Tesla begins early Optimus production, but Musk admits hands, safety and AI are unfinished

Tesla has begun early production of its Optimus humanoid robot on a high-volume line in Fremont, California, CEO Elon Musk announced on the company's earnings call, according to The Verge. The milestone comes with large caveats: Musk himself listed the robot's hands, safety, reliability, durability and the underlying AI as areas that still need work before Optimus becomes genuinely useful.

Production has started, the robot is not finished

Tesla called the machines coming off the Fremont line its first "production-quality bots" and said the manufacturing process was suitable for a commercial product. As The Verge points out, the company did not disclose how many robots it has built, which tasks they can perform autonomously, or how often they finish those tasks without a human stepping in.

Musk expects the remaining pieces to come together around the end of 2027, when Tesla could start delivering Optimus to outside companies. He did not say which companies, or what the robots would actually do for them. He described dangerous, repetitive industrial work performed around the clock as the first market, with home use arriving only once the robot has a far wider range of skills.

Hands and durability are the known hardware gaps

Musk acknowledged that Tesla is ramping production before fixing what he called Optimus's main hardware problem. "We're building robots that do not have perfected hands," he said. Unlike human hands, robotic ones cannot heal themselves, and they must survive repeated use while staying soft enough to handle delicate objects without damaging their sensors. Durability applies to the rest of the machine as well: "It can't, like, wear out after a month," Musk said.

Big scale ambitions, slow start

Most of the equipment for a Fremont line designed to produce a million robots a year is already in place, according to Musk, but reaching 20,000 robots per week will take "at least a year, maybe a little longer." He warned that output will follow an S-curve with a slow, bumpy beginning, that parts of the design are still changing, and that Tesla needs to build out a very large global supply chain. "There's like 10,000 things that have to go right for production of the robot," he said, "and it will move as fast as the slowest."

The software gap may be harder than the hardware

Closing that gap requires training data Tesla does not have. Unlike its cars, Optimus has no vast installed base feeding information back to the company, so Musk said Tesla needs to build roughly 10,000 to 20,000 robots to practice tasks at what he called "Optimus Academy," learning through self-play in the way a toddler gradually masters harder skills.

Ken Goldberg, a robotics professor at UC Berkeley, questioned whether that approach transfers to physical machines. Self-play is well established in software, he told The Verge, but "when you get into hardware, it's not so clear." He noted that robots have become adept at walking, running and jumping, while dexterity, such as manipulating wires and fabric or assembling products, remains far harder. In robotics labs, task success rates have stalled at around 70 to 80 percent, which he called unacceptable for real industry. He also described the "video emptor" problem: demos may show the one success in a thousand attempts or hide an operator offscreen. Some of Goldberg's colleagues think a truly useful humanoid is still 25 years away; he is more optimistic but does not expect the hardest problems to be solved on Musk's timeline.

A new chip and a promised demo

Musk said he hopes to integrate Tesla's AI5 chip into Optimus around April 2027, delivering roughly 10 times the computing power of AI4. He also promised a video within a month or two showing the robot performing tasks that are not preprogrammed, while criticizing humanoid demos that look impressive but would fail in real-world use. His longer-term predictions remain expansive: Optimus robots building future Optimus robots, and eventually more humanoids on Earth than humans.

Why it matters

The announcement separates two very different milestones: being able to manufacture humanoids at volume, and having a robot that can do dependable work. Tesla is betting it can scale the hardware first and let the growing fleet generate the data needed to fix the software. The risk, as Goldberg put it, is that "if you have a lot of hardware without the software, I think that's a big problem." The end-of-2027 delivery window and the promised unscripted demo give concrete checkpoints for judging whether that bet is paying off.