Altman Confirms OpenAI Will Develop Humanoid Robots
OpenAI CEO Sam Altman confirmed on the podcast Sources that the company is developing humanoid robots as well as other forms, stating, "We will definitely make humanoids." When asked whether they are focusing on humanoids or data center machines, he replied that they are working on both.
He explained that the world is built around human bodies: opening doors, typing, operating devices, and cleaning kitchens, so matching humanoid forms requires less environmental modification. However, he emphasized that creating physically capable intelligence is more important than just making a body. Having one robot per person is not the immediate priority; the focus is on infrastructure and manufacturing first. In the long term, he believes everyone should have a personal robot to do tasks they do not want to do.
This is not a new product launch. OpenAI shut down its early robotics team in 2020, restarted in 2024, and plans to officially expand its full-stack hardware, operations, systems, and machine learning positions by the end of May 2026. The number of job openings in San Francisco has increased from about 11 at the start of the division to about 19, with several positions focused on actuators. The team is integrated with the robotics business led by Aditya Ramesh, with short-term goals to assist skilled workers in building data centers and other infrastructure, where non-humanoid specialized machines can be used.
The company previously invested in Figure and collaborated on humanoid intelligence, with the partnership ending in 2025; it remains an investor in the home humanoid company 1X Technologies. There have also been internal discussions about spinning off robotics and consumer hardware for financing, but this was put on hold due to consolidation issues. Altman has mentioned that by 2027, robots capable of working in the real world may emerge, which is a time estimate, not a delivery commitment.
In the first half of 2026, global humanoid shipments are expected to be around 19,100 units, a significant year-on-year increase, with the market narrative shifting from concept to small-scale production. Tesla's Optimus, Boston Dynamics, and various startups are following a model of first creating a body and then layering on models; OpenAI states that hardware and machine learning are designed together to avoid purchasing off-the-shelf bodies and then adding a brain.
In terms of market mechanisms, the valuation premium is for "model companies entering the entity," while the old boundaries are for those only doing cloud-based conversations. Funding will flow towards actuators, remote operation data collection, and data center operations, rather than pre-sales for homes. Beneficiaries will be hardware suppliers that can provide parts and production lines to OpenAI, as well as data factories already running remote operations; those under pressure are humanoid startups that rely solely on OpenAI model licensing without controlling the entity. Personal robots remain a long-term demand, with current orders focused on data centers and production lines.
There are currently no models for walking products; the pricing is first set for narratives and recruitment structures.
Source: Public Information
ABAB AI Insight
OpenAI's serious approach to humanoid robots is due to the saturation of text data benefits, as physical world data needs to be collected using its own arms. The 2020 disbandment of the robotics team was due to cheaper internet corpus at that time; the 2024 restart indicates that models can now support operations. The investment in Figure and subsequent split shows that licensing a brain to someone else's body does not allow for hardware premium separation or control over data loops. The shift to self-developed entities integrates data collection, actuators, and model training into the same balance sheet.
Capital is diverting from GPU training clusters to actuators and factories. Short-term robots will first assist in data centers like Stargate by connecting cables and moving equipment, turning the manpower bottleneck in AI infrastructure into their own hardware's first customers. Consumer humanoids are lower priority because home scenarios have low fault tolerance, high liability insurance costs, and lack the repetitive actions found in data centers. Discussions about splitting for IPO have been retracted, indicating that hardware losses still need to be supported by software cash flow.
Comparing to Tesla's Optimus which spreads costs over automotive production lines, Figure which trades factory orders for data, and 1X which follows a home subscription model, OpenAI is in a remedial phase of "having a brain but lacking a production line," not a competitor to already mass-produced automotive companies. The industry is shifting from investing in others' bodies to fearing that bodies are not in their own hands.
Structural changes involve technological substitution combined with industry chain reconstruction. The mechanism is that after language models exhaust the internet, the next training data will be in kitchens and server cabinets; whoever owns the arms owns that data. Humanoids are just interfaces, and behind the interfaces is data ownership. The narrative of everyone having one is a consumer story, but the first to work will be specialized machines that do not enter living rooms.
ABAB News · Cognitive Laws
First build hands for data centers, then build humans for living rooms.
Bodies can be purchased, but data loops cannot be.
Everyone having one is the end narrative, but the first orders are in infrastructure.