Tesla's Optimus line is up tenfold — the hands are the bottleneck
The first hard production numbers on Optimus since the ramp restarted, and the reason Musk's timelines keep sliding, are both in one report today. Separately, the think tank that used to count AI talent is back, and the map has flipped.
Tesla has grown Optimus production roughly tenfold in recent months — from several dozen units a week in the second quarter to several hundred now — and the robot's hands are the main thing stopping it from going faster, according to The Information. The report names three blockers: supplier constraints, automated equipment failures, and the intricate mechanical assembly the hands require, which it calls the primary manufacturing bottleneck. The targets that sit on the other side of that gap are the story: Tesla wants 1,000 robots a week by the end of the year, on the way to an eventual 20,000 a week — the same 50,000-robot year it was auditing Chinese suppliers for last week.
The units being built are not the product Tesla wants to sell. The current Fremont-built Optimus V3 has a lighter chassis, a bigger camera array and a more human-like build than its predecessors, and it runs on a model trained on more than 500,000 hours of real-world data — but it is restricted to tightly programmed, supervised tasks in Palo Alto test zones and on factory floors. Early commercial units will be leased rather than sold, into warehouses that mirror Tesla's own, so the company can physically retrieve hardware for mechanical upgrades while it keeps harvesting factory data. Musk says full commercial sales could start in late 2027, "provided" being the operative word in that sentence.
Hands have been the humbling part of humanoid robotics for a decade. A robot that walks convincingly across a stage is doing a solved problem; one that picks a bolt out of a bin, over and over, at a rate a factory will pay for, is doing a hand problem — dozens of actuators, tendons and sensors in a package that has to survive millions of cycles. So the numbers read less like a stumble than like the moment the manufacturing physics arrives to price the demo. The lease-only rollout is the tell: Tesla is not confident enough in the durability of V3 to sell it to you. Our own coverage tracked the other half of this buildout — Tesla audits Chinese suppliers for 50,000 Optimus robots in 2026 — and China's humanoid assembly lines already run at a cadence Tesla is still chasing, as we noted in UBTECH switches on a factory that builds a humanoid every 10 minutes.
Carnegie China's revival of the discontinued MacroPolo talent tracker finds that China has overtaken the United States as the leading workplace for elite AI researchers: 41 percent of the tracked cohort works in China against 34 percent in the US, reversing 2022's 46–27 split in America's favor. The study samples authors of NeurIPS papers — 5,823 accepted papers and 25,677 authors for 2025 — and runs its analysis on the 10,280, about 40 percent, whose undergraduate, graduate and current employment histories are all recorded on OpenReview. Chinese-origin researchers also became the largest single bloc, at 57 percent against 13 percent US-origin, and the share of Chinese-origin researchers who stay in China rose from 57 percent to 69 percent. Peking University, not Google, is now the number-one institution for producing top AI talent.
The nuance is where the report is most useful, and it cuts against the headline. The United States still runs the flows: Carnegie measures a net gain of 2,145 researchers for the US and a net loss of 1,729 for China, a one-way imbalance of 30 to 1 — down from 46 to 1 in 2022, but still one-way. The US also retains 89 percent of its own talent, ahead of China's 69 percent. What has changed is where people already in the field choose to sit, and Carnegie attributes it to two forces pulling in opposite directions: China's AI industry now offers enough jobs and comparable pay to keep its graduates, while US visa friction keeps more of them from leaving in the first place. Two caveats belong in the record — this is a single-conference cohort built on self-reported profiles, and the authors disclose that they used Claude Code and OpenAI Codex to write and test the analysis scripts, with human rule-setting and quality control.
What to watch: whether Tesla puts a real weekly build number in its next earnings call, or leaves "several hundred" as the last published figure — and whether next year's talent edition keeps the 41–34 split or reverts.
Which bottleneck matters more for the next two years of AI — robot hands and factory throughput, or where the researchers who build the models decide to live? Tell us in the comments.
Sources: The Information · Investing.com · MT Newswires · Techmeme · Carnegie China — Who's Ahead in the Global AI Talent Race? · South China Morning Post · 观察者网 Guancha
