Software giant Oracle has a vast installed base of enterprise customers that it has agglomerated over the decades that gives it the cash flow to do many things. But what it does not have is the cash generation and top-notch credit ratings to afford to build all of the datacenters that its customers want it to operate, unlike its larger rivals Amazon Web Services, Microsoft Azure, and Google Cloud.

Which is one of the reasons why Oracle’s top brass have said again and again that the company does not own the datacenters that are at the heart of some its of biggest deals, importantly the $300 billion agreement that the company has with model builder OpenAI that runs from now out five years. Rather, Oracle owns the equipment, and it builds it as the money comes in and turns it around fast so it can get paid. But still, Oracle has to come up with cash every year to buy the stuff to make the gear, which we presume it will have in the field for more than five years and therefore get its profits as AI takes off among its hundreds of thousands of enterprise customers.

This seems to be the business model that Oracle co-founder, chairman, and chief technology officer Larry Ellison has come up with to differentiate Oracle Cloud Infrastructure from the larger clouds from Amazon, Microsoft, and Google and the handful of neoclouds (with Nvidia-backed CoreWeave at the front of the pack) who are also trying to get some of the AI compute action by focusing AI exclusively and bang for the buck in particular.

But it takes money to play this game – a lot more than Ellison, whose wealth largely comes from his 40 percent to 42 percent stake in the software giant, can come up with on his own.

Not that Oracle doesn’t generate plenty of cash, nor that it has not spent lots of money building its own megawatt-scale datacenters. Between the heart of the Great Recession in 2009 and the end of 2017 as the Age Of AI was dawning, Oracle extracted great profits from its vast enterprise customer base and grew its cash hoard from $11.3 billion to $71.6 billion.

The problem is that GenAI has expanded so fast it requires multi-gigawatt facilities for the largest model builders. A gigawatt AI datacenter costs anywhere from $45 billion to $60 billion – and somewhere around half of that goes to power and the datacenter facility, with the other half going to servers, storage, and switching.

With only $19.8 billion in cash in the bank and a debt load (both long term and short term) of $124.4 billion as the second quarter of fiscal 2026 ended last November, the company needs more money to prime the pump so it can start building the gear to service OpenAI a year from now in the Stargate datacenters that are actually being built by Crusoe in Abilene, Texas. The main reason that Oracle had $19.8 billion in cash last quarter is that it sold a set of bonds with maturities of between 3 and 40 years last September worth $18 billion.

In other words, Oracle could not afford to do business with OpenAI unless it borrowed money.

And so, today Oracle announced a plan to raise somewhere between $45 billion and $50 billion in calendar 2026 through a combination of debt and equity. (Now, you can see why this number is significant. This is enough money to cover the datacenter hardware for around 2 gigawatts of AI capacity.) Oracle has to go into the markets raise funds because it needs to prime the pump: OpenAI will not pay Oracle until it is using the machinery that Big Larry builds on its behalf, and even then, the payments will come in as services are rendered and taken off that $300 billion commitment from OpenAI, which is a majority portion of the $523 billion in revenue backlog that Oracle has.

The plan is for Oracle to issue $15 billion to $20 billion of new shares, which will dilute the existing shares by somewhere less than 5 percent by our math. The company will also issue mandatory convertible preferred securities, which could weigh in at around $5 billion. Another $25 billion or so will be raised through the sale of “investment-grade” senior unsecured bonds. The funny bit is that given Oracle’s relatively low but still investment-grade credit rating, which is rated at BBB by S&P, and the size of its debt load compared to its revenues (it is almost two to one in fiscal 2026 so far), Oracle is limited in the amount of debt it can take on and will have to dilute shares to meet its AI infrastructure goals. Or rather, those of OpenAI.

The question we have is will the cash flow from the rental of the first year’s worth of iron by OpenAI be enough to help pay for the next year’s tranche of datacenter expansion? OpenAI is in a hurry to get hardware, but it takes cash to do that. So Oracle is going to build up the Stargate systems estate over time, reinvesting its revenues from OpenAI to build out its fleet and try to get 40 percent gross margins. Exactly how that will all progress remains to be seen.

Oracle is starting with $70 billion in the bank plus the cash flow from its enterprise software business. Even assuming most of that software cash flow is invested in share buybacks, dividends, and development of AI tools to update and change its software estate, Oracle can plow its OpenAI revenues back into building up the GPU fleet until it hits the point the two companies have agreed upon.

There are too many variables to model this Oracle-AI deal from the outside. We tried three different ways and got three different answers, and a lot depends on the nature of the power used and the facilities and what that power cost is regionally. There is a lot of variation and too much assumption to be precise.

What we know is that Oracle has a five year deal that culminates with 4.5 gigawatts of datacenter facilities and AI hardware worth $300 billion in revenues to Oracle through rentals. Based on some rough math, we strongly suspect that OpenAI has bargained for a price around $10 per GPU-hour compared to the $14 to $18 per hours that AWS, Azure, and Google are charging for “Blackwell” GPU instances and the lower prices the neoclouds are charging – anywhere from $5 to $8 per GPU-hour.

But having said all of that, here is the basic math: At $60 billion per gigawatt, a good, high-end ballpark figure for a datacenter tricked out with the best AI gear, that 4.5 gigawatts costs $270 billion. About half of that goes to power and the facility, and half of that goes to datacenter gear. So an input of $135 billion in AI systems hardware by Oracle is going to net Oracle $300 billion in revenues, and it will have roughly half the money in the bank before calendar 2026 is over to start. And Oracle can use the rentals from the gear OpenAI uses to expand the fleet such that it peaks in year three or year four as OpenAI grows, and in year five all the gear is installed and all the OpenAI money coming to Oracle might be pure profit.

OpenAI still has to pay Crusoe and the datacenter operators a lot of money, and the power companies, too. It is not at all clear where OpenAI will come up with that money, which is also stacking up.

Anyway, the important thing is that OpenAI gets to use of the latest Nvidia and possibly AMD iron through OCI for five years, presumably with an option to extend, and each year that partnership gets OpenAI more and better stuff as well as access to the prior year’s stuff.

Eventually, the iron comes off contract with OpenAI and Oracle can rent the capacity out to its vast enterprise customer base, which has a staggering 430,000 customers. The iron will free up just about the time that GenAI becomes normal and is embedded in Oracle’s databases, middleware, and applications. And Oracle can support customers in a way that it would not have been able to, with massive amounts of AI oomph, had it not done the deal with OpenAI.

The other question we have is how much more will Oracle have to borrow? That all depends. If this $18 billion plus $45 billion to $50 billion is just to prime the OpenAI deal pump, this seems reasonable. But if Crusoe or SoftBank or other Stargate partners cannot do their parts, then Oracle may have to get into the datacenter facility business at a brand new scale, and that will require not only many tens of billions of dollars but also the right kind of land and power to support it.

Oracle may not have to borrow much more money to get $300 billion in revenues from OpenAI, particularly if its software and general purpose cloud businesses do well. It may have to wait for years four or five (and beyond) to see deep profits from its AI hardware investments. But after that, it has a vast network of AI capacity to sell to OpenAI, or any other model builder, or its enterprise customers base.

At that point, Big Larry might decide to be a model builder itself and take on OpenAI, much as it took on IBM in the relational database technology that Big Blue invented.

Why not? It will have an enormous fleet of tensor compute.