The most important number in humanoid robotics this month is not a funding valuation or a demo view count. It is a screw count. According to a report in The Information, relayed in detail by Electrek, the hand and forearm of Tesla’s Optimus contain more than 100 screws and small components, and workers still assemble them largely by hand. That figure sits at the center of the industry’s first true production ramp, and it explains more about the next two years of humanoid economics than any keynote demo ever will.
The ramp itself is real and it is fast. Tesla went from building a few dozen Optimus units a week during small-batch testing in the second quarter to several hundred a week in August, roughly a tenfold increase in a single quarter. Managers are aiming for a continuous automated line capable of more than 1,000 robots a week by the end of the year, against an eventual target of about 20,000 a week that Gizmodo highlighted earlier this month. Supply-chain reporting for September points to component orders sized for roughly 15,000 units across 2026. The line doing all this is not a new gigafactory: it is the former Model S and Model X line at Fremont, vacated when Tesla ended production of both flagship sedans in early May, with workers and engineers pulled from S/X and dozens more from Model Y onto the robot program.
Tenxing output in a quarter, on a converted car line, is a genuine industrial achievement. But the details of the ramp read less like the birth of a product and more like a live audit of everything the humanoid thesis has not yet solved. Three constraints stand out: the hand, the brain, and the human being asked to train it.
Where the robots actually go
Start with the awkward question: what are these hundreds of robots doing? Mostly nothing commercial. According to The Information’s sources, most units are used internally for testing, training and data collection. The robots working inside Tesla’s factories are confined to tightly controlled, supervised areas and are programmed for specific tasks rather than operating as general-purpose machines. The V3 robots now coming off the line are not even the version Tesla plans to commercialize; that variant still has to pass stricter durability and reliability thresholds.
This is the detail that separates a production ramp from a business. Tesla has solved, or at least industrialized, the first problem: building humanoids at hundreds-per-week cadence. It has not yet touched the second: what a humanoid does on the day it ships. A machine that requires days of task-specific programming and a supervised, fenced area is functionally a specialized robot at a specialized robot’s price premium, minus the specialized robot’s efficiency.
The hand is the choke point
The deepest technical problem in the Optimus program is also the smallest volume of its mass. Elon Musk has promised human-level dexterity, and the leaked Gen 3 assets extracted from the Tesla app by community members, as reported by Teslarati, show hands with 22 degrees of freedom, approaching the human hand’s roughly 27. But degrees of freedom are a design headline; manufacturability is the reality, and on that front the reporting is brutal.
The hand and forearm assembly, with its 100-plus screws and miniature components, defeats the fixtures meant to automate it. Stations for hands, joints and electronics testing cannot consistently align parts far smaller and tighter in tolerance than anything in a car, which produces a steady stream of robots needing rework after rolling off the line. Durability is no better: the hand’s touch sensors have shown reliability failures in the field, and Tesla’s fix, planned for next year, is a replaceable “sensing glove” so technicians can swap the sensor skin without replacing the entire hand.
That workaround deserves attention because it is an admission. A hand that eats itself through normal contact, and whose sensing surface is designed to be consumed and replaced like a brake pad, inverts the economics Tesla sells. The company’s pitch is that a general-purpose humanoid amortizes itself across unlimited tasks; a wear item on the most manipulation-critical subsystem instead creates a per-task maintenance meter that never stops running.
The supply chain compounds it. Tesla depends on outside suppliers, many in China, for the motors and precision gears that drive the joints. Some of those suppliers produce excellent prototype-quality parts and then struggle to hold tolerances at volume, the classic valley between craftsmanship and manufacturing. It is the same pattern seen across the component base, from harmonic reducers to the rare-earth magnet supply that this blog covered last week: the West can design humanoid parts at world-class level and cannot yet consistently make them at world-class scale.
The contrast with other programs is instructive. Toyota’s ELEY, the machine carrying the automaker’s plan for hundreds of thousands of factory robots, deliberately uses two-fingered hands, sacrificing anthropomorphic dexterity for a grasp that survives ten thousand cycles. Agility’s newly launched Digit 5 bets on cooperative safety and a 50-pound lift capacity, with no pretense of fine manipulation at all. The most disciplined buyers in the market keep choosing hands that do less, and do it forever, over hands that promise everything and break.
The brain learns one task at a time
The second constraint is generalization. Three people familiar with the system told The Information that Optimus’ AI cannot yet reliably handle a wide range of tasks and behaves unpredictably in situations it has not been trained on. Learning a new basic task still takes several days. Tesla’s approach is to build a library of primitive movements the robot can recombine for new jobs, which is a sound architecture, but the current state is that recombination is slow, supervised and narrow.
The data operation behind it is already enormous. Tesla has accumulated over 500,000 hours of training data and wants to double that by year-end. It has shifted much of its self-driving data annotation workforce onto Optimus, hired dedicated data collectors wearing camera helmets and motion-capture suits, and opened training hubs in Colorado, Arizona and Florida.
The commercialization plan follows directly from that machine: Tesla reportedly intends to lease, not sell, Optimus to a short list of companies whose factories and warehouses physically resemble Tesla’s own, precisely so the robots can adapt more easily, and to harvest fleet data from those deployments to improve the models. Anyone who followed the Full Self-Driving saga will recognize the playbook: ship the hardware, collect the fleet’s data, and trust the software to converge later. It worked well enough to build a data moat in cars. It also produced a decade of supervised-driving promises, and Musk’s January 2025 claim that “several thousand” Optimus robots would be doing useful work by the end of that year collapsed into his January 2026 admission that zero were. The mid-2026 V3 reveal date has already slipped past late September.
The trainers balked first
The most quietly consequential passage in the reporting concerns the humans in the loop. To feed imitation learning, Tesla had factory workers in Texas and California wear motion-capture suits to record their movements as they worked. According to The Information, workers complained because, as Ars Technica puts it, “they knew the robots were designed to eventually replace them.” Tesla has since moved data collection to dedicated teams staffed at the new training hubs.
Every discussion of humanoid labor displacement imagines the moment a robot takes over a station on a line. The actual first displacement is stranger and more immediate: it is the refusal of skilled workers to digitize their own craft for a system built to erase it. That refusal has a cost, and Tesla is now paying it in payroll, because dedicated mocap teams and multi-state training hubs are simply the price of doing imitation learning when your existing workforce declines to volunteer. Any company planning to train generalist robots on the tacit knowledge of its own employees should price in the same dynamic. The trainers hold a veto, and they know it.
The China problem cuts both ways
The geopolitical backdrop makes the supplier issue sharper. In July, the Federal Communications Commission added foreign-made humanoid robots, quadrupeds and even robot vacuums to its Covered List, blocking new commercial imports on national security grounds. The intent is to build a protected domestic robotics supply chain. The immediate effect is a tariff wall around a domestic industry whose flagship program still buys its joint motors and precision gears from Chinese suppliers, while Chinese makers scale on their own timeline: XPeng started running an automated IRON humanoid production line in Guangzhou this month, targeting commercial sales in 2027.
The result is a strange asymmetry. Chinese humanoids are locked out of the American market by regulation, while American humanoids are locked into Chinese components by necessity. If Washington’s industrial policy succeeds, that dependency loosens over years, guided by exactly the domestic component funding now flowing to chips and actuators. If it does not, the FCC Covered List protects a market whose leading manufacturer still cannot source its own gear train at scale. Either way, the supply chain, not the demo, is the battlefield on which the humanoid race’s first decade will be decided.
What to watch now
Four checkpoints will tell investors whether this ramp is a product or a program consuming a car company’s best factory. First, the sensing glove: if the replaceable sensor skin ships next year and cuts hand rework, the worst of the manufacturability story is survivable. Second, the 1,000-per-week automated line: hitting it by January proves the fixtures finally beat the tolerances; missing it means hands remain a craft product. Third, the commercial V3 and its reveal, already months late, which must clear durability thresholds the current line does not. Fourth, the first external lease to a non-Tesla facility, because a customer whose warehouse does not look like Fremont is the first honest test of generalization.
The competitive clock is meanwhile being set elsewhere. Hyundai plans to deploy up to 25,000 Boston Dynamics Atlas robots over the coming years. Agility is taking orders for a cage-free Digit 5 with $300 million booked and 65,000 hours of production experience behind it. Toyota, the largest committed buyer in history, has chosen wheels and two-fingered hands, a walking rebuke of exactly the design space where Tesla is spending its hardest manufacturing dollars.
Musk still calls Optimus potentially “the biggest product ever,” and the Fremont line’s tenfold ramp is evidence that Tesla is building it as if that were true. But the screw count is the story. Until a machine can be assembled by the same automation that would build it at 20,000 a week, with hands that survive contact and software that generalizes beyond yesterday’s task, Optimus remains the world’s most advanced internal data collection fleet. The humanoid era will not arrive when Tesla can build a thousand robots a week. It will arrive when the thousandth robot does something nobody trained it to do.