Stargate Data Center Explained: Who Builds It, Who Gets Paid

EXPLAINED · Mechanism · Last reviewed: 9 Nov 2026 · Next review: after Oracle’s second-quarter fiscal 2027 results in December 2026 and OpenAI’s next Stargate site update

Stargate is OpenAI’s programme for getting the AI data center capacity it needs without owning most of it: developers such as Crusoe, Vantage and SoftBank’s SB Energy build and power the campuses, Oracle or SB Energy leases or runs them, NVIDIA supplies the chips, and OpenAI pays as the capacity goes live. OpenAI said in April 2026 that it had already secured more than the 10 GW it promised in January 2025, but at the flagship campus in Abilene, Texas, Oracle said in September 2026 that 618 MW had been delivered, so the gap between signed and live is the part to watch.

Why an investor cares: Stargate is the biggest single source of new demand for AI data centers, and it runs through other companies’ balance sheets. Whether a landlord, a cloud provider or a chip maker gets paid on time depends less on the headline gigawatts than on when each hall is handed over. This page builds on what an AI data center is and which AI data centers are the biggest, and answers one question: what is Stargate, and who actually gets paid?

What is Stargate?

Stargate started as a company. On 21 January 2025, OpenAI announced The Stargate Project, a new company that intended to invest $500 billion over the next four years building AI infrastructure for OpenAI in the United States, with $100 billion to be deployed immediately. The equity funders were SoftBank, OpenAI, Oracle and MGX. SoftBank took financial responsibility and OpenAI operational responsibility, with Masayoshi Son as chairman. Arm, Microsoft, NVIDIA and Oracle were named as technology partners.

It has since become a label for a programme. By September 2025, OpenAI was calling Stargate its “overarching AI infrastructure platform”, and the sites it added were deals with partners rather than buildings owned by the Stargate company. In April 2026 OpenAI described Stargate as its long-term effort to build the compute it needs, said it had already passed the 10 GW US goal, and noted that “the financing models and partnership structures may evolve”.

So the useful definition is functional. Stargate is the set of contracts through which OpenAI gets computing capacity that someone else builds, powers and often owns. That makes it closer to a very large tenant than to a landlord, which is why it matters so much to the companies in our AI Data Centers layer.

How does Stargate work?

Each Stargate campus is a stack of separate deals. The names change from site to site, but the roles repeat.

• OpenAI is the customer. It signs for capacity it expects to need for training and for running its products. At PORTS-Pike in Ohio, OpenAI says it “will begin paying only as completed capacity becomes available for lease”, funded by revenue and money raised from investors.
• A cloud operator or landlord sits in the middle. At the Oracle sites, Oracle leases the campus, buys and owns the chips, and sells the computing to OpenAI through Oracle Cloud Infrastructure. At PORTS-Pike there is no cloud layer: SB Energy builds, owns and operates the data center and leases it to OpenAI for 20 years. Our Oracle Stack page covers that business model.
• A developer builds and powers the shell. Crusoe developed Abilene, Vantage is Oracle’s partner in Wisconsin, and SB Energy provides “powered infrastructure” in Milam County, Texas. See what Crusoe does and who builds AI data centers.
• NVIDIA supplies the chips and, increasingly, the credit. Abilene runs NVIDIA GB200 systems. At PORTS-Pike, NVIDIA will be the only compute supplier, will invest $1.5 billion in SB Energy and will give credit support for the land, power and shell behind the first 4.25 GW-IT.
• A utility connects it. Project Camellia in Georgia has contracted 3.2 GW from Georgia Power, delivered in phases from 2028 to 2032, and OpenAI says it will pay the full cost of the infrastructure needed to serve it. Our contracted power explainer shows how those agreements work.

Stargate sites and partners

Here is what OpenAI and its partners have said about each campus, with who builds it, who uses it and where it stands.

Sizes mix grid power, IT power and utility contracts, so they should not be added up. Status lines are from OpenAI, Oracle, Epoch AI, NEXTDC and Firmus documents dated January 2025 to September 2026. Full list under References.

Two numbers in OpenAI’s own announcements give a rough price. Its January 2025 plan was $500 billion for 10 GW, about $50 billion per gigawatt (our arithmetic). Its September 2025 update was over $400 billion for nearly 7 GW, about $57 billion per gigawatt (our arithmetic). Those totals bundle chips, buildings and power. Our page on what an AI data center costs to build splits the bill.

Who gets paid first, and who pays last?

Follow the cash in time order. The companies at the bottom of the stack are paid before the building runs; the ones at the top are paid only after it does.

• Chip and equipment makers, first. At Abilene the chips are owned by Oracle, so NVIDIA and the server and gear suppliers are paid when the hardware ships, long before OpenAI’s usage pays it back. Oracle spent $28.5 billion on capital expenditure in the three months to 31 August 2026, against $8.5 billion a year earlier. Our page on who makes data center power equipment covers the electrical side of that bill.
• Developers and landlords, when each lease starts. Oracle reported $288 billion of additional lease commitments at 31 August 2026, “substantially all related to data center arrangements”, expected to start between its second fiscal quarter of 2027 and fiscal 2029, for terms of 15 to 19 years. That is about 4.6 times the $63 billion of lease payments already on its books (our arithmetic, operating plus finance leases). Those signed-but-not-started leases are the developers’ future rent. Our data center lease explainer shows how that rent works.
• Oracle, as it delivers. Oracle’s remaining performance obligations, the contracted revenue it has not yet earned, were $664 billion at 31 August 2026, up from $455 billion a year earlier. OpenAI says its partnership with Oracle exceeds $300 billion over five years. If all of that sits in the backlog, it would be about 45% of it (our arithmetic); Oracle does not give the split. See our RPO explainer.
• OpenAI, last. OpenAI pays as capacity goes live, from its revenue and from the money it raises. That is the point of the structure: the partners carry the build, and OpenAI’s bills start when the halls do.

The order explains who carries the risk in the gap. In the same quarter, Oracle’s operating cash flow was $23.1 billion, of which $11.4 billion came from customer prepayments, about 49% (our arithmetic). Its free cash flow was minus $5.4 billion, and it sold $20 billion of new shares through an at-the-market programme. At its Investor Day on 28 October 2026, Oracle said capital spending peaks in fiscal 2027 and 2028, free cash flow turns positive in the second half of fiscal 2028, and it will limit further money raised through fiscal 2030 to $20 billion, split between equity and debt, leaning on vendor financing, customer-supplied hardware and prepayments. Our Analysis on Oracle and the power stuck step works through what that means.

Where is the stuck step?

At delivery: the moment a hall has power, cooling and chips, and is handed to the customer. Everything before that is a promise; everything after it is rent and revenue.

Abilene shows how long that takes even at the fastest site. On 10 September 2026, Oracle said six of the eight campus buildings, 618 MW and 75% of total capacity, had been delivered to the customer. That implies about 824 MW for the whole campus, so about 206 MW, two buildings, was still to come (our arithmetic). Oracle gave no date for the last two on that call.

Oracle’s backlog shows the same gap at company scale. Of the $664 billion, Oracle expects to recognise about 13% as revenue in the next 12 months, 37% in months 13 to 36, 34% in months 37 to 60 and the rest after that. In dollars that is roughly $86 billion within a year and about $578 billion later (our arithmetic). The contracts are signed; the halls that turn them into revenue are not built yet.

The newer sites push the clock further out, and power is the reason. At PORTS-Pike, OpenAI expects the first 800 MW in 2028, using existing AEP grid infrastructure. That is about 10% of the planned 8 GW-IT (our arithmetic). The rest needs new power plants, including natural gas, and new transmission lines, and depends on permits, environmental reviews and financing. Project Camellia’s power arrives between 2028 and 2032. Our Wire on Oracle’s force majeure notice on Project Jupiter and our page on how long an AI data center takes to build show why the building is the quick part.

Our view: the useful question about any Stargate headline is not how many gigawatts were announced, but how many megawatts have been delivered, and to whom the bill goes in the meantime. A gigawatt that is secured but not delivered pays the chip maker and the builder; it does not yet pay the landlord’s rent in full, or Oracle’s cloud revenue.

Is Stargate in Asia? Australia, Malaysia and Singapore

Australia. On 5 December 2025, NEXTDC told the ASX it had agreed a memorandum of understanding with OpenAI under the OpenAI for Countries programme. The two will collaborate on the planning, development and operation of a hyperscale AI campus and GPU supercluster at NEXTDC’s S7 site at Eastern Creek in Sydney, with OpenAI intending to be an initial offtaker. The first phase is expected in the second half of 2027, subject to approvals. A memorandum is not a lease, so in the language of this page it is neither secured nor delivered yet. See what NEXTDC does and our Australia Pulse on the bid for NEXTDC.

Malaysia. On 8 September 2026, Firmus said OpenAI will contract dedicated AI computing from two Firmus AI Factory sites in Malaysia and become an anchor customer. Firmus did not say how many megawatts are OpenAI’s. Its more than 900 MW of contracted capacity covers all customers, across seven AI factories in Australia, Singapore, Indonesia and Malaysia, of which two are operating and five target ready-for-service within 24 months. Our Wire on Firmus pulling its ASX listing and Malaysia Pulse on TNB’s new power plants show the delivery side.

Singapore. None of the documents we read names Stargate capacity for OpenAI in Singapore. The closest link is Firmus, which already runs one of its two operating AI sites there for its customers. Singapore’s data center capacity is rationed by government allocation, which makes it a hard place to land a gigawatt-scale campus; see what Nxera does and our Singapore door.

What would change this view?

• Faster delivery. If Abilene’s last two buildings and the first Shackelford and Wisconsin halls arrive on time, more of Oracle’s $288 billion of leases start, and the stuck step moves from the building to the power plant.
• A change in who carries the build. OpenAI says the financing models may evolve. More deals like PORTS-Pike, where a developer owns the site and NVIDIA backs the credit, would shift risk away from Oracle and toward developers and suppliers.
• OpenAI’s ability to pay. The whole chain rests on the customer at the top. OpenAI says it will fund its commitments from revenue and investor capital; slower growth or harder fundraising would show up first in delayed or resized sites.
• Memoranda turning into leases. A binding agreement at NEXTDC’s S7, or a disclosed OpenAI megawatt figure at Firmus, would move the Asia sites from intention to secured.

How to read a Stargate headline

• Secured, delivered or running? OpenAI’s 10 GW is secured. Oracle’s 618 MW at Abilene is delivered. They are different stages.
• Grid power or IT power? PORTS-Pike is stated in GW-IT, Camellia in grid power from Georgia Power. The two are not the same unit; see how much power an AI data center uses.
• Lease, capacity contract or memorandum? A 20-year lease, a cloud contract with Oracle and a memorandum with NEXTDC carry very different certainty.
• Whose balance sheet? Find out who buys the chips, who builds the shell and who lends against it. That company carries the gap until delivery.
• All customers or OpenAI only? Oracle’s $664 billion and Firmus’s 900 MW include other customers.

FAQ

Related: what is an AI data center; the biggest AI data center in the world; data center financing; take-or-pay GPU contracts; where AI data centers are located; sovereign AI; Oracle’s 850 MW quarter.

Stack doors

Layer: AI Data Centers. Also: Hyperscalers, Neoclouds, Semiconductors. Markets: US, Australia, Malaysia, Singapore.

Get the Wire — short daily notes when a Stargate hall is delivered, a lease starts or a site slips.

References