Abilene brought online for GPT-5

This marks the moment Stargate stopped being a financing story and became a live compute asset for OpenAI. By mid 2025, Abilene was not just under construction, it had energized buildings, installed first GB200 racks, and was already running frontier workloads. That means OpenAI was using owned and partner controlled infrastructure for both model training and serving, instead of waiting for a fully finished campus before putting capacity to work.1 2 3 4

• Abilene moved unusually fast for a hyperscale AI site. Construction started in June 2024, the first phase covered two buildings and 200 plus megawatts, and Oracle had begun delivering Nvidia GB200 racks by June 2025. That is the hardware layer behind the early GPT-5 work.2 3 4
• Partially operational means a staggered ramp, not a ribbon cutting. OpenAI and Crusoe both describe parts of the facility as up and running before the full campus was complete, with early training and inference already live. In practice, some clusters were usable while more buildings, power, and networking were still being added.1 2 3
• The strategic point is vertical control of scarce compute. Crusoe builds the campus, Oracle operates the cloud layer, Nvidia supplies the systems, and OpenAI consumes the capacity for frontier research. That setup looks less like ordinary cloud renting and more like a dedicated AI factory built around one customer’s roadmap.1 2 4 5

Going forward, this staged bring up becomes the template. OpenAI can start training and serving on each new site before the whole campus is finished, which shortens the gap between capex commitments and usable model capacity. That is how Stargate compounds into a real moat, faster model iteration, more reliable inference, and less dependence on any single cloud provider.1 3 5