OpenAI Abandons Stargate Data Center Plans, Shifts to Rented Computing

When President Trump stood alongside OpenAI CEO Sam Altman and SoftBank's Masayoshi Son in January 2025 to announce Stargate — a $500 billion AI infrastructure initiative they called "the most important project of this era" — the vision was unmistakable: America would build its way to AI dominance, one massive data center at a time [1].

Fourteen months later, that vision is cracking. OpenAI has walked away from expanding its flagship Stargate campus in Abilene, Texas, abandoned plans to build its own data centers entirely, and is instead renting computing power from cloud providers including Amazon Web Services, Google Cloud, and CoreWeave [2][3]. The pivot represents one of the most significant strategic reversals in the AI industry's brief but explosive history — and raises hard questions about whether the economics of AI infrastructure were ever as solid as the hype suggested.

The Abilene Collapse

The first visible fracture appeared on March 6, when Bloomberg reported that Oracle and OpenAI had ended plans to expand the Abilene data center from 1.2 gigawatts to approximately 2.0 gigawatts [4]. The expansion, which would have been developed in partnership with data center operator Crusoe, had been under discussion since mid-2025.

The reasons for the collapse were multiple and compounding. Financing proved elusive — lenders were unwilling to back billion-dollar construction projects for a company still carrying heavy operating losses [5]. OpenAI's demand forecasts shifted repeatedly, frustrating Oracle and Crusoe. A multi-day outage caused by winter weather that damaged Crusoe's liquid-cooling infrastructure strained relations further [6]. And critically, power at the expansion site would not be available for at least a year, by which point OpenAI hopes to deploy Nvidia's next-generation Vera Rubin chips rather than the Blackwell GPUs currently going into Abilene [7].

Oracle pushed back against reports characterizing the situation as a cancellation. The company confirmed that its existing 1,000-acre, eight-building campus continues to operate, with two buildings already running training and inference workloads on Nvidia GB200 Blackwell racks and the remaining six scheduled for completion by mid-2026, bringing total capacity to roughly 1.2 gigawatts [8]. Oracle also insisted that its broader 4.5-gigawatt, $300 billion agreement with OpenAI remains on track.

But the damage was done. Oracle's stock, already under pressure, sits 56% below its 52-week high of $344.21, trading around $151 as of mid-March [9]. RBC Capital cut its price target, and analysts flagged that Oracle carries $131.7 billion in total debt with negative levered free cash flow of $13.2 billion [10]. Blue Owl, Oracle's largest Stargate financing partner, had already declined in late 2025 to back the 1-gigawatt Michigan Stargate campus, citing less favorable lease terms and rising leverage concerns [9].

From Owner to Renter

The Abilene episode is a symptom of a deeper strategic shift at OpenAI. The company spent much of 2025 trying to build its own data center empire, only to discover that the capital requirements were staggering and the financing terms punishing. Rather than owning physical real estate, OpenAI has pivoted to what amounts to a "renter" model — controlling what goes inside the data centers while leaving the construction risk to others [5].

The numbers tell the story of this diversification. OpenAI has assembled a web of cloud partnerships that would have been unimaginable even two years ago:

• Amazon Web Services: A seven-year, $38 billion deal providing access to hundreds of thousands of Nvidia GPUs, with full capacity targeted for deployment by the end of 2026 [11]. This was later expanded by $100 billion over eight years, with AWS becoming the exclusive third-party cloud distribution provider for OpenAI's enterprise platform.
• Microsoft Azure: A $250 billion purchase commitment established during a 2025 restructuring, though OpenAI ended its Azure exclusivity to work with competitors [12].
• Oracle: The $300 billion, five-year agreement for 4.5 gigawatts of capacity, though its execution is now under scrutiny [8].
• CoreWeave: A $22.4 billion deal for dedicated GPU infrastructure from the specialized AI cloud provider [12].
• Google Cloud Platform: A deal announced in June 2025 to support ChatGPT's global expansion, with financial terms undisclosed [12].

In total, OpenAI has committed more than $1 trillion in hardware and cloud infrastructure agreements over the next decade, with current cloud commitments approaching $600 billion — down from an earlier $1.4 trillion figure that reflected a more ownership-heavy strategy [12].

On March 16, OpenAI reorganized its Stargate computing initiative into three groups covering technical design, commercial partnerships, and facility management, with former Intel AI chief Sachin Katti among the new leaders appointed to oversee the restructured infrastructure teams [3]. The reorganization signals that Stargate is not dead but is being fundamentally reimagined — less a constellation of OpenAI-owned mega-campuses and more a managed network of rented capacity.

Meta Swoops In

One company's retreat is another's opportunity. Meta Platforms is now in talks to lease the capacity that OpenAI abandoned in Abilene, with Nvidia reportedly paying a $150 million deposit to Crusoe to secure the site and facilitate negotiations [6][7]. The deal, if completed, would underscore how the AI infrastructure gold rush is redistributing assets even as it stumbles.

Meta's interest makes strategic sense. The company plans to invest up to $135 billion in 2026 capital expenditures, primarily for GPU compute capacity [13]. Unlike OpenAI, which generates revenue primarily from API access and subscriptions, Meta can justify massive infrastructure spending against its $160 billion advertising business.

Nvidia's role as intermediary is also telling. The GPU maker has a vested interest in ensuring its chips get deployed regardless of which AI company occupies the data center — and reportedly intervened specifically to ensure Nvidia processors, rather than rival AMD chips, would power the expanded facility [6].

Source: Goldman Sachs / Futurum Group / Company Earnings Reports

Data as of Mar 17, 2026CSV

The $690 Billion Question

OpenAI's pivot unfolds against the backdrop of what may be the largest infrastructure buildout in human history. The five largest U.S. cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026, nearly double the approximately $380 billion spent in 2025 [14][15].

The individual figures are staggering: Amazon at $200 billion, Alphabet at $175-185 billion, Meta at $115-135 billion, Microsoft tracking toward $120 billion or more, and Oracle targeting $50 billion [14]. Worldwide spending on AI is forecast to total $2.5 trillion in 2026, a 44% increase over 2025, according to Gartner [15].

But the sheer scale of spending is creating its own problems. Power availability has emerged as the binding constraint, with data center construction timelines extending 24 to 72 months due to grid limitations [16]. PJM Interconnection, the largest U.S. grid operator serving over 65 million people across 13 states, projects it will be a full six gigawatts short of reliability requirements by 2027 [16].

The energy crisis triggered by the Iran war has compounded these challenges. Oil prices have surged from approximately $67 per barrel in late February to nearly $95, with the Strait of Hormuz effectively closed [17]. While data centers primarily run on electricity rather than oil, the broader energy price shock is rippling through the power market, raising costs for natural gas generation and increasing political pressure on data center power consumption.

Source: FRED / U.S. Energy Information Administration

Data as of Mar 9, 2026CSV

Residential electricity prices in the U.S. have already risen more than 36% since 2020, from 12.76 cents to 17.44 cents per kilowatt-hour, and are expected to hit 19.01 cents by September 2027 [16]. A growing political movement — spanning Senator Bernie Sanders to conservative state legislators — is pushing back against data center power consumption, with more than 230 environmental groups demanding a national moratorium on new construction [18].

What This Means for the AI Race

OpenAI's shift from builder to renter carries significant strategic implications. On the positive side, renting provides flexibility: the company can scale capacity up or down without being locked into specific hardware generations or geographical locations. In an industry where the next GPU architecture can render current infrastructure suboptimal within 18 months, this agility has real value.

But renting also means dependency. OpenAI is now reliant on AWS, Google, and Oracle — companies that are simultaneously its infrastructure providers and, in Google's case, a direct competitor in the AI model market. The company's compute costs will be subject to the pricing decisions of providers who have their own AI ambitions to fund.

The broader Stargate project is not dead, but it is smaller and more uncertain than the grand vision announced at the White House in January 2025. The $500 billion headline figure always depended on assumptions about financing, power availability, and demand growth that are now being stress-tested by reality. Oracle's debt load, the Abilene setback, Blue Owl's reluctance to finance the Michigan campus, and partner disagreements about control all point to a project that will likely be built incrementally rather than at the breathtaking pace originally envisioned [9][10].

For the AI industry as a whole, the Stargate stumble is a warning sign — not that artificial intelligence is failing, but that the physical infrastructure required to power it cannot be willed into existence by press conferences and press releases. The chips need power. The power needs grid connections. The grid needs investment. And all of it needs financing that makes economic sense for companies that, in OpenAI's case, still do not turn a profit.

The trillion-dollar AI infrastructure buildout will continue. But it will look less like a centrally planned moonshot and more like a messy, multi-party negotiation between companies with misaligned incentives, constrained power grids, and investors who are starting to ask harder questions about when — and whether — these massive bets will pay off.