Tesla has worked out how to build humanoid robots in volume. It has not yet worked out how to make them useful. According to reporting this week from The Information, relayed by Electrek, Gizmodo and Interesting Engineering, the Fremont line that used to build the Model S and Model X was turning out several hundred Optimus V3 units a week in August. That is roughly ten times the few dozen a week it managed during small-batch testing in the second quarter. The robots leaving that line still need days of training to learn basic jobs, and they can behave unpredictably once they leave what they were trained on.
The gap between those two facts is now the main question over Tesla's biggest bet. In July, Elon Musk told investors that Optimus could be "the biggest product ever." The factory is starting to keep up with that claim. The software is not.
A Car Line, Rebuilt for Robots
Tesla stopped building its two flagship vehicles in early May and moved workers and engineers from the S/X program, along with dozens more from Model Y, onto Optimus. Managers are reportedly aiming for a continuous automated line that can build more than 1,000 robots a week by year-end. The eventual target is about 20,000 a week.
Almost none of those robots are going to customers. Most units are used internally for testing, training and data collection. The ones working in Tesla's own plants, per The Information, operate in "tightly controlled areas where their movements and tasks can be closely supervised." They run specific programmed tasks, not general-purpose work. The V3 now in production is also not the version Tesla plans to commercialize, which still has to pass stricter durability and reliability tests.
The hardware is where the ramp hurts most. Each hand and forearm contains more than 100 screws and small components that workers still assemble by hand. Fixtures at several stations, including hands, joints and electronics testing, do not reliably line up parts with tighter tolerances than anything on a car, so more robots need rework after they come off the line. Some of the hand's touch sensors have had reliability problems. Tesla's fix is a replaceable "sensing glove," due next year, so a failed sensor no longer means swapping the whole hand. Suppliers of motors and precision gears, many of them in China, can reportedly make good parts at prototype volumes but struggle to keep quality consistent at scale.
The Brain Is the Bottleneck
The more serious problem is the AI. Three people familiar with the system told The Information that Optimus cannot yet reliably handle a wide range of tasks. Even basic new jobs reportedly take several days to learn. Tesla is trying to build a library of basic movements such as grasping, lifting, placing and walking, which the model could recombine for new work. The company has more than 500,000 hours of training data and wants to double that by year-end. It has moved much of its self-driving annotation team onto the robot, hired data collectors who wear camera helmets and motion-capture suits, and opened training hubs in Colorado, Arizona and Florida.
Tesla's go-to-market plan follows the same approach it used for Full Self-Driving. Rather than selling robots, it reportedly plans to lease them to a short list of companies whose factories and warehouses resemble its own, then use the data from those deployments to improve the models. Musk has said Tesla could begin selling Optimus by late 2027.
Electrek editor Fred Lambert, a longtime Tesla critic, put the stakes bluntly: "A humanoid that has to be trained for days on each task and kept in a fenced-off area is just an expensive, less efficient specialized robot."
Why It Matters
Humanoid robots are expensive because their pitch is generality. One machine shaped like a person should be able to work anywhere people already work, without the retooling that fixed industrial arms or wheeled warehouse bots require. If every new task takes days of training, and every deployment needs a supervised enclosure, that advantage disappears and the unit economics look much like any single-purpose robot, only worse.
Tesla's track record makes the timeline hard to trust. In January 2025, Musk said Tesla would build about 10,000 Optimus robots that year and that "several thousand" would be doing useful work by year-end. A year later, he acknowledged that no Optimus units were doing useful work at Tesla. A V3 reveal promised for mid-2026 still has not happened. Tesla has now given up two vehicle lines to a product whose core capability depends on an AI breakthrough nobody in the industry has achieved yet.
Competitors are not waiting. Figure says its BotQ facility in California has scaled to producing about one Figure 03 per hour. XPeng started running an automated line for its IRON humanoid in Guangzhou this month and is targeting commercial sales in 2027. Unitree, which shipped more than 5,500 units in 2025, sells a full-size humanoid online for about $16,000. None of them has solved generalization either.
For AI more broadly, humanoids are the clearest test yet of whether the scale-the-data approach that worked for language models carries over to the physical world. Tesla is betting fleet data will close the gap, as it has for years with FSD.
What to Watch
The next checkpoints are close. Watch whether Tesla holds the 1,000-per-week target by December, whether the long-delayed V3 demo finally appears and shows unscripted, multi-task behavior rather than choreographed routines, and whether Tesla names its first leasing customers. Two numbers matter most over the next two quarters: how long Optimus takes to learn a new task, and how many robots are doing paid work outside Tesla's own fences. Tesla did not respond to Gizmodo's request for comment. If those metrics do not move, a faster production line will simply mean more robots waiting for their software.
“A humanoid that has to be trained for days on each task and kept in a fenced-off area is just an expensive, less efficient specialized robot.”— Fred Lambert, Editor in Chief, Electrek