OpenAI’s big new deal with Oracle for massive AI computing power

OpenAI seals the vast computing muscle it needs for its next-generation models.

OpenAI—the San Francisco research company behind ChatGPT and the GPT-4o family—has agreed to rent 4.5 gigawatts of new U.S. data-centre capacity from Oracle, locking in a bill that could reach about $30 billion a year once the build-out is complete in 2028. That amount of power equals the output of several nuclear reactors and secures the raw compute OpenAI considers essential for training successors to GPT-4o and for running ever-larger versions of its multimodal systems.

Oracle commits to constructing multiple high-density data centres under the Stargate banner.

To meet the contract, Oracle will expand an existing 1.2 GW campus in Abilene, Texas, and develop new sites in states such as Wyoming, Michigan and Georgia, all clustered around low-cost electricity and favourable permitting regimes. Internally named “Stargate,” the programme pools financing from OpenAI, Oracle, SoftBank and Abu-Dhabi-backed MGX and carries an overall investment target of $500 billion by 2029.

The 4.5-gigawatt figure dwarfs every previous cloud contract and illustrates an unprecedented scaling curve.

Until now, even hyperscale leases rarely exceeded a few hundred megawatts; jumping to multi-gigawatt territory signals that training frontier models has moved beyond boutique supercomputers and into an era of industrial-scale AI infrastructure. Analysts note that a single 1 GW campus can host roughly half a million cutting-edge GPU or AI-accelerator boards.

A projected $30 billion annual spend rewrites the math of recurring cloud revenue.

If the ramp-up proceeds as outlined, Oracle’s cloud-infrastructure line—which brought in about $10 billion in 2025—could more than triple, pushing overall corporate revenue beyond $100 billion early in the next decade. Industry watchers call the contract “net-new” rather than a renewal, meaning every dollar counts as fresh growth instead of replacement business.

Stargate’s joint-venture roots reveal how OpenAI is spreading risk and bargaining power.

OpenAI remains closely tied to Microsoft Azure, yet the Oracle pact, plus earlier deals with Google Cloud and specialist provider CoreWeave, shows a deliberate multi-cloud strategy designed to avoid single-supplier bottlenecks and secure better pricing on GPU clusters. The broad partnership also unlocks capital from SoftBank and MGX, reducing the cash OpenAI must raise on its own balance sheet.

Oracle’s share price reacts because investors see pure incremental cash flow.

News of the agreement sent Oracle stock up nearly 4 percent to fresh record highs, with trade-desk commentary emphasising the contract’s size and duration rather than short-term earnings beats or misses. Fund managers frame the deal as a multi-year visibility event that could lift Oracle’s growth rate above 50 percent in fiscal 2028.

OpenAI’s decision marks a calculated move beyond the comfort of Microsoft’s Azure umbrella.

While Azure still hosts ChatGPT’s live traffic, research leadership inside OpenAI argues that frontier-model training benefits from adding bespoke clusters with different networking fabrics and supply-chain footprints. By layering Oracle’s forthcoming GB200-based GPU farms on top of its existing Azure estate, OpenAI can iterate faster and hedge against component shortages.

Gigawatt-class clusters draw in Nvidia’s new accelerators and spur custom-chip talk.

Oracle has signalled intentions to order roughly $40 billion in Nvidia GB200 systems and to spend $25 billion on capital expenditure next year, a scale that may reshape GPU allocation across the industry. In parallel, OpenAI continues co-design work on proprietary ASICs, aiming to supplement third-party hardware with internally optimised inference chips.

Host states welcome the economic windfall while grappling with electricity demand.

Local authorities in Texas, Wyoming and Michigan highlight incoming construction jobs and long-term tax revenue, yet utility planners warn that data-centre loads approaching gigawatt levels will force upgrades to transmission lines and, in some cases, dedicated on-site generation.

________

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org