Building the compute infrastructure for the Intelligence Age — Stargate update
Key Points
- Stargate surpassed its 10 GW U.S. goal ahead of schedule
- GPT‑5.5 trained at Abilene on OCI using NVIDIA GB200 systems
- Closed-loop cooling minimizes ongoing water use; community and workforce investments underway
Summary
OpenAI’s Stargate program is accelerating compute capacity to meet surging AI demand. The program has already exceeded the original 10 GW U.S. target (set in Jan 2025), adding more than 3 GW in the last 90 days. Stargate is partner-centric, working with cloud providers, data center operators, chipmakers, utilities, trades unions, and local communities to bring capacity online quickly while emphasizing local benefits and responsible resource use.
Key Points
- Capacity and pace: Surpassed the 10 GW U.S. commitment ahead of schedule; >3 GW added in the past 90 days.
- Partner and ecosystem model: Projects delivered via partnerships (cloud, data center, chip, energy, construction, finance, skilled trades) to reduce execution risk and preserve flexibility.
- Abilene site specifics: Flagship Stargate site in Abilene, TX runs on Oracle Cloud Infrastructure with NVIDIA GB200 systems; GPT‑5.5 was trained there.
- Water and cooling: Uses closed-loop cooling; initial fill ≈ 2 Olympic pools per building; ongoing annual water use at full build comparable to a medium-sized office (~4 households).
- Community & workforce investments: Launched community engagement (donation to Port Washington‑Saukville Education Foundation with Vantage and Oracle) and partnerships with NABTU for skilled-trades workforce development.
- Strategic implications: Focus on capacity coming online at scale, on time, and with flexibility as hardware, energy, and demand evolve.
Implications for engineers
- Expect increased on‑prem/colocated/OCI capacity and more GPU availability (NVIDIA GB200 families) for model training and inference.
- Plan workloads for scale and cost improvements over time; anticipate variability in site constraints (power, transmission, permitting, workforce).
- Consider operational constraints: cooling design (closed‑loop), one-time water fills, energy procurement and local permitting when scheduling deployments.
Actionable takeaways
- Prepare to leverage growing OCI/NVIDIA GB200 capacity for large-model training and evaluation.
- Coordinate with infra, procurement, and facilities teams early for power/water/permitting needs on new sites.
- Engage with partner and community programs where possible to align deployment timelines with local workforce and regulatory realities.