OpenAI Secures 10GW U.S. AI Compute Capacity

Per a detailed blog post, OpenAI's Stargate program has secured 10 GW of U.S. AI data center capacity, a milestone the company originally set for 2029 and says it reached just over a year after the program launched in January 2025 (OpenAI, "Building the compute infrastructure for the Intelligence Age"). The blog states that more than 3 GW came online in the prior 90 days and identifies the first active training and serving site in Abilene, Texas, with additional Stargate sites under development in Texas, New Mexico, Wisconsin, and Michigan (OpenAI, "Stargate Community").

A press release from NVIDIA describes a strategic partnership and a letter of intent with OpenAI to deploy at least 10 GW of NVIDIA systems representing millions of GPUs for OpenAI's next-generation infrastructure (NVIDIA press release). The release says NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed, and that the first gigawatt of NVIDIA systems is targeted to come online in the second half of 2026 on the Vera Rubin platform (NVIDIA press release). The release includes quotes from Jensen Huang, Sam Altman, and Greg Brockman on the collaboration (NVIDIA press release).

Companies building large-scale AI compute capacity typically combine long-term capacity commitments, co-investment from hardware partners, and staged deployment plans, because procuring tens of thousands of GPUs and matching power and networking takes multi-year coordination. Industry-pattern observations: large-scale GPU rollouts commonly require synchronized logistics across chip suppliers, power utilities, and data center construction firms, and hardware partners often provide both equipment and long-term financing to align incentives.

Securing 10 GW of capacity is materially significant for practitioners because sustained access to GPU-scale compute affects training cadence, model size choices, and total cost of ownership. Observers tracking the compute supply chain will view a combined deployment and investment commitment from a major hardware vendor as reducing execution risk for large training programs, while also concentrating demand pressure on particular GPU architectures and supply chains. For practitioners: aggregated capacity at this scale generally lowers per-unit training cost and enables larger experiments, but it also raises operational requirements for power, cooling, and software orchestration.

Editorial analysis: Watch the deployment cadence and the first NVIDIA Vera Rubin systems coming online in H2 2026 as reported by NVIDIA, since early system performance, interconnect characteristics, and software stack integration will shape how quickly that capacity translates into larger training runs (NVIDIA press release). Also monitor local permitting and power-procurement details at announced Stargate sites, because the blog post states OpenAI will tailor community energy commitments site-by-site and, in some cases, fund incremental generation or transmission to avoid raising local electricity prices (OpenAI, "Stargate Community").

Editorial analysis: Practitioners should note that access to a steady multi-gigawatt fleet typically shifts tradeoffs toward larger distributed training jobs, more aggressive model scaling experiments, and increased emphasis on software for scheduling, checkpointing, and multi-node performance tuning. Industry-pattern observations: teams running at this scale commonly invest concurrently in systems software, data pipelines, and cost-allocation tooling to maintain throughput and utilisation across heterogeneous hardware as new platforms arrive.

The NVIDIA release includes direct quotes attributed to executives, including Sam Altman: "Everything starts with compute," and Greg Brockman: "Weve been working closely with NVIDIA since the early days of OpenAI," which the release frames as describing the partners' history and intent for co-optimization of hardware and software roadmaps (NVIDIA press release).

What is public is focused on planned and contracted capacity and partner commitments; there is no public, independently audited breakdown of the specific GPU models, exact timelines for each gigawatt beyond the first-phase target, or detailed cost terms in the NVIDIA release and OpenAI posts. Observers should treat the announced figures as headline commitments documented in OpenAI and NVIDIA materials.