OpenAI confirms it will build a humanoid robot; Altman says the brain is the hard part

OpenAI CEO Sam Altman has confirmed the company will build a humanoid robot, telling the Sources podcast in two words: "We will." His timeline is concrete and modest. "I think we'll have a cool demo in 2027 ... I don't think there will be robots walking the streets for a few years longer," Altman told Fortune.

The confirmation matters because it moves OpenAI, the dominant vendor of frontier AI software, into direct competition with Tesla's Optimus program and XPENG's IRON humanoid, which recently walked off the production line autonomously after assembly. It also raises consumer privacy questions the company has not yet answered: OpenAI has announced no robot, no price, no subscription model, no release date and no robot-specific privacy framework for a machine that would carry cameras and microphones through someone's house.

Why a human form?

Altman argued the shape follows from the built world. "The world is very much designed for people," he said on the Sources podcast. Doors, keyboards and tools all assume a certain kind of body, so a robot working in those spaces should share it. "Matching that form factor seems good," he said.

The body itself is not what keeps him up at night. Questions about the robot's appearance are "less important than really figuring out ... the brain that makes the robot work," Altman said. Two arms and two legs accomplish nothing if the machine does not know what to do with them.

That framing sets OpenAI's bet apart in a crowded humanoid market. The company is treating robotics as a software and intelligence problem first, an electromechanical engineering problem second.

What must the brain actually do?

Calling it a "brain" understates the difficulty. A capable humanoid has to:

• interpret what its cameras and sensors see;
• parse an instruction;
• decide how to execute it;
• coordinate its body;
• correct itself when the real world interferes.

Consider a mug. The robot must identify the mug, find the handle, estimate the grip and apply the right force to lift it. Fail any step and the result is spilled coffee and a floor to clean.

OpenAI is already hiring against this problem. Its robotics openings span motion control, state estimation, actuator control, whole-robot behavior and the machine-learning infrastructure required to train physical AI systems. The careers page currently lists 27 robotics openings.

The company has prior art. Its Dactyl research project taught a robotic hand to manipulate real objects, trained entirely in simulation. OpenAI later applied the technique to a Rubik's Cube. That work exposed the failure mode that still plagues robotics companies: a model can learn in software what should happen, but physics has the final say in the real world.

Is OpenAI skipping the hardware?

No. Altman's emphasis on intelligence does not mean the body is an afterthought. Job listings indicate OpenAI is designing custom electromechanical actuators that integrate motors, transmissions, sensing and thermal systems. Other posted positions cover manufacturing and the supply chain required to produce hardware at scale. That is substantial engineering for the half Altman called "less important."

The competitive context reinforces the point. Tesla has spent years on Optimus. XPENG's IRON humanoid walked itself off the production line after assembly. Both demonstrate that even advanced software must coax a very complex set of motors and joints into motion.

OpenAI's stated ambitions start smaller than a robotic roommate. One job posting defines the short-term objective as general-purpose robots that help skilled workers build infrastructure. The long-term goal is a personal robot that assists with a much wider variety of tasks.

What is still unknown?

OpenAI has not disclosed what the 2027 demo will showcase, what the humanoid will look like, or when anyone outside the company will get near one. There is no announced product, price, subscription model, release date or privacy framework for a camera- and microphone-equipped machine roaming a home.

The near-term signals are hiring and Altman's own dates. If the 2027 demo arrives as promised, OpenAI will have moved from foundation models to physical machines in roughly the same window its rivals have spent building entire humanoid product lines — and the differentiator, by the CEO's own account, will be the brain, not the body.