Intro

The Abilene “Stargate” campus was billed as the flagship of a $500 billion joint‑venture to give the United States a sovereign AI‑compute backbone. Two data‑center buildings went live in September 2025 and a further six were slated for 2026, targeting roughly 1.2 GW of capacity. Yet, by March 2026 both Oracle and OpenAI announced that the planned expansion to 2 GW would not move forward, citing financing headwinds and shifting demand forecasts.

We break down the financing web—equity, cloud contracts, debt lifts, and the power‑supply bets that keep the chips running—so investors and engineers can see where the megawatts and the dollars really stand.

Key takeaways

  • Oracle‑OpenAI’s original 2 GW expansion was cancelled in March 2026; the site now sits at ~1.2 GW.
  • Nvidia is investing up to $3 bn for a 20‑30 % equity stake in power‑developer Lancium, locking in 1.2 GW for the Abilene campus.
  • OpenAI’s compute budget has been trimmed to $600 bn by 2030, with up to $300 bn earmarked for Oracle cloud services.
  • SoftBank is leaning on debt to help fund OpenAI’s expansion, seeking $40 bn in loans; Oracle is raising $50 bn in debt/equity and cutting staff to free cash.
  • Power delivery remains the chokepoint: Lancium has secured 4 GW on the Texas grid, but a state‑level suspension of new grid‑connection applications adds uncertainty.

1. Funding the Joint‑Venture: Equity, Cloud Deals and Debt

Stargate LLC was incorporated on January 21, 2025 as a joint venture between OpenAI, SoftBank, Oracle and MGX, with a declared asset base of US$100 bn in 2025. Ownership is split 40 % OpenAI, 40 % SoftBank, with Oracle and MGX holding the remainder. The venture’s public goal is to spend up to US$500 bn on AI infrastructure in the United States by 2029.

The financing mix is a patchwork:

Metric Original Plan (early 2025) Current Status (mid 2026) Source
Total capital pledged $500 bn joint‑venture budget $500 bn still advertised, but many projects delayed Source 2 (announcement) & Source 1 (cancellation)
Cloud spend by OpenAI $300 bn to Oracle cloud (2025‑2030) $300 bn still contracted, Oracle now raising debt to fund it Source 1 (cloud deal)
Oracle financing $50 bn debt/equity raise slated for 2026 $50 bn raise underway, accompanied by layoffs of thousands Source 1 (debt raise) & Source 1 (layoffs)
SoftBank financing $40 bn loan hunt for joint‑venture $40 bn loan pursuit ongoing Source 1 (SoftBank loans)
Equity stakes OpenAI operational, SoftBank financial responsibility (per Altman) Same split, but OpenAI has shifted toward more cloud contracts Source 2 (Altman quote)

OpenAI’s own compute budget has been trimmed to around $600 bn by 2030, down from a previous $1.4 tn pledge for 2033. The reduction aligns with the decision to park a large chunk of spend in Oracle‑run cloud services.

2. Power‑Infrastructure: Lancium, Nvidia and the 1.2 GW Lock‑In

AI training clusters are power‑hungry; the bottleneck has moved from silicon to megawatts. Nvidia, the primary silicon supplier, is buying a slice of the power supply chain. According to The Information, Nvidia has agreed to invest up to $3 bn in Lancium, with an initial $2 bn for roughly a 20 % equity stake and a further $1 bn contingent on grid‑connection milestones. The deal values Lancium at an enterprise value of about $10 bn and could push Nvidia’s ownership to ≈30 % if the second tranche is delivered.

Lancium’s power portfolio in Texas already includes 4 GW of contracted capacity, with another 15 GW of projects awaiting grid connection. Of the secured 4 GW, 1.2 GW is earmarked for the OpenAI‑Oracle Abilene campus—the core of Stargate 1. A separate 900 MW is slated for a data center Crusoe is building in Abilene for Microsoft, and QTS Data Centers has signed for 1 GW in Turkey, Texas. Another 1 GW is under construction in Childress by Crusoe.

Lancium, founded in 2018 in Houston, originally built demand‑response tech for Bitcoin mining before pivoting to AI in 2024. In July 2024 the company and Crusoe announced a multibillion‑dollar, 200 MW AI data‑center deal that forms the first phase of a 1.2 GW build‑out at the Clean Campus. The 1,000‑acre “Stargate 1” site is engineered for ≈400,000 Nvidia AI chips across eight buildings, each capable of supporting up to 50,000 GB200 NVL72 units.

Blackstone, a private‑equity firm, already owns ≈50 % of Lancium and has poured > $500 m into the developer. Lancium is eyeing an IPO in 2027 to further monetize its power assets.

3. Capacity Roadmap: From 1.2 GW to the Cancelled 2 GW Goal

When the campus first broke ground, the plan was to launch two buildings in September 2025, then add six more in 2026, reaching ≈1.2 GW. The venture later announced a 2 GW target for the Abilene site. By March 2026, Oracle and OpenAI announced the 2 GW expansion was cancelled due to financing challenges and OpenAI’s volatile demand forecasts.

A winter‑weather‑driven multi‑day outage that crippled liquid‑cooling equipment further strained the relationship between OpenAI and Crusoe. Timing worked against the expansion too: with roughly a year‑long wait before additional power comes online at Abilene (per reporting from The Information), OpenAI’s hardware roadmap will have moved past today’s Blackwell GPUs to the next‑gen Vera Rubin generation — silicon it would rather stand up at a fresh campus than retrofit into the old one.

Despite the setback, OpenAI’s compute‑scaling executive Sachin Katti emphasized the site's size: “Our flagship Stargate site is one of the largest AI data center campuses in the United States”, and added that the company is shifting new capacity to other locations, including a Wisconsin site where steel beams have just gone up.

4. Partner Dynamics and Recent Moves

  • Oracle operates the Abilene facilities on behalf of OpenAI. To fund the broader AI‑infrastructure push, Oracle is raising up to $50 bn in debt and equity in 2026 and planning thousands of layoffs to preserve cash flow for AI data‑center spend.
  • OpenAI now leans more heavily on cloud contracts; its $300 bn cloud commitment to Oracle remains a cornerstone of its compute budget.
  • SoftBank is also turning to debt to help fund OpenAI’s expansion, currently seeking $40 bn in loans.
  • Nvidia is not only the chip supplier but also a power‑provider investor through Lancium, hedging against electricity scarcity that could throttle the Vera Rubin rollout.
  • Meta is in talks with Crusoe to pick up the remaining capacity at Abilene after Nvidia’s $150 m deposit was placed to secure the site for a potential Meta lease.

5. Risks, Governance Gaps and Timeline Uncertainty

The power‑supply landscape in Texas has become a regulatory risk factor. Governor Greg Abbott ordered a suspension of new data‑center grid‑connection applications, putting projects like the pending 15 GW Lancium pipeline on hold. While Lancium already has utility‑level studies completed for many of those projects, the policy pause could delay power readiness for any future Stargate expansions.

Finally, the operational relationship between OpenAI and Crusoe remains fragile after the liquid‑cooling outage, and OpenAI’s leadership turnover (the departure of director of physical infrastructure Keith Heyde earlier in 2026) hints at possible strategic re‑orientation toward cloud‑based solutions rather than in‑house builds.

Conclusion

Stargate’s Abilene campus illustrates how the AI‑compute boom is now a financing puzzle as much as a hardware one. The megawatt‑level power contracts, equity stakes by Nvidia, massive cloud spend locked into Oracle, and SoftBank’s loan hunt collectively keep the lights on at a campus engineered for some 400,000 Nvidia chips. Yet the cancellation of the 2 GW expansion, power‑grid regulatory headwinds, and a shift toward newer chip architectures all suggest the next phase of the build‑out will migrate to newer sites—Wisconsin among them—rather than simply scaling the original Texas footprint.

Sources

This article was researched and fact-checked against the following sources: