Oracle sold $18 billion in bonds to bankroll the biggest cloud deal ever signed
Key takeaways
- Oracle issued $18 billion in bonds with maturities to 2065 to fund AI data centre construction
- The $300 billion Stargate contract with OpenAI targets 4.5 GW of capacity across four US states
- Half of the hardware needs deploying before end of 2026, creating intense supply chain pressure
Oracle has sold $18 billion in bonds in a single offering, and the money is there to start delivering on the largest cloud infrastructure contract ever signed. According to Data Center Dynamics, the debt is being raised ahead of a huge AI data centre build-out.
The $300 billion Stargate deal with OpenAI, announced last September, commits Oracle to building 4.5 gigawatts of AI data centre capacity over five years. The sites span Texas, New Mexico, Wisconsin and Michigan. Bond maturities stretch from 2030 to 2065, which tells you Oracle is planning in decades, not quarters.
What did Oracle actually sell?
A bond is a loan split into small pieces and sold to investors. The company receives the cash now, pays interest along the way and repays the principal on a fixed date, called the maturity. Oracle's offering was one large sale with several maturities, the earliest in 2030 and the latest in 2065.
Spreading maturities matters. Short-dated bonds are repaid sooner, while a bond due in 2065 locks in funding for roughly 40 years. A data centre is built to last for many years, so matching long-lived assets with long-dated debt is a standard way to avoid a refinancing crunch.
What is the Stargate deal?
Stargate is the name for the data centre programme at the heart of the agreement between Oracle and OpenAI. Data Center Frontier describes the $300 billion contract as an effort to build AI infrastructure at national scale. The 4.5 gigawatts it targets is a measure of electrical power, not computing speed. For comparison, a gigawatt is roughly what a large power station produces, so the plan amounts to several of those dedicated to running AI systems.
If you want a sense of what that kind of demand means for the grid, our explainer on how much power AI data centres use covers where the electricity comes from. For the wider picture, see what the global AI infrastructure build-out looks like from the inside.
The key numbers at a glance
| Item | Figure | What it means |
|---|---|---|
| Bond offering | $18 billion | Raised in a single sale |
| Stargate contract | $300 billion | Oracle and OpenAI cloud deal |
| Capacity | 4.5 GW | Across Texas, New Mexico, Wisconsin and Michigan |
| Timeline | Five years | Build period for the full capacity |
| AI systems hardware | $135 billion | Roughly half to deploy before the end of 2026 |
| Cloud backlog | $664 billion | Contracted work not yet delivered |
| Bond maturities | 2030 to 2065 | Long-dated funding |
Why does a software company need this much debt?
Oracle is best known for databases and business software, but its cloud arm is now the part of the company that needs the most capital. Renting out computing power to an AI lab means owning the buildings, the power connections and the servers first, and only then collecting payments over the life of the contract. The cash goes out years before it comes back in.
Bonds are one way to bridge that gap without handing over ownership. Unlike issuing new shares, borrowing does not dilute existing shareholders, though it does create fixed interest payments that must be met every year. That trade-off sits behind the whole deal.
Why the urgency matters
Roughly half of the $135 billion in AI systems hardware needs to be deployed before the end of this year. That means Oracle has to source, install and commission an enormous volume of servers, networking gear and electrical infrastructure in a window that is already closing.
The bottleneck is not compute chips. It is transformers, switchgear and the electrical equipment that sits between a signed lease and a live megawatt of delivered power. Transformers step voltage up or down so power can be used safely, and switchgear is the equipment that routes and protects electrical circuits. Neither can be conjured quickly.
Six hyperscalers (the very largest cloud providers) have now committed over $690 billion in AI data centre capital expenditure, and they are all competing for the same pool of contractors and components. That competition is what makes the deadline so demanding.
How does the money fit with OpenAI's own funding?
Oracle is not the only company taking on debt around OpenAI's ambitions. We have reported that Nvidia is reportedly backing OpenAI's debt with a credit guarantee, and that OpenAI plans to burn through 278 billion dollars of cash by 2030, according to the Financial Times. Those stories show how much of the AI build-out rests on borrowing and on a customer whose spending is still ahead of its income.
Oracle has also been watched for how it pays for growth. Our report on Larry Ellison cancelling a planned 7.5 billion dollar Oracle stock sale, as disclosed in Oracle's filing is a useful companion, because it shows a company weighing debt against selling shares.
The deployment timetable deserves a second look. Half of $135 billion is a very large sum to install in months, and every delay in a transformer delivery or a grid connection pushes back the point at which a site can earn money. Because Oracle has already borrowed, interest accrues whether or not the servers are switched on.
What does it mean for readers?
For most people the effect is indirect but real. The AI tools that millions use every day run in buildings like these, and the cost and availability of those buildings shape what the tools cost and how quickly they improve. A contract of this size also pulls on local electricity supply, construction labour and equipment makers in the four states involved.
It is also a signal about confidence. Investors who buy a bond maturing in 2065 are effectively betting that demand for AI computing will still be strong long after today's models are obsolete. Whether that proves right is the open question behind the figures above.
What to watch
Oracle's $664 billion cloud backlog gives it revenue visibility that most infrastructure companies would envy. But backlog is not cash flow, and $18 billion in new debt needs servicing regardless of whether OpenAI's compute demand materialises on schedule.
The bond market is pricing Oracle as if the AI buildout is a certainty. If deployment slips, or if cheaper inference models compress the demand curve, those 40-year bonds start looking very different. Inference is the work of running a trained AI model to answer requests, and cheaper models would mean less hardware is needed for the same output.
The first test is the end of 2026. If Oracle gets roughly half of its hardware live on time, the long-dated funding looks well judged. If supply of electrical equipment holds it back, the gap between a signed contract and delivered capacity will be the story to follow.
