OpenAI is building its own humanoid robot, and Sam Altman promises a demo in 2027

The question took one line. Alyson Shontell, Fortune’s editor-in-chief, asked Sam Altman whether OpenAI was building a humanoid robot. “Yeah,” the chief executive replied, adding: “I think we’ll have a cool demo in 2027—a great demo.” Then he dialled expectations back down. “I don’t think there will be robots walking the streets for a few years longer, but we’ll have something impressive.”

The exchange, broadcast on September 15 in Fortune’s video series on business leaders, resurfaced over the weekend, notably in Engadget on Saturday, October 17. It sheds light on a project OpenAI has been pursuing quietly for months, and whose shape, price and release date remain unknown.

A post on X, a podcast, a stack of job ads

The first signal came on May 31. In a post on X, Altman announced OpenAI Robotics and called for hardware, systems, operations and machine-learning engineers. “AI should be able to help people in the physical world,” he wrote. He sketched a two-stage plan: “In the short term, we are focused on robots to support skilled workers to build our future infrastructure; in the long term, we imagine everyone having a personal robot doing anything they need.” The team grew out of OpenAI’s research on simulating the physical world and is led by Aditya Ramesh, according to The AI Insider.

On September 1, in the first episode of journalist Alex Heath’s Sources podcast, Altman was blunter. Asked whether OpenAI would build a humanoid, he said: “We will.” His reasoning is practical: with its doors, stairs and tools, “the world is very much designed for people.” But he quickly shifted the emphasis. What the robot looks like, he argued, matters less than “really figuring out ... the brain that makes the robot work.”

The job listings say more than the interviews. Engadget counted 27 open robotics roles on OpenAI’s careers site. The company is looking for specialists in motion control, in state estimation (knowing where every part of the robot’s body is), and in custom electromechanical actuators, the motors that drive the joints, as well as manufacturing and supply-chain staff. In other words, OpenAI is not just writing models for other people’s robots: it is preparing to produce its own.

A comeback after an earlier retreat

OpenAI has done robotics before. In 2018 and 2019, its Dactyl project taught a mechanical hand to manipulate objects and then solve a Rubik’s Cube, training almost entirely in simulation. The team was later disbanded, around 2020 and 2021, as the company bet on language models, where training data was plentiful and progress faster. It stayed in the field through investments. In 2023 it led a $23.5 million round in the start-up 1X; in February 2024 it took part in Figure’s $675 million raise and signed a deal to supply models for its robots. Figure ended that partnership in February 2025 to build its own model, Helix.

Those two stakes leave OpenAI in an awkward spot: it is becoming a rival to companies it partly owns. 1X is already taking preorders for NEO, its home humanoid, at $20,000 or $499 a month, with first US deliveries promised for 2026. Figure, valued at $39 billion in its September 2025 round, is targeting factories and warehouses first.

Against Tesla and the Chinese, OpenAI starts far behind

The competition has not waited. Tesla is replacing its Model S and Model X lines in Fremont, California, with a first Optimus production line designed, according to its April quarterly update, for one million robots a year; a second line in Texas is being designed for ten million in the long run. Those are stated capacities, not robots delivered, and Elon Musk has already had to push back several deadlines on the project.

In China, the volumes are already real. Unitree, whose initial public offering on Shanghai’s STAR Market got under way in August, shipped more than 5,500 humanoids in 2025 according to its prospectus, on revenue of 1.70 billion yuan (about $240 million), more than four times its 2024 figure. Carmaker XPeng, for its part, has had its IRON humanoid walk off its assembly line.

OpenAI’s bet comes down to one idea: robot bodies will look increasingly alike, and the difference will lie in the software that drives them, where the company believes it has an edge. That is also the hardest part. Keeping a robot upright on a building site, or having it grasp an unfamiliar tool without breaking it or hurting anyone, requires data the internet does not supply, which explains OpenAI’s focus on simulation.

So far the company has released no photo, no spec sheet, no price and no rules on the data a robot in someone’s home would collect. Altman has not said what the 2027 demo will show. He has only set an order of battle: the building site first, the home after.