Tesla is finally building Optimus humanoid robots at something resembling scale — several hundred a week at its Fremont factory, roughly ten times what it was making just months ago. There’s just one problem: the company reportedly can’t sell you one yet, because the robots still aren’t smart enough to be useful.
According to Electrek, which covered a new paywalled report from The Information, Tesla has gone from a few dozen Optimus units a week during small-batch testing in the second quarter to several hundred a week in August. Managers are reportedly aiming for a continuous automated line capable of more than 1,000 robots a week by year-end, with an eventual ambition of around 20,000 a week.
The Hundreds Have Nowhere to Go
So where are all these robots going? Mostly nowhere. According to the report, most units are being used internally for testing, training, and data collection. The ones working inside Tesla’s factories are confined to tightly controlled, supervised areas, and they’re programmed for specific tasks rather than running as general-purpose machines.
The V3 robots Tesla is building now aren’t even the version it plans to commercialize. As Gizmodo reported, the current design — lighter, with more cameras and a more human-like appearance — still has to pass stricter durability and reliability thresholds before it faces a paying customer.
That line lives where the Model S and Model X used to be built. Tesla ended production of its two flagship vehicles in early May and converted the Fremont line to Optimus, pulling workers and engineers from S/X, and dozens more from Model Y, onto the robot program.
100 Tiny Screws Per Hand
The hardware bottleneck is embarrassingly human: hands. Each hand and forearm contains over 100 screws and small components that workers still assemble by hand. Fixtures on several stations — hands, joints, electronics testing — can’t consistently line up parts that are far smaller and tighter-tolerance than anything on a car, so more robots need rework after coming off the line.
Durability is its own problem. Some of the hand’s touch sensors have had reliability issues, and Tesla’s planned fix is a replaceable “sensing glove” arriving next year, so it won’t have to swap the whole hand when a sensor fails.
Then there’s the supply chain. Tesla relies on outside suppliers, many in China, for motors and precision gears. Some of them can make good parts in prototype quantities but struggle to keep quality consistent at higher volume.
Days to Learn a Basic Task
The bigger issue is the brain. Three people familiar with the system told The Information that Optimus’ AI can’t yet reliably handle a wide range of tasks, and that its behavior can be unpredictable in situations it hasn’t been trained on. It still reportedly takes several days for Optimus to learn even basic tasks.
Tesla is fighting that with data: over 500,000 hours of training data, a goal to double that by year-end, and much of its self-driving data annotation team moved over to Optimus, plus dedicated data collectors in camera helmets and motion-capture suits at training hubs in Colorado, Arizona, and Florida.
And here’s the part nobody else is emphasizing: Tesla reportedly intends to lease, not sell, Optimus to a short list of companies whose factories and warehouses look like its own — so the robots have an easier time adapting, and so Tesla can harvest their deployment data to improve the AI. That’s the FSD playbook, applied to humanoids: ship the hardware, collect the fleet data, promise the software will catch up.
Why This Matters
Step back and look at the scoreboard. In January 2025, Musk said Tesla would build about 10,000 Optimus robots in 2025 and that “several thousand” would be doing useful work by year-end. A year later, he admitted that zero Optimus robots were doing useful work at Tesla. The production ramp is real — 10x in a few months — but the hardware is outrunning the software, and Tesla’s answer is to make the ramp itself a data-collection exercise.
It’s also the same bet Tesla made in 2016, when it told customers every car it built had the hardware for full self-driving. That software still requires supervision nearly a decade later. Now Tesla has killed the Model S and Model X to build a robot whose AI can’t do what the product needs it to do — while competitors like XPeng are already running an automated humanoid production line in Guangzhou, targeting commercial sales in 2027.
Meanwhile in Robotics: meet O-ID, the Tokyo startup betting humanoid robots should be easy to fix — a bet that looks smarter by the day — and Feather Robotics is pitching a $30,000 humanoid as the “Android of robotics”, selling developer hardware today instead of promises for 2027.
