OpenAI has signed a lease on a 202,371-square-foot industrial building at 1411 Harbour Way South in Richmond, California — a waterfront site spanning 25.9 acres with more than 14,000 amps of electrical power, a capacity built for battery manufacturing that is almost certainly overkill for servers and profoundly unusual for a warehouse. The site covers 25.9 acres according to Mercury News and property records; the San Francisco Chronicle, citing county property records, reported 15.75 acres at the same address — the gap likely reflects different definitions of what counts toward the total parcel. The company is moving into physical robotics manufacturing, and the infrastructure tells you how serious this is.
The lease was recorded on March 9, 2026, with OpenAI OpCo LLC as the signatory, according to county property records reviewed by the Grand View Independent. In December 2025, OpenAI told employees it planned to open a second robotics-focused facility in Richmond — its first East Bay footprint and the first known location outside San Francisco for a robotics operation the company has quietly been building since early 2025, the San Francisco Chronicle reported.
The power specification is the most technically revealing detail in the deal. Most industrial buildings offer between 200 and 4,000 amps of electrical capacity. This one offers more than 14,000 amps — a figure that begins to make sense only when you know what the site was built for. The Portside Commerce Center property was originally developed for Moxion Power, a battery startup that filed for bankruptcy in 2024 and never occupied the space. Moxion was designing high-density energy systems. The electrical infrastructure was engineered for that use case, not for a standard data center or warehouse. For a company planning to run robot assembly lines, battery testing, and AI compute simultaneously, that infrastructure is not a bonus — it may be the reason the site was selected.
OpenAI declined to comment on the lease or its plans. But the signals have been accumulating for more than a year.
OpenAI's move toward in-house robotics hardware began, by most accounts, with partnerships. The company worked with Figure AI, the Bay Area humanoid startup, starting in 2023. Figure CEO Brett Adcock said in February 2025 that the partnership was ending — the two companies had designed the arrangement to build next-generation AI models for humanoid robots, but the collaboration did not survive long enough to produce a commercial result, Business Insider reported. OpenAI then brought in Chengshu Li from Stanford University in June 2025, where he had worked on humanoid robotics benchmarks, Wired reported. Li now leads a robotics effort inside OpenAI that has grown substantially in the months since.
The scale of the San Francisco lab operation has been documented by Business Insider: OpenAI runs three shifts with roughly 100 data collectors teaching robotic arms via teleoperation — a labor-intensive process in which humans control robots in real time to generate training data for AI models. The lab has more than quadrupled in size since launching in February 2025. The data collectors use Franka robotic arms equipped with custom 3D-printed controllers called GELLOs. The work is not glamorous; it is the foundation on which any autonomous robotics capability will eventually be built. Richmond, by all indications, is the scale-up.
The company has also posted job listings through staffing agencies that list Richmond as the location — the clearest external evidence that the facility is not theoretical. And in a request for proposals posted publicly, OpenAI sought US manufacturing partners to supply what it called "critical inputs for advanced robotics (e.g., gearboxes, motors, power electronics)," according to the company's website. The document signals that whatever is being assembled in Richmond is meant to scale beyond a research lab.
The financial stakes are large and widely acknowledged. Morgan Stanley reckons the humanoid robot industry could be worth $5 trillion by 2050, Wired reported. Venture capitalists have invested more than $5 billion in humanoid startups since the start of 2024, the same outlet noted. OpenAI is entering a crowded competitive field that already includes Figure AI, Physical Intelligence, and 1X Technologies — Bay Area companies that have been building and deploying humanoid robots for months. What OpenAI brings that the others do not is the compute scale and the AI model development capability. Richmond is the physical infrastructure to match.
There is also the Stargate context. OpenAI has announced planned capacity that puts it "well over halfway to meeting our 10-gigawatt commitment", the company said in a supply chain statement — part of the broader infrastructure buildout that the Trump administration framed as a cornerstone of US AI competitiveness. Physical AI — AI systems that interact with and manipulate the physical world — has become a stated priority for an organization that built its reputation on language models. The Richmond lease is the real estate line item for that ambition.
The East Bay location is not incidental. Richmond offers waterfront logistics access, relative affordability compared to San Francisco or the Peninsula, and — because of the Moxion development — electrical infrastructure that most industrial buildings cannot match. JLL commercial real estate executives Jason Ovadia, Greg Matter, Michael Murray and Ada Wong have been seeking tenants for the property. For a company planning to build battery-powered robots at scale, the power may have been as important as the square footage. A conventional converted warehouse would have required new electrical service before it could support the kind of operation OpenAI appears to be planning. The Moxion site needed almost none of that work.
OpenAI has also been expanding its South Bay presence. The company has leased 439,000 square feet in Mountain View — its first major South Bay footprint — according to Mercury News and property records. The Mercury News listing puts the Mountain View offices at 439,000 square feet, capable of accommodating 1,800 to 2,200 workers; the San Francisco Chronicle, citing county property records, reported 450,000 square feet at the same address. Both figures derive from property records; the reason for the discrepancy is not immediately clear — it may reflect multiple leases, common area inclusions, or measurement differences.
What to watch next is straightforward: whether OpenAI can staff the Richmond facility at the pace its ambitions require. The teleoperation data collection operation currently running in San Francisco is a bottleneck for progress — more robotic arms generating training data means faster model improvement, assuming the data quality holds. Richmond gives them room to scale that operation significantly. Whether that translates into a robot that can do something useful outside a lab remains an open question. But the company has now signed a lease that says, in the clearest possible real estate language, that it intends to find out.
