tracker, SAG, estimates [about 19,100 shipped in the first half of 2026](https://theroboticsmedia.com/article/unitree-18000-cumulative-bipedal-humanoid-production-g1-star-market-ipo-august-12-2026), more than 97% of them from Chinese makers. The dedicated American humanoid factories are rated in the low five figures. Figure's BotQ first line is ["capable of manufacturing up to 12,000 humanoids per year"](https://www.figure.ai/news/botq). Agility's RoboFab has ["the capability to produce more than 10,000 robots per year"](https://www.agilityrobotics.com/content/opening-robofab-worlds-first-factory-for-humanoid-robots), with "hundreds" planned for its first year. One Fremont line is designed to make about 77 times what the whole world shipped in 2025, and about 26 times the first half of 2026 at an annual rate.

On August 11 a friend sent me a few messages about this, and I think his read is the right one. "A factory that makes 6k can't ramp to 1M," he wrote. "It's almost a different kind of problem." And: "Assuming intelligence in robotics gets solved and ends up like today's models, the bottleneck is in production." Most of the humanoid debate is about whether the robots are smart enough. His point is that this is only the first question. The second one, who can build a million of them, decides who wins, and almost nobody is set up to answer it.

This essay is his thesis with the evidence I could find for and against it. After that comes a layer I added the same night, about how anyone should hold a bet like this when the data can't settle it.

## The argument

The first messages were short (translated from Portuguese):

> Elon said he's converting the Fremont factory to get to 1M Optimus a year. Unitree's factory makes 6k. A factory that makes 6k can't ramp to 1M. It's almost a different kind of problem. Assuming intelligence in robotics gets solved and ends up like today's models, the bottleneck is in production. Anyways, bullish Tesla. Not to mention Elon has Tesla's self-driven miles.

A few minutes later he wrote out the full chain:

1. Robotics will be massive. At the speed we're solving intelligence, solving robotics is just a consequence.
2. Solving robotics can solve the problem of the American dream: paying down debt and generating productivity.
3. Assign a high probability to the US–China conflict escalating.
4. Assume the Chinese won't manage to scale production for the US and Europe.
5. Assume the West will need one winner that takes most of the market.
6. Assume that winner has a high chance of being Tesla.
7. Assume it comes much faster than we expect, through the conversion of a factory and the compression of intelligence.
8. So Tesla is mispriced.

His "6k" is a round number, and the real figures bracket it: Unitree's 5,500+ (Omdia, with its own definitions, counts 4,200), Figure's 12,000 nameplate, Agility's 10,000. The only company I found that discloses a 6,000 figure is UBTECH, and it's a capacity number. Its [2025 annual results](https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0331/2026033102607.pdf) report "an annualized production capacity of over 6,000 full-size embodied intelligent humanoid robots" at the end of the year, and a "sales volume of 1,079 units" for the year. Even at 6k, capacity and output are different numbers. All of them sit two orders of magnitude below a million. "Mispriced Tesla" is my friend's investment frame, and none of this is investment advice.

## A 6k factory is not a small 1M factory

![Announced humanoid capacity against disclosed output, log scale](https://future-seems-so-good.com/blog/assets/robotics-is-a-factory-problem/charts/capacity-vs-output.svg)

A million a year is about 2,740 robots a day, or one every 32 seconds if the line never stops. Electrek, summarizing a paywalled report from The Information, gives the Fremont managers' eventual target as ["about 20,000 a week"](https://electrek.co/2026/09/25/tesla-optimus-production-ramp-hands-ai-generalization-problems/), which is the same number. The same report puts Tesla at "a few dozen Optimus units a week" in Q2 and "several hundred a week in August," with a goal of more than 1,000 a week by the end of the year. The distance left is still a factor of 40 or more.

The reason it's a different kind of problem shows up in how parts get made. Figure's [BotQ post](https://www.figure.ai/news/botq) is the clearest public description. Figure 02 was built with "high complexity, tight tolerance, slow computer numerical control (CNC) machining processes." For Figure 03 they moved to injection molding, die casting, metal injection molding and stamping: "Parts that previously spent over a week on a CNC machine can now be manufactured in under 20 seconds with complex steel molds." The switch "comes with a high capital cost." Below some volume you machine parts and assemble by hand. Above it you commit to tooling, and every mold is a bet that the design won't change. A 6k plant and a 1M plant don't sit on one curve. The second one is built out of decisions the first one never had to make.

The fact that points him at Tesla is in the same data. In 2025 Tesla [produced 1,654,667 vehicles](https://www.sec.gov/Archives/edgar/data/1318605/000162828026000016/exhibit9914.htm). Tesla has already run a ramp to that order of magnitude, in cars. A Hacker News commenter made the same argument in April, replying to someone who doubted the Fremont number: Boston Dynamics ["can make on the order of 25k robots a year though... There is one US company that can scale this kind of manufacturing currently... You don't need them to have the best robot now. Or ever really if they're the first to scale."](https://news.ycombinator.com/item?id=47711621)

Musk gave the counterweight himself, on the same January call:

> Now because it is a completely new supply chain, it's just -- it's -- there's really nothing from the existing supply chain that exists in Optimus... So that means the normal S-curve of manufacturing ramp will be longer for Optimus than it is for products that have at least some portion of an existing supply chain. Like when everything is new, the production rate will be proportionate to the least lucky, least confident part of the entire supply chain. And if there's 10,000 things that need to go right, it's -- it only takes one to be slow to lag that.

Tesla's car capability is real, and it doesn't bring the car supply chain along. In April Musk said that replacing the S and X line in four months would be ["an insanely fast speed,"](https://www.businessinsider.com/elon-musk-tesla-optimus-humanoid-robot-unveiling-date-2026-4) and that he didn't know what Optimus's 2026 production rate would be.

## What ramps actually look like

The historical record is thin, and it's all survivors. I'll come back to why that matters.

**Model 3: the robots were the bottleneck.** Tesla's first goal for the Model 3 was 5,000 a week by the end of 2017. It [delivered fewer than 2,000 in all of 2017](https://www.wired.com/story/tesla-q3-production-numbers/) and hit 5,031 in the final week of June 2018, in an end-of-quarter push. Over the 13 weeks of Q3 the average was 4,095 a week, short of the 6,000 promised by late August. The [Q2 2018 shareholder letter](https://www.sec.gov/Archives/edgar/data/1318605/000156459018018490/tsla-ex991_6.htm) explains what broke. GA3, the main general assembly line, "was designed to work with hundreds of robotic lifters that bring components to the line. Due to the density of the line and the relatively high downtime of the lifters, ramping GA3 became substantially more complicated than we had anticipated." The fix was GA4, built quickly because its "layout and processes... are similar to those of the Model S and X assembly line." The part of the ramp that failed was the automation, and the patch was a simpler line. The same letter says the plan was to reach 10,000 Model 3s a week "sometime next year." Seven years later, Tesla made 1.6 million Model 3s and Ys in a year. Late, and then very large.

**Liberty ships: many yards, one frozen design.** The American wartime cargo ship is the canonical fast ramp. Completions went from [2 in 1941 to 542 in 1942 to 1,294 in 1943](https://www.mathscinotes.com/2018/05/liberty-ship-production-data/), and median build time from keel to completion fell from 225 days for ships laid down in 1941 to 39 days in 1943. The ramp was spread over 16 shipyards, because "the design of the Liberty Ship was very simple, which allowed it to be built by many shipyards." And quality broke at volume: 30% of the fleet eventually had the cracking problem traced to brittle steel and new welding practice. The 600x ramp came from a frozen, simple design copied across many sites, with a defect rate nobody would accept in a robot that works next to people.

**Figure 03: 24x in 120 days, from a small base.** In April, Figure said it had [gone from one robot a day to one an hour](https://www.figure.ai/news/ramping-figure-03-production), "a 24x throughput improvement in under 120 days." End-of-line first-pass yield was "over 80% and improving weekly," and it had built over 9,000 actuators across more than 10 SKUs. On July 23, Brett Adcock posted: ["Proud to share BotQ has manufactured our 1,000th humanoid robot for F.03."](https://www.humanoidsdaily.com/news/a-golden-milestone-figure-manufactures-its-1-000th-figure-03-humanoid) One an hour, around the clock, is about 8,760 a year, close to BotQ's 12,000 nameplate. That's what a real ramp to the 10k class looks like. It's fast and well instrumented, and it's still two orders of magnitude from a million.

Musk put the idea in one line at the Giga Texas opening in 2022: ["The factory is the product."](https://www.statesman.com/story/business/2022/04/08/elon-musk-tesla-future-plans-2022-vision-austin-manufacturing-facility/9511074002/) My friend's thesis is that sentence applied to robots.

## The supply chain is the ramp

I tweeted in March that integrating two different systems is one of the hardest problems in engineering. Musk's "10,000 things" is that problem thousands of times over. Figure says there's no mature supply base to buy from: "Humanoid robots, unlike most other industries, do not have well established supply chains with various tiers of manufacturers building modules of the system." They ended up designing "almost the entire robot from scratch including the actuators, motors, sensors, battery pack, and electronics," and they say their suppliers "can easily scale to 100,000 robots or 3,000,000 actuators in the next four years." That ratio, 30 actuators per robot, is the number to keep in mind. A million robots a year is on the order of 30 million precision actuators a year.

**Actuators and reducers.** Those actuators need precision gearing, and that market is already tight before humanoids arrive. An unsigned editorial in a trade publication, ManufacturingMag, [put it this way in March](https://www.manufacturingmag.com/article/14-month-wait-actuators-2026-robotics-supply-chain): "Fourteen months. That's the current lead time on precision harmonic drive actuators from the three dominant Japanese suppliers — Harmonic Drive SE, Nabtesco, and Nidec-Shimpo — for standard industrial robot joint configurations." The same piece says "Nabtesco controls roughly 60% of the global market for RV reducers used in six-axis industrial robots." That's one trade source, and I couldn't find a second one for either number, so read them as an industry estimate. If they're close, a humanoid maker waits in the same queue as Fanuc and ABB. Tesla's own problem is quality at volume. Per The Information via Electrek, Tesla "relies on outside suppliers, many in China, for motors and the precision gears that drive the joints. Some of them can make good parts in prototype quantities but struggle to keep quality consistent at higher volume." Morgan Stanley's [humanoid note](https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050) says it plainly: "there are few U.S.-based alternatives for many humanoid components, such as screws, reducers, motors and batteries. Nearly every robot developer in the world still requires critical components sourced from China and other parts of Asia."

**Rare earths.** On April 4, 2025, China's Ministry of Commerce and customs administration issued [Announcement No. 18](https://english.mofcom.gov.cn/Policies/AnnouncementsOrders/art/2025/art_0dd87cbee7b045bf93fabe6ab2faceee.html), putting seven medium and heavy rare earths under export licensing: samarium, gadolinium, terbium, dysprosium, lutetium, scandium and yttrium, including the [permanent magnet materials](https://www.mofcom.gov.cn/cms_files/filemanager/policySummary/viewcore_f3a1432ba20248eca12ff7b91bc73fda.html) made from the first four. Less than three weeks later Musk said on an earnings call that [Optimus production "had been affected"](https://www.reuters.com/business/autos-transportation/musk-says-teslas-optimus-humanoid-robots-affected-by-chinas-export-curbs-rare-2025-04-23/) and that Tesla was applying for a license: "China wants some assurances that these are not used for military purposes, which obviously they're not. They're just going into a humanoid robot." Reuters noted that licenses could take "six or seven weeks to several months." Every motor built on those magnets inherits that wait.

**The Chinese cost curve.** Unitree's prospectus shows what vertical integration inside the supply chain buys. Its average humanoid price [fell from 593,400 yuan in 2023 to 167,600 yuan in 2025](https://restofworld.org/2026/unitree-china-humanoid-robot-shanghai-ipo/), about $25,000, while gross margin rose to nearly 60%. It "self-develops and manufactures core components," and imported raw materials make up about 20% of its supply chain. Musk has suggested Optimus would cost around $20,000. Unitree was already near that price, at a profit, at 5,500 units. By August 12 it said it had produced [about 18,000 bipedal humanoids](https://theroboticsmedia.com/article/unitree-18000-cumulative-bipedal-humanoid-production-g1-star-market-ipo-august-12-2026) in total.

My friend's fourth point, that the Chinese won't manage to scale production for the US and Europe, is a claim about politics and market access. The components point the other way: today the Western humanoid depends on Chinese parts more than the Chinese humanoid depends on Western ones.

## Miles are the data side

My friend's afterthought was the self-driven miles. Tesla's fleet passed [10,010,684,206 cumulative miles on FSD (Supervised)](https://eletric-vehicles.com/tesla/tesla-fsd-hits-10-billion-miles-matching-musks-threshold-for-unsupervised-driving/) on May 3, adding about 28.8 million a day. In 2025 alone owners drove [4.25 billion](https://www.teslarati.com/tesla-fsd-supervised-8-billion-miles/). In January Musk had set the bar: "Roughly 10 billion miles of training data is needed to achieve safe unsupervised self-driving."

The miles haven't become scale yet, even in driving. In June, Tesla announced ["Unsupervised Robotaxi now in the entire Austin Metro area,"](https://www.reuters.com/business/autos-transportation/tesla-rolls-out-unsupervised-robotaxis-austin-2026-06-03/) and Reuters reported, from a presentation by Austin officials, that Tesla had roughly 50 vehicles there against more than 250 for Waymo.

Driving data doesn't teach a hand to fold a shirt. What carries over to Optimus is the machine that produces the data. Tesla has moved much of its self-driving annotation team to Optimus, hired data collectors in camera helmets and motion-capture suits, and has over 500,000 hours of robot training data with a goal of doubling it by year-end. The plan is to lease Optimus to companies whose factories look like Tesla's own and feed the deployment data back into the model. Electrek's summary: "That's the FSD playbook: ship the hardware, collect the fleet data, and promise the software will catch up."

In *How to achieve superintelligence* I argued that the strongest data moats come from loops where users produce the training data while getting value from the product. A robot fleet is that loop in hardware. Figure says it almost in those words: "Every robot that rolls off the BotQ line at our new hourly cadence is more than just a unit of hardware - it is a data-collection engine."

This joins the two halves of his argument. If the policy is general enough to learn from any deployment, units shipped are also the data rate. The factory stops being downstream of the intelligence and becomes the pipe that feeds it. That's the strongest version of the thesis, and it depends on a premise I'll hold against the evidence below.

## The West needs one winner

My friend's middle points are geopolitical: high odds of US–China escalation, a Western market closed to Chinese robots, and one Western champion taking most of it.

The Chinese side of the board is already crowded. A Counterpoint analyst told Rest of World there are "more than 100 humanoid companies in China," likely to consolidate to a few dozen after the first IPOs. Unitree plans to produce 75,000 humanoids a year within five years. Morgan Stanley's forecast for 2050 has 302.3 million humanoids in use in China against 77.7 million in the US, and its head of industrials research said "China could catch up when humanoids reach downstream application and mass production, riding on its strong self-sufficient supply chain."

He doesn't spell out why there should be one winner. My reading is that tooling and supply contracts pay back only above some volume. If Western demand splits across ten companies at 10,000 units each, none of them gets there, and all ten buy their actuators from China. Concentration is how a smaller market can still reach the volume where a domestic supply base makes sense.

The Liberty ships complicate this: one design, 16 yards. The Western winner could be a design and a supplier base that many plants build to. That's a guess.

## Data-driven decisions are less rational than they look

My reply that night was a few lines (translated, lowercase kept):

> interesting how data-driven actions aren't that rational. inhibited synthesis -> Kant's argument about it. survivorship bias -> the airplane thing. elon building this thing (I think it's pretty hard to find this action in the current data..., it's more elon's good taste for both predicting and defining (the reflexivity principle soros uses))
>
> a/b is a thing for low-ontology processes. high-ontology processes have to be decided this way, which is hard to algorithmize

I wasn't disagreeing with him. I was saying that nothing in a spreadsheet in 2025 contained the instruction "convert the S and X line into a million-robot line."

**Survivorship.** Abraham Wald's wartime memos were titled [A Method of Estimating Plane Vulnerability Based on Damage of Survivors](https://www.cna.org/reports/1980/0204320000.pdf). The title carries the lesson, and the first memo defines its data as the planes that "received exactly *i* hits but have not been downed, i.e., have returned from combat." The only planes you can inspect are the ones that came back, so the damage you see is the damage a plane can survive. In one of his hypothetical worked examples the conclusion runs against the eye: "the engine area is the most vulnerable in the sense that a hit there is most likely to down the plane." Every ramp I used above is a survivor. We study the Model 3 because Tesla made it through. The factories that tooled up for a product that never found demand aren't in the dataset, and neither are Musk's own misses. A model trained on successful ramps learns that ramps succeed.

**Synthesis.** Kant's question in the *Critique of Pure Reason* was "How are synthetic a priori judgments possible?" A passage the [Stanford Encyclopedia entry](https://plato.stanford.edu/entries/kant-hume-causality/) quotes puts the problem well: "Appearances certainly provide cases from which a rule is possible in accordance with which something usually happens, but never that the succession is necessary." Data gives you what usually happens. A bet like Fremont needs a judgment that holds together a supply chain, a model trajectory, a labor market and a geopolitical risk as one thing. That unity isn't a column in the table. Someone has to make it, and making it is an act rather than a measurement.

**Reflexivity.** George Soros, in the first chapter of [*Open Society*](https://archive.nytimes.com/www.nytimes.com/books/first/s/soros-open.html): "the relationship between thinking and reality is reflexive—that is, what we think has a way of affecting what we think about." And: "Situations that have thinking participants do not inertly wait to be studied; they are actively shaped by the participants' decisions." The way I use it: a forecast made by a participant feeds back into the fundamentals it's trying to describe. Musk's forecast works that way. The announcement moved workers from the S and X line onto Optimus, and it's part of a 2026 capex guidance "in excess of $20 billion" that pays for six factories, the Optimus factory among them. It tells suppliers what volume to plan for, and tells engineers where the interesting work is. A forecast like that predicts the future and is also one of its causes. That's what I meant by predicting and defining.

The forecasts show it. Goldman Sachs raised its global humanoid forecast this month to [about 890,000 units in 2030 and 6.5 million in 2035](https://finvaulta.com/research/goldman-sachs/global-physical-ai-framing-the-forward-progress-of-humanoids-2026-09-16). The same report, [per 24/7 Wall St.](https://247wallst.com/investing/2026/09/14/goldman-sachs-just-supercharged-its-humanoid-robot-prediction-5x-to-6-5-million-by-2035/), raised the 2026 estimate from 51,000 to 75,000, the 2030 one from 256,000, and the 2035 one from about 1.4 million. Citi's 2024 model counts robots in use rather than shipped, so it isn't directly comparable: ["we estimate humanoids could total 13.3m worldwide by 2035, growing to 648m by 2050,"](https://www.citifirst.com.hk/home/upload/citi_research/rsch_pdf_30297368.pdf) in "a $7 trillion humanoid market by 2050." Tesla's announced capacity, Fremont's million plus a Texas line with a long-term output of [10 million a year](https://www.businessinsider.com/elon-musk-tesla-optimus-humanoid-robot-unveiling-date-2026-4) per its shareholder deck, is larger than Goldman's forecast for the entire world in 2035. At full rate it would build Citi's whole 2035 installed base in under fifteen months.

![Tesla's announced Optimus capacity against global humanoid shipment forecasts](https://future-seems-so-good.com/blog/assets/robotics-is-a-factory-problem/charts/capacity-vs-forecast.svg)

Either the plan is too big, or the forecast is too small, or the plan is one of the things that will move the forecast. In a reflexive system you can't tell those apart ahead of time, because the answer depends on what people do after reading both numbers.

The people betting on Tesla specifically are less impressed than the forecasts. On September 25, a Manifold market on whether [Tesla officially reports building 10,000 or more Optimus robots in 2026](https://manifold.markets/BarryJones/will-tesla-officially-report-buildi) stood at 14%, though the market is thin. Another, on whether [Tesla sells humanoid robots to the public in 2027](https://manifold.markets/nonnihil/tesla-sells-humanoid-robots-to-the), stood at 35%.

Reflexivity also runs in reverse. Unitree listed in Shanghai in August, and a month later Nikkei reported its shares had ["declined more than 50% from the opening price."](https://asia.nikkei.com/business/markets/equities/china-s-unitree-robotics-stock-falls-by-half-in-month-after-listing) The narrative that pulls capital into a production bet can pull it out before weekly output has proved anything. A Western champion that funds a decade of tooling with its share price is exposed to the same loop.

**Low and high ontology.** A/B testing works when the categories are fixed. You vary a button or a price inside a world whose objects don't change, and you measure. I call those low-ontology decisions. A high-ontology decision changes what exists: it decides that a car line becomes a robot line, and that a humanoid is a product category at all. There's no control group, because the treatment creates the population you'd have to sample. In [Bottom-up is one ontological level higher](https://future-seems-so-good.com/blog/bottom-up-is-one-ontological-level-higher) I wrote about the same split in organizations. Here it shows up in a single capital allocation decision.

His answer was immediate: "Totally, it's just that I think you can reach that conclusion with more intelligence and data." I think he's right about the conclusion and I'm right about the bet. His eight-point chain is reasoning, and a better model could produce it. More intelligence can get you to "someone should build a million-unit humanoid line." It can't tell you whether this line, run by these people, becomes true, because the data about that gets produced by the commitment. The prediction can be computed. The definition has to be done. Where I'd hold my own view loosely: things that are high-ontology today can get cached into low-ontology procedure later. Once a few million-unit robot lines exist, the next one will look like an optimization problem.

## Where the thesis leaks

My friend's premise is the conditional: *assuming* robot intelligence gets solved like language models were. Most counterarguments go straight at it. The strongest one I found goes at his fourth point instead.

**China is already at factory scale.** He frames the race as who gets to a million first. The race to the first tens of thousands already has a leader. SAG's estimate for the first half of 2026 is about 19,100 humanoids shipped worldwide, more than 97% by Chinese makers, with Unitree alone at 5,900. Unitree has produced about 18,000 bipedal humanoids in total, and AgiBot [announced its 15,000th robot in June](https://www.humanoidsdaily.com/news/a-golden-milestone-figure-manufactures-its-1-000th-figure-03-humanoid). Figure's 1,000th Figure 03 a month later is real progress, and it's more than an order of magnitude behind. So the Western bet is catch-up against a factory base that already exists, with its tooling, suppliers and yield curves. That is a harder bet than a blank slate. It turns back into a blank slate only if politics keeps Chinese units out of Western factories, through tariffs, security rules or export controls. Then his fourth point carries the whole thesis. The question changes from who can build a million robots to who can build a million without Chinese reducers, motors and magnets. The supply-chain section suggests nobody can yet. Morgan Stanley expects adoption to be "relatively slow until the mid-2030s." If it's slow, a Western champion has to finance a decade of low utilization against competitors whose supply chain is in-house.

**Dexterity may not come from this data.** Rodney Brooks, whose companies built the Roomba, PackBot and Baxter, argued a year ago in [Why Today's Humanoids Won't Learn Dexterity](https://rodneybrooks.com/why-todays-humanoids-wont-learn-dexterity/) that the industry collects the wrong data. "Both Figure and Tesla are all in on videos of people doing things with their hands are all that is needed to train humanoid robots to do things with their hands." Human dexterity runs on touch, and "We as a species have not developed technologies to capture touch, to store touch, to transmit touch over distances and time, nor to replay it." His conclusion: "we are more than ten years away from the first profitable deployment of humanoid robots even with minimal dexterity." If he's right, a million-unit line in 2027 builds a million expensive machines that can do little. The factory would be solving the second problem before anyone has solved the first.

**The current evidence leans his way.** The same Electrek report that has Tesla at hundreds a week says the robots "still can't handle generalized tasks," that it "takes several days for Optimus to learn even basic tasks," and that most units are used internally for testing, training and data collection. The hand and forearm contain over 100 screws and small parts assembled by hand, and touch sensors have had reliability problems. The V3 robots on the line now "aren't even the version it plans to commercialize." Electrek's own view is that humanoids will be "a small fraction of the overall robotics industry," because a machine built for one job beats a human-shaped one at it.

**Intelligence might commoditize in a way that doesn't favor the biggest factory.** Physical Intelligence's [π0](https://www.pi.website/blog/pi0) is a single policy trained on data from 8 distinct robots, and it "can control a variety of different robots." If one general policy runs on many bodies, intelligence commoditizes the way the thesis assumes. But the value might go to whoever owns the policy and the most diverse data, and bodies become a contract-manufacturing business. That still makes production matter. It doesn't make one Western brand the winner.

**Tesla's timelines.** Musk said in January 2025 that Tesla would build about 10,000 Optimus robots that year and that "several thousand" would be doing useful work. A year later, per Electrek, he admitted none were. He promised a V3 reveal by mid-2026, and it hasn't happened. The autonomy record is worse: [full autonomy by 2018, a million robotaxis by 2020, unsupervised FSD by June 2025](https://eletric-vehicles.com/tesla/tesla-fsd-hits-10-billion-miles-matching-musks-threshold-for-unsupervised-driving/). The fair reading of the Model 3, though, is that the deadline was wrong and the outcome arrived later at a larger scale. Which of those two patterns Optimus follows is the actual question.

## Speculation

This part is speculation, and I'll mark where it stops being evidence.

The evidence that points toward the premise: π0 runs one policy across 8 robots. Figure trained a whole-body controller in simulation that climbs real stairs zero-shot, "no real-world fine-tuning." Tesla is building a library of movements the robot can recombine for new jobs. These are the early signs of general policies that transfer across tasks and bodies, which is what happened with language models before they became interchangeable.

From here on it's speculation. If robot policies do commoditize, the humanoid business starts to look like the phone business: the brain becomes a component, and the margin sits with whoever builds hardware at volume and owns the deployment data. Figure also says its robots will help build robots on its own lines. If that works at volume, each robot shipped lowers the cost of the next and adds to the data that makes the next one better. That's the positive feedback loop from [Positive feedback eats the world](https://future-seems-so-good.com/blog/positive-feedback-eats-the-world), running through a factory. Nobody has shown it at scale. In January, Tesla said Optimus wasn't used in its factories "in a material way."

If the loop closes, my friend's "faster than we expect" follows, because loops compound and forecasts extrapolate the part of the curve before the loop closed. If it doesn't, the million-unit line is a large bet on a software timeline.

## This is not investment advice

My friend's conclusion was "mispriced Tesla." That's his view, stated in a private chat, and I haven't checked Tesla's price or done any valuation. This essay is about what kind of problem humanoid robotics is. Nothing here is a recommendation to buy or sell anything. A good thesis about production can still be a bad trade, for reasons outside the thesis. The price might already reflect it, or the timing might be off by years.

## A scorecard

If the problem is a factory problem, the evidence will show up as factory numbers, and those are the ones to watch. Demos and reveal dates tell you much less.

| What to watch | Last reading | Source | The thesis gains if | It loses if |
|---|---|---|---|---|
| Optimus units per week, averaged over a quarter | Several hundred (Aug 2026); >1,000 goal by year-end | The Information via Electrek | >1,000 a week for 13 weeks, not a final-week burst | Under 1,000 a week by mid-2027 |
| Units doing paid work outside Tesla | Mostly internal testing and data collection | Electrek | Named external lessees with uptime numbers | Fleet stays internal through 2027 |
| Actuator and reducer suppliers | Many in China, struggling with quality at volume | Electrek; Morgan Stanley | In-house or non-China reducers and motors | Quality problems grow with volume |
| Magnet exposure | Heavy rare earths and their magnet materials under Chinese licensing since Apr 2025 | MOFCOM; Reuters | Magnets sourced outside the licensing regime | A licensing delay shows up in weekly output |
| Cost per unit | Unitree: 167,600 yuan (about $25k) average price, ~60% gross margin, 2025 | Rest of World | Tesla's unit cost or lease price at or below Chinese peers | Tesla stays well above Chinese peers at volume |
| Uptime and yield | Figure: >80% first-pass yield; Optimus hands needing rework | Figure; Electrek | Yield and hours between interventions trending up | Rework and recalls grow with volume |
| Built versus sold | UBTECH: >6,000 capacity, 1,079 sold (2025) | UBTECH annual results | Units sold and working track units built | Capacity keeps outrunning sales |
| Time to teach a new task | Several days for basic tasks | Electrek | Hours, then minutes | Still days in 2027 |
| Chinese output | ~19,100 global shipments in H1 2026, >97% Chinese; Unitree ~18,000 bipeds produced in total | SAG and Unitree via The Robotics Media | Chinese units stay out of Western factories | Chinese robots enter Western factories at a fraction of Tesla's price |
| Forecasters on Tesla | 14% on ≥10,000 Optimus built in 2026; 35% on public sales in 2027 (Sep 25, 2026) | Manifold | Both rise as weekly output rises | Both drift down through 2027 |
| Texas line | Start around summer 2027, 10M-a-year long-term plan | Business Insider | Starts on time | Slips past 2028 |

If I had to pick one number, it would be sustained weekly output together with the uptime of the units. One good week is a press release. A quarter is a factory. And the GA3 lifters showed that a robot's downtime becomes the factory's downtime.

What would falsify the thesis by 2028? Two things together: a general policy that runs well on cheap Chinese bodies, and Tesla still under a few thousand units a week. Then the brain would have commoditized without the factory mattering. What would confirm it is the reverse: the policies converge, and the company with the line is the only one that can ship them in volume outside China.

## Sources

- Tesla, [Q4 2025 earnings call transcript](https://earningscalls.dev/transcripts/tesla-inc_tsla_earnings_call_transcript_2026-01-28), 2026-01-28
- Tesla, [Q4 2025 production, deliveries and deployments (EX-99.1)](https://www.sec.gov/Archives/edgar/data/1318605/000162828026000016/exhibit9914.htm), 2026-01-02
- Tesla, [Second Quarter 2018 Update (EX-99.1)](https://www.sec.gov/Archives/edgar/data/1318605/000156459018018490/tsla-ex991_6.htm), 2018-08
- Lloyd Lee, Business Insider, [Elon Musk says copycats are why Tesla can't unveil Optimus sooner](https://www.businessinsider.com/elon-musk-tesla-optimus-humanoid-robot-unveiling-date-2026-4), 2026-04-23
- Fred Lambert, Electrek, [Tesla ramps Optimus to hundreds a week, but the robots can't generalize](https://electrek.co/2026/09/25/tesla-optimus-production-ramp-hands-ai-generalization-problems/), 2026-09-25 (summarizing a paywalled report in The Information)
- Laurie Chen, Reuters, [Unitree plans Shanghai IPO, testing interest in humanoid robots](https://www.reuters.com/world/asia-pacific/unitree-plans-shanghai-ipo-testing-interest-humanoid-robots-2026-03-20/), 2026-03-20
- Kinling Lo, Rest of World, [The world's largest humanoid robot maker is going public](https://restofworld.org/2026/unitree-china-humanoid-robot-shanghai-ipo/), 2026-03-31
- Longbridge / TechNode, [China's AgiBot leads global humanoid robot shipments in 2025, Omdia says](https://longbridge.com/news/272030450), 2026-01-09
- Figure, [BotQ: A High-Volume Manufacturing Facility for Humanoid Robots](https://www.figure.ai/news/botq), 2025-03-15, and [Ramping Figure 03 Production](https://www.figure.ai/news/ramping-figure-03-production), 2026-04-29
- Humanoids Daily, [A Golden Milestone: Figure Manufactures Its 1,000th Figure 03 Humanoid](https://www.humanoidsdaily.com/news/a-golden-milestone-figure-manufactures-its-1-000th-figure-03-humanoid), 2026-07-26 (quotes Brett Adcock's 2026-07-23 post; also the AgiBot 15,000th-robot report)
- Kaan Tınmaz, The Robotics Media, [Unitree Passes 18,000 Bipedal Humanoids Produced As STAR Market IPO Nears](https://theroboticsmedia.com/article/unitree-18000-cumulative-bipedal-humanoid-production-g1-star-market-ipo-august-12-2026), 2026-08-17 (Unitree's cumulative figure and SAG's H1 2026 estimate)
- UBTECH Robotics, [Annual results announcement for 2025](https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0331/2026033102607.pdf), HKEX, 2026-03-31
- Kohei Fujimura, Nikkei Asia, [China's Unitree Robotics stock falls by half in month after listing](https://asia.nikkei.com/business/markets/equities/china-s-unitree-robotics-stock-falls-by-half-in-month-after-listing), 2026-09-21
- Agility Robotics, [Opening RoboFab: World's First Factory for Humanoid Robots](https://www.agilityrobotics.com/content/opening-robofab-worlds-first-factory-for-humanoid-robots), 2023-09-18
- Jack Stewart, WIRED, [Tesla Made a Lot of Cars in Q3. But It Needs to Make More](https://www.wired.com/story/tesla-q3-production-numbers/), 2018-10-02
- Math Encounters, [Liberty Ship Production Data](https://www.mathscinotes.com/2018/05/liberty-ship-production-data/), 2018-05-22
- Kara Carlson, Austin American-Statesman, [Elon Musk lays out Tesla's big vision for its Austin operations](https://www.statesman.com/story/business/2022/04/08/elon-musk-tesla-future-plans-2022-vision-austin-manufacturing-facility/9511074002/), 2022-04-08
- MOFCOM and GACC, [Announcement No. 18 of 2025](https://english.mofcom.gov.cn/Policies/AnnouncementsOrders/art/2025/art_0dd87cbee7b045bf93fabe6ab2faceee.html) and [policy summary (Chinese)](https://www.mofcom.gov.cn/cms_files/filemanager/policySummary/viewcore_f3a1432ba20248eca12ff7b91bc73fda.html), 2025-04-04
- Hyunjoo Jin, Reuters, [Musk says Tesla's Optimus humanoid robots affected by China's export curbs on rare earths](https://www.reuters.com/business/autos-transportation/musk-says-teslas-optimus-humanoid-robots-affected-by-chinas-export-curbs-rare-2025-04-23/), 2025-04-23
- ManufacturingMag Editorial, [The 14-Month Wait for Actuators: Who Really Wins and Loses in the 2026 Robotics Supply Chain](https://www.manufacturingmag.com/article/14-month-wait-actuators-2026-robotics-supply-chain), 2026-03-13 (single trade source for the lead time and the Nabtesco share)
- Morgan Stanley, [Humanoids: A $5 Trillion Market](https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050), 2025-05-14
- Goldman Sachs, Global Physical AI: Framing the Forward Progress of Humanoids, 2026-09-16, via [Finvaulta summary](https://finvaulta.com/research/goldman-sachs/global-physical-ai-framing-the-forward-progress-of-humanoids-2026-09-16) and Rich Duprey, 24/7 Wall St., [Goldman Sachs Just Supercharged Its Humanoid Robot Prediction 5X to 6.5 Million by 2035](https://247wallst.com/investing/2026/09/14/goldman-sachs-just-supercharged-its-humanoid-robot-prediction-5x-to-6-5-million-by-2035/), 2026-09-14 (2026 estimate and prior figures)
- Citi GPS, [The Rise of AI Robots: Physical AI is Coming for You](https://www.citifirst.com.hk/home/upload/citi_research/rsch_pdf_30297368.pdf), Rob Garlick et al., December 2024
- Manifold, [Will Tesla officially report building 10,000+ Optimus robots in 2026?](https://manifold.markets/BarryJones/will-tesla-officially-report-buildi) and [Tesla sells humanoid robots to the public in 2027](https://manifold.markets/nonnihil/tesla-sells-humanoid-robots-to-the), read 2026-09-25
- Simon Alvarez, Teslarati, [Tesla posts updated FSD safety stats as owners surpass 8 billion miles](https://www.teslarati.com/tesla-fsd-supervised-8-billion-miles/), 2026-02-19
- Cláudio Afonso, [Tesla FSD Hits 10 Billion Miles, Matching Musk's Threshold for Unsupervised Driving](https://eletric-vehicles.com/tesla/tesla-fsd-hits-10-billion-miles-matching-musks-threshold-for-unsupervised-driving/), 2026-05-04
- Reuters, [Tesla rolls out unsupervised robotaxis in Austin](https://www.reuters.com/business/autos-transportation/tesla-rolls-out-unsupervised-robotaxis-austin-2026-06-03/), 2026-06-03
- Rodney Brooks, [Why Today's Humanoids Won't Learn Dexterity](https://rodneybrooks.com/why-todays-humanoids-wont-learn-dexterity/), 2025-09-26
- Physical Intelligence, [π0: Our First Generalist Policy](https://www.pi.website/blog/pi0), 2024-10-31
- Abraham Wald, [A Reprint of "A Method of Estimating Plane Vulnerability Based on Damage of Survivors"](https://www.cna.org/reports/1980/0204320000.pdf), Center for Naval Analyses, 1980 (memos from 1943)
- Stanford Encyclopedia of Philosophy, [Kant and Hume on Causality](https://plato.stanford.edu/entries/kant-hume-causality/)
- George Soros, *Open Society: Reforming Global Capitalism*, PublicAffairs, 2000, [chapter one excerpt via The New York Times](https://archive.nytimes.com/www.nytimes.com/books/first/s/soros-open.html)
- Hacker News, [thread on the Fremont conversion](https://news.ycombinator.com/item?id=47711621), 2026-04
- Private chat with a friend, 2026-08-11, quoted and translated