OpenAI Confirms It Will Build Humanoid Robots, Starting With Industrial Models for Data Centers

OpenAI has confirmed for the first time that it will build humanoid robots itself, a decision that transforms the company from the AI industry's dominant software provider into a would-be hardware maker and puts it in direct competition with the startups it once backed. Chief executive Sam Altman said on the Sources podcast on September 2, 2026 that OpenAI will definitely build humanoid robots and also robots of other forms, and he signaled that the company's short-term priority is not the home companion robots that dominate consumer robotics marketing but industrial machines that can assist skilled workers building and operating data centers. The remark ends years of strategic ambiguity in which OpenAI acted as robotics investor and model provider while leaving the bodies to companies like Figure AI and 1X.

The confirmation is a return to territory OpenAI explored and then abandoned. The company ran a robotics research program from 2017 to 2020, producing the Dactyl robotic hand in 2018 and a one-handed robot that solved a Rubik's cube in 2019, before shutting the team down around 2020 because training data for physical robots was scarce and expensive. 36Kr reported on September 3 that Altman now believes the situation has changed, because advances in multimodal models, video generation, simulation and synthetic data have removed the bottleneck that killed the earlier effort, and the company has quietly rebuilt a robotics division around the world-simulation research that grew out of that era.

Key Facts

The podcast remarks were unambiguous about intent. When asked directly whether OpenAI would build humanoid robots, Altman replied that the company will definitely build humanoids and will also build robots with other body shapes, according to 36Kr's September 3 report. He explained the humanoid form as a practical choice rather than an aesthetic one, noting that the world is designed around the human body, with doors, stairs, keyboards and factory equipment all built for people, so a machine that works in that world should match that shape. The same logic, he said, argues for non-humanoid robots in settings where a different body is more efficient, which is why OpenAI is also pursuing specialized forms for data center work.

The industrial focus is the most consequential detail for the competitive landscape. Yahoo Tech reported on September 3 that Altman said home companion robots are not the top priority right now and that the near-term emphasis is on robots that can assist skilled workers in building and operating data centers, the physical infrastructure of the AI boom itself. The strategic logic is direct, OpenAI's own growth depends on data centers, the industry faces a shortage of the skilled workers who build and maintain them, and a robot that extends those workers' productivity would address a constraint OpenAI feels as acutely as any hyperscaler. The company also said it will develop specialized non-humanoid robots for data center tasks, acknowledging that some jobs are better done by machines shaped for the task rather than shaped like a person.

The organizational pieces are already in place even though no hardware has been shown. The Paper reported on September 4 that OpenAI's robotics work is now led by Aditya Ramesh, the former head of the Sora video-generation project, and that the division emerged from the company's world-simulation research, which began hiring publicly around late May and early June 2026. Job postings reviewed in the September coverage show OpenAI recruiting across robotics software, electrical engineering, firmware and product safety, with some roles advertised at salaries up to $445,000 plus equity, a sign that the company is paying frontier rates to staff a hardware effort from scratch. As of the announcement, OpenAI had shown no prototype, named no manufacturing partner and given no release timeline.

Analysis

What this really means is that the partnership model at the heart of the humanoid robot industry is breaking down, and OpenAI has decided it cannot afford to leave the body of the robot, and therefore the physical interface to the real world, in someone else's hands. For the past two years OpenAI has been the intelligence layer behind several high-profile robot startups, supplying the models that let their machines perceive and act, while the startups supplied the hardware. Altman's confirmation that OpenAI will build its own robots inverts that relationship and makes OpenAI a direct competitor to the very companies that made it the default brain for the category, a shift that will force every robotics startup to reconsider its dependence on a single model provider that is now also a rival.

The bigger picture here is that OpenAI is following the same playbook it used to dominate software, integrate vertically before a competitor locks in the standard. In chatbots, OpenAI won by pairing a frontier model with a consumer product and an API ecosystem before rivals could establish the default; in robotics, Altman appears to believe that the winner will be the company that controls both the intelligence and the physical platform, because the feedback loop between real-world robot data and model improvement will be the moat that matters. The focus on data centers reinforces that logic, because it gives OpenAI a captive first customer for its robots, its own infrastructure buildout, where it can iterate on hardware without needing to win over skeptical enterprises first.

The history of OpenAI's earlier robotics retreat hangs over the announcement as a cautionary tale. The company's 2017 to 2020 program produced genuinely impressive research, the Dactyl hand and the Rubik's cube solver were landmark demonstrations, but it was shut down because the data problem made progress too slow and too expensive. Altman's argument that multimodal models and simulation have solved that problem is plausible but unproven, and the gap between a model that can plan a manipulation task and a robot that can perform it reliably for years in a data center is exactly the kind of gap that has humbled well-funded robotics efforts before. The absence of a prototype or a timeline, reported by The Paper on September 4, is a reminder that this announcement is a statement of direction rather than a product launch, and the real test will come when OpenAI has to ship hardware that works outside a lab.

Why It Matters

For the robotics industry, OpenAI's entry as a hardware maker reshapes the competitive map, because the startups that positioned themselves as OpenAI's physical partners must now decide whether to compete with their former backer, pivot to other model providers or focus on niches OpenAI will ignore. For the AI infrastructure sector, the announcement points to a future in which the data center labor shortage is addressed partly by robots, which could change how new facilities are designed and how quickly they can be brought online. For OpenAI, the move is a hedge and an offensive at the same time, a hedge because the company needs physical capacity to continue growing, and an offensive because controlling the robot hardware market would extend its dominance into the physical economy. And for workers and unions in construction and data center trades, the prospect of humanoid coworkers raises questions about training, safety and job displacement that the industry has barely begun to address.

Next Up

In the coming weeks, watch for OpenAI to publish technical details or hiring milestones that reveal the maturity of its robotics program, since a company that confirms robot plans without showing hardware is usually still early in the development cycle. Watch also for reactions from the robotics startups that have relied on OpenAI models, because any of them announcing a switch to alternative model providers would confirm that the partnership model is fracturing. The most important signal will be a prototype reveal or a manufacturing partnership, because OpenAI has neither, and the first concrete evidence that the company can actually build a robot body, rather than just the brain inside one, will determine whether this announcement marks the start of a new hardware era or another ambitious robotics program that ends in retreat.