Oracle’s Reported $300B OpenAI Deal Needs Lenders

Oracle once added Nvidia A100 support to its cloud catalog and revised the price list. In November 2026, it filed to raise $20 billion through stock and bonds to pay for chips and data centers. The company still sells compute; the thing it must deliver has changed.

In 2020, Oracle could present a new accelerator as a cloud-product update. By September 2025, OpenAI had reportedly agreed to buy $300 billion of Oracle computing power over roughly five years, a commitment requiring 4.5 gigawatts of capacity beginning in 2027. Neither the reported contract value nor the power requirement means that the sites are operating today.

Competition in AI infrastructure has become a financing business: cloud providers must bind customer commitments to chips, energized buildings and construction capital, then decide who carries the risks that a cluster arrives late, sits idle or ages before its contracts do. The reported Oracle–OpenAI agreement makes that change unusually visible, but banks, data-center developers, chipmakers and other cloud buyers are making the same calculation.

Oracle’s instance catalog became a construction schedule

Oracle’s September 2020 cloud-compute update offered A100 support alongside Arm-powered processors and new pricing schemes. It answered the question cloud vendors had built their catalogs to answer: what equipment can a customer rent, and at what price? The provider owned the machinery, while the customer could buy capacity without committing to a building.

AI clusters changed the scale of that promise. Oracle was reported in May 2025 to be spending about $40 billion on roughly 400,000 Nvidia GB200 chips for a planned 1.2-gigawatt Texas data center serving OpenAI. The Texas chip-and-site plan joined a hardware purchase to a particular customer and a particular power requirement. Oracle could not make the contracted service useful merely by adding a machine type to a menu.

OpenAI has other suppliers and architectures to choose from. It agreed to buy 750 megawatts of Cerebras capacity over three years in a deal reported at more than $10 billion. AWS plans to add two million Nvidia Blackwell Ultra, Rubin and Rubin Ultra GPUs to its fleet in 2027 and 2028. Those plans put established cloud operators, specialist chip companies and customers that buy from several of them in different positions along the same supply chain.

Demand is not hypothetical. Oracle’s cloud revenue increased 27% year over year, and its reported fourth-quarter revenue reached $15.9 billion, up 11%. OpenAI said it had signed contracts for 10 gigawatts of US AI compute capacity, including more than three gigawatts added in the preceding 90 days. Current growth helps explain why Oracle is pursuing so much capacity. It does not establish when each contracted megawatt will be ready or what Oracle must pay while it waits.

A GPU cannot fulfill a contract without an energized site

A cloud vendor can procure chips before a developer finishes the building that will hold them. The developer can finish a building before the grid connection is ready. Each missing complement keeps the others from producing the compute a customer reserved. The AI infrastructure bottleneck therefore has an address: a site awaiting power, construction or equipment, rather than an abstract shortage of GPUs.

Meta’s Project Walleye in Ohio sought $3 billion in loans structured to fund both the data-center building and its power assets. The proposed building-and-power financing, described as a first-of-its-kind arrangement, recognized that lenders cannot assess delivery by looking at the server room alone. PJM likewise reduced its summer-2027 peak-demand forecast from about 164 gigawatts to about 160 gigawatts because some projects, including data centers, lacked firm service or construction commitments. A request for power was not yet a finished load.

Local decisions can interrupt that sequence. Digital Realty, QTS and NTT Data warned that more than 24 US data-center projects had been blocked in January 2026 alone. That count does not measure how many gigawatts were lost, but it identifies another decision a procurement team cannot settle by ordering chips. Equinix’s planned $5 billion data center near London faced a local challenge as the UK changed land rules to encourage construction.

Delivered capacity occupies a different position. Digital Realty agreed to acquire a majority stake in three fully leased Northern Virginia data centers from Blackstone-managed funds in a $7.8 billion deal. The price alone cannot establish a premium over unfinished projects; the distinction is that a buyer can inspect operating, contracted assets instead of financing a promise to complete them. Data-center developers and operators bring buildings, sites and power relationships to a cloud market that once emphasized the software interface above them.

Customers’ reservations bring lenders to the site

A long-term customer commitment gives a developer expected revenue to present to a lender before the facility opens. Banks were marketing more than $56 billion in investment-grade construction loans tied to Oracle’s future leases. Oracle also moved $66 billion of AI-data-center construction debt off its balance sheet through special-purpose vehicles. Those are distinct financing arrangements, not figures that can simply be added into one project total. Oracle’s November 2026 filing to raise $20 billion—$10 billion in common stock and $10 billion in bonds—for data-center construction and chips shows that it also sought capital directly.

A lease can help a lender fund construction, but moving debt into a vehicle does not by itself reveal who ultimately pays if a project is delayed or its customer leaves. Oracle’s reported capacity deals do not reveal who pledges collateral, who guarantees payment, who has recourse or cancellation rights, or who bears residual-value risk. Those terms determine how much risk each financing structure actually transfers. The broader project-bankability question turns on them, not on the size of the announced contract alone.

OpenAI has begun making a similar bargain with its own customers. Its Guaranteed Capacity offering discounts compute access in exchange for one- to three-year spending commitments. OpenAI gains a more predictable claim on future customer spending; customers gain assured access at the cost of some flexibility. The contracted megawatt can support financing at one end of the chain only because another party has accepted an obligation at the other.

Nvidia proposed a different form of support for younger cloud providers: it said it would rent back unused GPUs in exchange for a share of their cloud revenues. That arrangement gives a provider a potential customer for otherwise idle hardware while giving Nvidia a claim on future utilization. It does not remove the cost of constructing an energized site. It allocates one risk—the chance that purchased chips cannot immediately find renters—to a chipmaker with an interest in getting them deployed.

Lenders distinguish a leased building from a proposed one

Data-center developers began seeking credit ratings while projects were still under construction, as ratings agencies expanded their coverage. A rating can open a project to more buyers of debt; it cannot finish the project. Capital markets have also assigned very different prices to different claims on AI capacity. A CoreWeave-tied data center raised $1.1 billion in junk bonds yielding 9.25%, about 2.7 percentage points above similarly rated debt. That spread does not isolate which risk investors priced most heavily, but it cautions against treating every reserved cluster like a fully leased, operating building.

CME Group and Silicon Data announced futures contracts based on daily benchmarks for on-demand GPU rental rates. A provider exposed to changing rental prices may eventually use such a contract to hedge part of that exposure. A daily on-demand benchmark, however, does not settle whether a long-term customer may cancel, whether construction finishes on schedule or what an older chip will fetch after a lease ends. The hedge and the physical asset live on different clocks.

Creditors are watching even the largest cloud balance sheets. The cost of insuring against debt defaults at Amazon, Alphabet, Microsoft, Meta and Oracle reached its highest level since 2021. That measure does not establish that any one buildout will fail. It does show that creditors are paying attention to obligations cloud companies must fund before all the associated revenue arrives.

A long contract can outlive the hardware—or the permission to use it

Microsoft reportedly began turning off rented Oracle capacity after declining to renew a 2023 arrangement because it used older Nvidia chips. Microsoft’s nonrenewal is evidence about a particular hardware vintage, not evidence that buyers no longer want current-generation capacity. It exposes a problem that a construction loan cannot solve: a building may remain useful for years while the equipment purchased for an earlier customer loses that customer’s business.

Policy can break the expected use of capacity for a different reason. Oracle reportedly scrapped a roughly $7 billion, five-year agreement to lease as many as 100,000 advanced AI chips to Tencent after pressure from the Trump administration. The reported Tencent cancellation concerns a specific cross-border customer, not a general ban on AI construction. For a financing plan built around future rent, though, a customer lost to policy presents much the same immediate question as one lost to newer hardware: who owes the next payment?

Construction timing raises a third version. Oracle reportedly sent Blue Owl a force-majeure notice that could delay payments on the 2.45-gigawatt Project Jupiter in New Mexico if it fails to launch in 2028. The notice does not establish that the site will miss that planned date. It shows why a clause about delay can be as consequential to a developer and its lenders as a clause about price.

In the Oracle coverage record, regulatory framing rose 8.2 percentage points while enterprise framing fell 8.1 points across the 2024–2026 comparison periods. That measures a change in coverage, not the probability that a project will be blocked. The contracts remain the better place to look for an answer: which party keeps paying when a grid connection is late, a chip ages or a customer can no longer take delivery?

Oracle’s 2020 price list needed a machine customers could rent. Its OpenAI commitment needs chips inside a powered building—and a lender willing to wait for the lights to come on.