Stargate Abilene — OpenAI × Oracle × SoftBank's 1.2 GW Flagship

1 · The unusual corporate structure

Stargate is not an OpenAI project. It's a Delaware joint venture (Stargate LLC) with four equity owners: OpenAI, SoftBank, Oracle, and Abu Dhabi investment firm MGX. The initial capital commitment is $500B by 2029, equivalent to more than the total market cap of every data center REIT combined.

The Abilene site itself is built by Crusoe Energy Systems, the Denver-based neocloud known for running GPU clusters off stranded natural gas in West Texas oilfields. It's financed by Blue Owl Capital in a JV with Primary Digital Infrastructure. Oracle leases the completed campus for 15 years and hosts OpenAI's workloads on Oracle Cloud Infrastructure (OCI).

2 · Why Abilene?

Three reasons: stranded wind power, fast permitting, Texas sovereignty.

• Wind surplus. West Texas has some of the cheapest wholesale power in the US — ERCOT wind + gas peakers frequently clear below $0.04/kWh. The Lancium Clean Campus is specifically designed to consume excess wind generation before it's curtailed.
• ERCOT's isolated grid + fast permitting. Texas's grid operates independently of the federal reliability regime (FERC). Site selection and power contracts move in months, not the 5+ years typical of PJM or CAISO.
• Political fit. Trump announced the project from the White House with explicit promises of "emergency declarations" to expedite development. Texas was the obvious political home.

3 · 450,000 GB200 GPUs

Abilene is the largest GB200 NVL72 deployment publicly disclosed. 450K Blackwell GPUs on a 15-year Oracle lease means roughly 6,250 NVL72 racks, consuming ~750 MW of IT power alone (120 kW/rack × 6,250). The rest of the 1.2 GW envelope covers cooling, networking, and power overhead.

All buildings are designed for both liquid and air cooling: DLC for GPU racks, air for networking/storage/head-node infrastructure. The first two buildings span 980,000 sq ft and hold 200+ MW of IT load.

Source: Crusoe Energy Systems press release (Sept 2025) · OpenAI announcement blog (Jan 2025) · Wikipedia "Stargate LLC" (maintained from filings).

4 · The 1.2 GW cap (March 2026)

Oracle and OpenAI had been negotiating to expand Abilene to 2.0 GW. In March 2026 they publicly walked away: additional grid capacity would not be available until 2027+, and neither party would wait. Microsoft stepped in and leased the remaining 900 MW of the adjacent Lancium tract — meaning OpenAI and Microsoft are now literal neighbors pursuing independent compute strategies (despite Microsoft's ~27% stake in OpenAI).

The strategic pivot: OpenAI is redirecting all new Stargate capacity toward NVIDIA's next-generation "Vera Rubin" platformacross 5 newly-announced sites. Abilene remains the GB200 flagship; Rubin sites will handle the next 7+ GW of the 10 GW commitment.

5 · Lessons from Stargate Abilene

• Utility interconnect is the hard ceiling. Abilene went from 1.2 GW → planned 2 GW → capped back at 1.2 GW because the grid couldn't deliver on any timeline OpenAI would accept.
• The JV-lease-build triangle is the new template. A tech customer that can't own the capex, a cloud tenant with a balance sheet, a builder-operator, and a private-equity financier. All four Stargate expansion sites now use variations of this structure.
• Silicon roadmaps drive site strategy. The GB200-Rubin generational shift is why the 8 buildings stop at 1.2 GW — the next generation wants purpose-built power + cooling, not retrofit.
• Stranded renewables matter. Abilene exists because Lancium's site was engineered to consume curtailed West Texas wind. The clean-energy framing is also political cover for natural-gas backup.