Altman Says OpenAI Will Build Humanoids, Opening a Path Beyond Software

OpenAI CEO Sam Altman has made the company’s ambitions around OpenAI humanoid robots explicit for the first time. In an August 26 TIME interview, journalist Alex Heath reported that Altman confirmed OpenAI would “definitely” build humanoid robots, adding that he believes everyone should eventually have access to a personal robot of their own.

That commitment, later covered in The Rundown’s September 3 newsletter, signals a broader ambition to build physical systems around OpenAI’s existing intelligence capabilities. Still, the interview itself establishes neither a specific release date nor an actual shipping product, leaving the announcement firmly in the realm of stated intent for now.

Why Physical Hardware Matters to OpenAI’s Strategy

Building physical robotic systems could give OpenAI significantly greater control over how its models act in the real world, along with more direct influence over how it gathers robot training data. That shift would also bring new responsibilities, specifically for the physical machinery that translates a model’s decisions into actual movement.

What OpenAI Is Already Assembling

An OpenAI robotics operations job posting, visible as of September 6, describes an effort focused on integrating hardware and software across multiple robot forms. The role specifically covers data collection facilities, equipment operators, testing rigs, equipment readiness, throughput, downtime, and overall data quality.

These responsibilities offer concrete evidence that operational work on the robotics side is already underway. However, they don’t establish the actual size of the program, its current output, or whether it has led to measurable improvements in any specific model. Since the posting itself is undated, it also can’t confirm exactly which details were already public when The Rundown’s newsletter first covered the story.

The emphasis on dedicated collection facilities is particularly significant. Training robots effectively requires a reliable way to gather useful examples of physical behavior, maintain testing equipment properly, and assess the quality of resulting data. OpenAI appears to be actively hiring someone specifically to manage that process alongside the broader integration of intelligence and physical machinery.

Why Owning the Hardware Could Matter

Robot intelligence remains central to OpenAI’s broader strategy, but building the physical body itself could give the company far more control over how that intelligence gets developed and deployed in practice. Sensing, model behavior, and physical movement would become interconnected parts of a system OpenAI could refine together, extending its engineering responsibilities into mechanical reliability and physical control systems as well.

Figure’s February 2025 Helix announcement helps illustrate why these connections matter so much. Figure described an architecture separating slower visual and language processing from faster motor control systems, while also reporting roughly 500 hours of teleoperated training behaviors. While these remain Figure’s own technical disclosures rather than independently verified results, they demonstrate how robot learning connects directly to examples of real physical action.

For OpenAI humanoid robots, owning a dedicated robot platform could support a similar development cycle: gathering observations and actions suited to a specific training goal, revising the underlying model, then testing improvements on the same physical hardware. That approach would give OpenAI considerably more control over what data it collects and how it evaluates progress over time. Whether this ultimately produces better robot performance, or meaningfully benefits OpenAI’s broader software models, remains genuinely uncertain based on current available information.

What Success Would Actually Look Like

Measuring success for OpenAI humanoid robots will depend heavily on how the rollout actually unfolds. The September 3 newsletter described infrastructure work as the likely starting point. That sequence, however, remains unconfirmed as an official product roadmap.

If OpenAI does begin with infrastructure-focused tasks first, prospective operators would need concrete measures of success, including which specific jobs a machine can reliably complete, how often human intervention becomes necessary, and how much genuinely useful work gets delivered between interruptions. The operational priorities outlined in OpenAI’s job posting, readiness, downtime, throughput, and data quality, offer relevant early indicators, even though the role itself focuses specifically on data collection rather than final product development.

Personal robots, if they eventually materialize, would extend this ambition directly into people’s homes. Altman’s stated aspiration still leaves major open questions around actual capabilities, safety validation, manufacturing logistics, and realistic timing largely unresolved. A stated commitment to eventually build humanoids gives potential customers relatively little concrete basis yet for determining when such a machine might reliably handle their specific everyday tasks.

Growing Competition in Humanoid Robotics

A fully realized version of OpenAI humanoid robots would also place the company in more direct competition with existing players like Figure and Tesla. Figure has already described its own integrated learning and control approach in detail. Tesla’s current Optimus product page, observed as of September 6, describes a general-purpose autonomous humanoid robot, specifically identifying balance, navigation, perception, physical interaction, controls, and mechanical engineering as essential areas of ongoing work.

That overlap establishes clearly competing ambitions across the industry, though it doesn’t necessarily establish comparable levels of technical readiness between the various companies involved.

As OpenAI humanoid robots development continues progressing, any eventual official announcements from the company will likely need to demonstrate genuine evidence of reliable physical execution alongside strong underlying model capability. Metrics like task completion rates, frequency of human intervention, equipment reliability, and the overall usefulness of collected training data will likely prove essential in showing whether owning physical hardware truly delivers the control and learning advantages OpenAI’s broader strategy seems to be betting on.