Who is paying for OpenAI’s data centers?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
OpenAI’s data centers are being paid for upfront by cloud providers, infrastructure funds, strategic investors and lenders, while OpenAI takes on the long-term contracts and expects future customers to cover the final bill.
The old answer, “Microsoft pays,” is no longer enough. OpenAI now spreads its demand across Microsoft, Oracle, AWS, CoreWeave and Google Cloud, while suppliers such as Nvidia, AMD, Amazon and Broadcom help finance or equip the build.
OpenAI does not need $750 billion sitting in cash to announce $750 billion of computing plans. Most of the immediate spending happens on partners’ balance sheets, then comes back through leases, cloud-service payments and long-term purchase commitments.
The largest disclosed cloud agreements already add up to roughly $710.4 billion. That puts Microsoft, Oracle, AWS and CoreWeave at the center of OpenAI’s near-term compute plan, even before every campus, chip order and power project is counted.
Oracle shows how the model works. It is borrowing and issuing equity to build capacity before collecting the full revenue, while customers involved in its largest AI contracts have prepaid for or directly supplied tens of billions of dollars of GPUs.
Stargate is not one fully funded $500 billion pot. It is a collection of project-level financings in which developers, property funds, banks, Oracle, SoftBank and equipment suppliers each fund a different layer.
Supplier financing creates a circular feature that deserves attention. Microsoft, Amazon and Nvidia invest in OpenAI, while OpenAI signs huge contracts to buy their cloud capacity, chips or related services.
OpenAI’s current business is growing extraordinarily fast, but it still cannot fund the planned build on its own. At roughly $24 billion of annualized revenue, the projected compute spending through 2030 is more than 31 times the company’s present revenue run rate.
The debt often sits one step away from OpenAI. Oracle bondholders, SoftBank’s banks, CoreWeave’s lenders and data-center investors advance the capital, while OpenAI’s contracts make the borrowing possible.
Tax incentives and utility investment can reduce costs or move some risk outward, but they are not the main source of funding. The direct build is still overwhelmingly financed by private companies, investors and lenders.
The plan works only if OpenAI becomes one of the world’s largest technology revenue platforms. If that happens, subscriptions, enterprise contracts, API usage, advertising and commerce will gradually repay the infrastructure network. If it does not, the losses will fall first on OpenAI’s shareholders and its most leveraged suppliers.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
Who is paying for OpenAI’s data centers?
Why is this question suddenly so important?
OpenAI’s data-center bill has now grown far beyond what its current business can fund by itself.
The Wall Street Journal recently reported that OpenAI expects to spend roughly $750 billion on computing through 2030, up from an earlier internal projection of around $600 billion. OpenAI, meanwhile, says it is currently generating about $2 billion in revenue each month, or $24 billion at an annualized rate.
Those figures belong to companies of completely different sizes. The planned compute spending is more than 31 times OpenAI’s current annual revenue. The money will be paid over several years rather than immediately, but the gap remains enormous.
The question has also become harder because OpenAI keeps changing how it gets its computing power. For years, the simple answer was Microsoft. OpenAI trained and ran its models on Azure, while Microsoft supplied investment capital and built the infrastructure.
That arrangement has widened into something much bigger. OpenAI now buys capacity from Microsoft, Oracle, AWS, CoreWeave and Google Cloud. It works with Nvidia, AMD, Amazon and Broadcom on chips. It is developing Stargate campuses with Oracle and SoftBank. It has also committed $20 billion to begin Project Camellia, a Georgia data center that OpenAI is designing and developing more directly.
The old answer, “Microsoft pays,” no longer explains what is happening. OpenAI has built a financing network in which dozens of companies and investors advance money against the expectation of much larger AI revenue later.
When we say “paying,” who are we actually talking about?
Paying for OpenAI’s data centers can mean funding construction, buying the chips, signing the cloud contract or absorbing the loss if OpenAI falls short.
Imagine that an OpenAI campus costs $15 billion. A property investor may pay for the buildings. Oracle may purchase the servers. Banks may lend against the project. OpenAI then signs a long-term agreement to use the computing capacity.
All four parties have “paid” in some sense, but at different moments.
The property investor and Oracle provide much of the immediate cash. The banks provide borrowed money. OpenAI promises years of future payments. OpenAI’s customers are then expected to generate the revenue that eventually covers those payments.
That is how OpenAI can announce hundreds of billions of dollars in infrastructure without holding hundreds of billions in its bank account. Most of the capital arrives through its partners’ balance sheets.
It also tells us who carries the risk. OpenAI’s shareholders lose first if the company fails. Its cloud providers may be left with underused servers. Lenders may face defaults. Property funds could end up owning highly specialized buildings that are difficult to lease to another customer.
| Part of the bill | Who usually pays first | How the money comes back |
|---|---|---|
| Land and buildings | Developers and infrastructure funds | Rent and long-term leases |
| Chips and servers | Cloud providers, customers and hardware financiers | Cloud-service payments |
| Electricity infrastructure | Utilities and energy developers | Power contracts and minimum bills |
| OpenAI’s operating losses | Equity investors | Future company value |
| Long-term compute use | OpenAI | Revenue from AI products |
| Final economic cost | Consumers and businesses | Subscriptions, API usage, advertising and enterprise contracts |
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
Is OpenAI building the data centers itself now?
OpenAI currently controls more of the design, but partners still provide most of the buildings, hardware and upfront capital.
Most OpenAI infrastructure still follows the cloud model. Microsoft, Oracle, AWS and CoreWeave build or lease facilities, install computing equipment and sell the resulting capacity to OpenAI over several years.
The flagship Stargate campus in Abilene, Texas, is a good example. Crusoe developed the site. Funds managed by Blue Owl and Primary Digital Infrastructure helped finance the property. Oracle agreed to operate the cloud infrastructure and buy hundreds of thousands of Nvidia chips. OpenAI became the main user.
Project Camellia in Georgia represents a real change. OpenAI says it is designing and developing the campus itself and has contracted with Georgia Power for 3.2 gigawatts of electricity, delivered in stages between 2028 and 2032. The Wall Street Journal reported that OpenAI has committed $20 billion to begin the project.
That gives OpenAI more control over the building schedule, power supply and technical design. It also moves more responsibility onto OpenAI than a normal cloud-services agreement would.
Still, developing a campus does not require OpenAI to fund every part alone. The company can bring in property investors, utilities, equipment lenders and operating partners. OpenAI itself says its infrastructure strategy depends on cloud providers, chipmakers, construction companies, energy suppliers, investors and public-sector partners working together.
OpenAI is becoming a data-center developer, but not a self-funded one.
If you want more recent data on this point, please see our latest AI infrastructure market report.
How much has OpenAI promised to spend?
OpenAI has already made roughly $710 billion of publicly disclosed or reliably reported cloud commitments, covering only part of its broader computing plan.
Microsoft says OpenAI has agreed to buy an additional $250 billion of Azure services. Amazon says its original $38 billion AWS agreement was expanded by another $100 billion, bringing the combined AWS commitment to $138 billion.
CoreWeave puts the maximum value of its OpenAI contracts at approximately $22.4 billion. The Wall Street Journal reported that OpenAI also agreed to purchase about $300 billion of Oracle computing capacity over roughly five years, beginning in 2027.
Together, those agreements reach approximately $710.4 billion.
That total needs careful interpretation. The contracts run for different lengths of time. Some represent maximum values. Some depend on capacity being delivered. OpenAI will probably pay gradually as servers come online and usage grows.
Even with those qualifications, the order of magnitude is clear. OpenAI has promised more future cloud spending than almost any company has ever committed to a group of technology suppliers.
The latest estimate of $750 billion through 2030 sits only around $40 billion above the four large agreements below. That suggests Microsoft, Oracle, AWS and CoreWeave currently account for most of OpenAI’s expected near-term compute bill.
| Provider | OpenAI commitment | Main service |
|---|---|---|
| Oracle | Approximately $300B | Large-scale Oracle Cloud capacity |
| Microsoft | $250B incremental | Azure computing services |
| AWS | $138B total | AWS infrastructure and Trainium capacity |
| CoreWeave | Up to $22.4B | GPU cloud capacity |
| Combined | Approximately $710.4B | Multi-cloud computing |
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
Why is Oracle borrowing so much money for OpenAI?
Oracle is borrowing because it must build OpenAI’s computing capacity years before collecting the full cloud revenue.
The reported OpenAI agreement is worth around $300 billion over roughly five years. An even distribution would equal $60 billion annually, although actual payments will depend on when facilities become available.
Oracle cannot wait for those payments before buying land, constructing buildings, installing networks and ordering chips. It has to finance the infrastructure first.
Oracle’s latest financial results show the cost. The company generated $32 billion in operating cash flow during its latest fiscal year, but free cash flow fell to negative $23.7 billion as infrastructure spending surged.
To cover that gap, Oracle raised $43 billion in debt and $5 billion in equity during the year. It expects to raise approximately another $40 billion through debt and equity in its following fiscal year.
Oracle openly says it is raising capital to meet contracted demand from customers including OpenAI, Meta, Nvidia, xAI and TikTok. OpenAI may be the most closely watched customer, but Oracle is spreading the new capacity across several large buyers.
Customers are also helping with hardware. Oracle says companies involved in its large AI contracts have prepaid for or directly supplied $75 billion of GPUs. Oracle has not revealed how much came from OpenAI.
The arrangement is pretty clear. Oracle’s bondholders and shareholders advance the capital. Oracle builds the computing capacity. OpenAI pays Oracle over time, provided its own revenue grows fast enough.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Who is writing the checks for Stargate?
Stargate is being funded project by project by SoftBank, Oracle, developers, infrastructure funds, banks and equipment suppliers.
When Stargate was announced, OpenAI and SoftBank described a plan to invest $500 billion in US AI infrastructure over four years. The announcement named SoftBank, OpenAI, Oracle and MGX as the first equity funders, with SoftBank taking financial responsibility and OpenAI taking operational responsibility.
That announcement was widely read as though the group had already secured the whole amount. In practice, the $500 billion is a target spread across many campuses, companies and financing rounds.
Abilene shows what this looks like on the ground. Crusoe, Blue Owl-managed funds and Primary Digital Infrastructure formed a $15 billion joint venture to finance the 1.2-gigawatt site. The package included both debt and equity.
Oracle then agreed to buy roughly 400,000 Nvidia GB200 chips for the campus, according to the Financial Times. The publication estimated the hardware bill at around $40 billion, with Oracle leasing the resulting computing power to OpenAI.
OpenAI’s promise to use the capacity ties the whole arrangement together. Banks are more willing to lend when a major customer has signed a long contract. Property funds are more willing to build when Oracle has agreed to become a tenant or operator. Oracle is more willing to buy chips when OpenAI has committed to pay for the service.
OpenAI’s demand makes Stargate financeable. Its partners still have to find and advance the money for each individual project.
This chart, included in our AI infrastructure market deck, shows why CoreWeave is winning in AI infrastructure
How much is SoftBank really paying?
SoftBank is currently OpenAI’s largest named direct financial backer, with an expected cumulative investment of $64.6 billion.
Its latest commitment adds another $30 billion to OpenAI through SoftBank Vision Fund 2. Once completed, SoftBank expects to own approximately 13% of OpenAI.
That sounds like SoftBank is simply transferring cash from its own reserves. The actual financing is more leveraged.
SoftBank split the investment into three $10 billion tranches. It borrowed $10 billion to fund the first tranche and another $10 billion to fund the second. A third tranche is scheduled later, subject to the agreement’s terms.
Banks therefore provided at least $20 billion of the immediate cash already sent under this investment. SoftBank remains responsible for repaying those loans and carries the risk if its OpenAI shares lose value.
This is normal for an investment company, but it changes how we should describe the payer. SoftBank is making the investment decision and accepting the equity risk. Its lenders are supplying much of the cash at the beginning.
SoftBank’s wider role may prove even larger. Its energy subsidiary, SB Energy, is developing major infrastructure projects and can bring together land, power, lenders, governments and technology partners.
That makes SoftBank both an OpenAI shareholder and one of the main organizers of the financing around OpenAI.
Are Microsoft, Amazon and Nvidia funding their own customer?
Microsoft, Amazon and Nvidia are all helping finance a customer that has promised to spend huge amounts on their products.
Microsoft’s OpenAI stake was valued at roughly $135 billion after OpenAI’s restructuring. OpenAI has also committed to buy an additional $250 billion of Azure services.
Amazon is investing $50 billion in OpenAI, beginning with $15 billion and followed by another $35 billion when agreed conditions are met. OpenAI, in return, has expanded its AWS commitments to $138 billion and agreed to consume around two gigawatts of Amazon Trainium capacity.
Nvidia invested $30 billion in OpenAI’s latest financing round. OpenAI says Nvidia GPUs currently power its training fleet and most of its inference systems.
A much larger Nvidia role may now be forming. The Wall Street Journal reports that Nvidia is discussing a guarantee of roughly $250 billion to support financing for a possible 10-gigawatt OpenAI data-center lease in Ohio. Nvidia is also reportedly considering ways to finance the chips inside the project.
The Ohio discussions remain unfinished and could still collapse. They are nevertheless revealing. Nvidia may use its own financial strength to help lenders fund a project that would later purchase hundreds of billions of dollars of Nvidia equipment.
Money is moving in a loop: suppliers invest in OpenAI, OpenAI signs large purchasing agreements, and those agreements help suppliers justify new infrastructure.
There is real economic activity inside that loop. Servers are built, electricity is consumed and cloud services are delivered. The danger starts when supplier financing runs far ahead of revenue from independent customers.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
How does CoreWeave afford all this OpenAI capacity?
CoreWeave affords OpenAI’s computing capacity by using long-term customer contracts to raise large amounts of debt.
Its agreements with OpenAI have a combined maximum value of approximately $22.4 billion. OpenAI also received $350 million of CoreWeave shares as part of the companies’ first large agreement.
CoreWeave can show those contracts to lenders as evidence of future revenue. It then borrows money to lease sites, buy GPUs and connect the equipment to power.
The company’s latest figures show both sides of that model. CoreWeave reached more than one gigawatt of active power and built a revenue backlog of $99.4 billion. It also recorded $536 million of net interest expense and a $740 million net loss during the quarter.
CoreWeave has since secured an $8.5 billion delayed-draw loan facility structured without recourse to the wider company. Nvidia has also invested another $2 billion in CoreWeave.
The model works while contracted revenue arrives faster than financing costs, chip depreciation and operating expenses. It gets painful when a customer delays deployment or needs less capacity than expected.
OpenAI is therefore paying CoreWeave through long-term service contracts. Before those payments arrive, CoreWeave’s lenders, shareholders and Nvidia are funding the machines.
Where do banks and private funds fit in?
Banks and private funds currently supply much of the money that turns OpenAI’s promises into physical data centers.
They often remain hidden behind the company named in the headline. Oracle issues bonds. SoftBank takes a bridge loan. CoreWeave secures equipment financing. Crusoe and Blue Owl create a property joint venture.
Each transaction gives investors a different type of exposure. A bank may lend against a building. Another lender may finance GPUs. Bondholders may rely on Oracle’s entire corporate balance sheet. A private infrastructure fund may own a campus and collect rent for 15 years.
OpenAI has its own bank support too. It recently expanded its revolving credit facility to approximately $4.7 billion, backed by a group including JPMorgan Chase, Citi, Goldman Sachs, Morgan Stanley, Wells Fargo, UBS and HSBC.
The facility remained undrawn when OpenAI announced it. Compared with the company’s planned infrastructure spending, $4.7 billion offers short-term flexibility rather than a full financing solution.
The bigger borrowing happens one step away from OpenAI. Stronger companies such as Oracle can issue bonds more cheaply. Data-center owners can borrow against property. CoreWeave can pledge equipment and customer contracts.
This structure keeps much of the debt away from OpenAI’s balance sheet, while OpenAI’s payment promises still support the borrowing.
This chart, included in our AI infrastructure market deck, compares the main business model options for AI cloud infrastructure providers
Are taxpayers and electricity customers helping pay?
Public support lowers some data-center costs, but private companies and investors still cover most of OpenAI’s direct construction bill.
Local governments may offer property-tax reductions, build roads, speed up permits or reserve land. Utilities may construct new power plants, substations and transmission lines before a data center begins consuming electricity.
The Abilene project received approval for large property-tax abatements. Business Insider reported that parts of the development could receive an 85% property-tax reduction if the project meets investment and employment requirements.
That reduces the developer’s future expenses. It also means the local government collects less tax than it would under normal rates.
Electricity creates a larger potential risk because a single AI campus can require as much power as a city. Project Camellia has contracted for 3.2 gigawatts from Georgia Power, delivered over several years.
Georgia’s Public Service Commission created special rules for new customers using more than 100 megawatts. Those rules allow longer contracts, minimum bills and charges for generation, transmission and distribution infrastructure.
The stated goal is to keep existing households and businesses from paying for infrastructure built mainly for data centers. OpenAI also says it will cover the full cost of the electricity infrastructure and service required for its Georgia project.
Whether those protections work perfectly will depend on the final contracts and the life of the project. A data center that closes early can leave behind power assets that still need to be paid for.
For now, public support looks more like a cost reduction and risk-sharing mechanism than the main funding source.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Can OpenAI’s current revenue cover the bill?
OpenAI’s current revenue cannot cover its planned data-center spending, even after its extraordinary recent growth.
OpenAI says it now generates about $2 billion in monthly revenue. It has more than 900 million weekly users, over 50 million paying subscribers and an enterprise business contributing more than 40% of revenue.
Those numbers make OpenAI one of the fastest-growing technology companies ever created. They still leave a wide gap between the business today and the infrastructure it has ordered.
A simple calculation shows the challenge. Fifty million subscribers paying an average of $20 per month would produce $12 billion annually before taxes, app-store fees, discounts, staffing, research and computing costs.
Enterprise contracts, API usage and higher-priced subscriptions add substantial revenue. OpenAI also says its early advertising pilot reached more than $100 million in annualized revenue in under six weeks.
The company recently raised $122 billion in committed capital, giving it a large cushion while the business grows. Equity funding cannot permanently cover cloud bills of this size, because investors eventually expect the company to produce cash rather than keep consuming it.
As seen above, the projected $750 billion compute bill is more than 31 times OpenAI’s current annualized revenue. OpenAI therefore needs several years of exceptional growth, better computing efficiency and much stronger revenue per user.
The plan can work, but only if OpenAI becomes one of the largest revenue-generating technology platforms in the world.
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
What happens if OpenAI cannot pay?
OpenAI’s shareholders and its most leveraged infrastructure partners would take the first serious losses if the company could no longer meet its data-center commitments.
OpenAI’s private valuation would fall first. Investors have committed capital at an $852 billion post-money valuation, largely because they expect OpenAI’s revenue to catch up with its infrastructure ambitions.
Cloud providers would then face the practical problem of unused capacity. Oracle could redirect some facilities to Meta, Nvidia, xAI or other customers, but finding another buyer for tens of billions of dollars of specialized computing power would take time.
CoreWeave has less room for error. Its high interest expense and heavy reliance on a small number of large customers make idle GPUs especially costly.
The private contracts probably contain minimum payments, milestones, prepayments and delivery conditions. Since the full terms are confidential, we cannot know exactly how easily OpenAI could reduce or delay its purchases.
A moderate shortfall would probably lead to slower deployments, contract extensions or capacity being moved to other customers. Suppliers have little reason to push OpenAI into collapse when a renegotiated agreement could preserve years of revenue.
A severe shortfall would reach lenders and property investors. Utilities and local governments should sit further from the loss when minimum bills, deposits and tax-clawback rules are strong enough.
| Who carries the risk? | What could go wrong? | Main protection |
|---|---|---|
| OpenAI shareholders | Valuation falls and new funding becomes harder | Continued revenue growth |
| Oracle shareholders and bondholders | Expensive capacity remains underused | Other large cloud customers |
| CoreWeave shareholders and lenders | Debt costs continue while GPU usage falls | Long-term contracts and collateral |
| Infrastructure funds | Specialized buildings lose tenants | Long leases and property ownership |
| Utilities | New grid assets become stranded | Minimum bills and long contracts |
| Local governments | Tax incentives produce fewer benefits than expected | Investment and job requirements |
So who is really paying for OpenAI’s data centers?
OpenAI’s partners, investors and lenders are paying now; OpenAI’s customers are supposed to repay them later.
Oracle, Microsoft, AWS and CoreWeave are building or acquiring much of the computing capacity. Crusoe, Blue Owl and other infrastructure investors are financing the buildings. SoftBank, Amazon, Nvidia, Microsoft and financial institutions are supplying OpenAI with equity.
Banks and bondholders are lending to nearly every layer. Utilities and energy developers are funding power infrastructure under long-term contracts. Local governments sometimes lower project costs through tax incentives.
OpenAI provides the commitment that connects these groups. Its agreements tell Oracle how much cloud capacity to build, help CoreWeave borrow against future revenue and give property investors confidence that their campuses will have a tenant.
OpenAI will eventually have to meet those commitments using money earned from ChatGPT subscriptions, enterprise products, API customers, advertising, commerce and other services.
The final answer has three layers.
The upfront money comes mainly from OpenAI’s suppliers, strategic investors, banks and infrastructure funds.
The contractual obligation belongs mainly to OpenAI.
The final economic bill is meant to be paid by the people and businesses that buy OpenAI’s products.
Today, those customers are not yet producing enough revenue to support the full infrastructure plan. Capital markets are covering the difference because they expect OpenAI’s business to become dramatically larger.
If that expectation proves correct, OpenAI’s customers will gradually repay the companies financing the build. If it fails, the remaining bill will land on the shareholders, lenders and suppliers that funded the bet.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
OUR METHODOLOGY
This analysis asks who is actually paying for OpenAI’s data centers. We separate four roles that are often blended together in headlines: who supplies the upfront capital, who owns or finances the physical infrastructure, who signs the long-term payment commitments, and who ultimately carries the economic risk.
We traced the money through cloud providers, chip suppliers, property investors, banks, utilities and OpenAI itself. Large cloud contracts show where OpenAI’s contractual obligations sit, while debt issuance, bridge facilities and project-financing structures show who is advancing the cash today.
We distinguish firm commitments from maximum contract values, conditional investments and broader spending targets. The approximately $710.4 billion combined cloud figure is therefore an aggregation of agreements with different durations, delivery conditions and payment schedules, rather than a single bill due at once.
Revenue, cash flow and financing data were used to judge how much of the build OpenAI’s current business can support. OpenAI’s reported monthly revenue was compared with its projected compute spending, while Oracle’s cash flow, SoftBank’s borrowing and CoreWeave’s debt model helped identify where external capital is covering the gap.
Project-level evidence was used to understand the physical build. The Abilene Stargate campus shows how developers, infrastructure funds, banks, Oracle and OpenAI can each finance a different layer. Project Camellia shows OpenAI taking a more direct development role while still relying on utilities, capital providers and operating partners.
We prioritized corporate announcements, investor materials, financial results, financing agreements and reporting from established financial publications. Key sources include The Wall Street Journal on OpenAI’s projected compute spending and Project Camellia commitment, OpenAI on its infrastructure and financing ecosystem, OpenAI on the Georgia project and its power-cost obligations, OpenAI on the original Stargate structure, Microsoft on the additional $250 billion Azure commitment, AWS on Amazon’s investment and expanded partnership, Oracle’s fiscal 2026 results, The Information on the reported Oracle agreement, Financial Times reporting on Abilene summarized by Reuters, and SoftBank’s disclosures on its OpenAI investment.
The conclusion comes from combining those layers. OpenAI’s partners and financial backers provide much of the money now, OpenAI carries the central long-term obligation, and future customers are expected to pay the final economic bill.
In our AI infrastructure market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Can OpenAI actually afford all this compute?
- Can OpenAI ever become profitable?
