The transition from perpetual licenses to subscription-based pricing was a difficult transition for all makers of system and application software, and it is one that Oracle steered through with a certain amount of grace. But that transition pales compared to the tectonic shift that Oracle is making as it is investing to become the fifth hyperscaler located in the United States.

There has been plenty of jumpiness about Oracle’s massive investments in infrastructure, with five major datacenters in various stages of development to support its AI processing aspirations. Wall Street was very jumpy once again about that this week as the company reported its results for the fourth quarter of its fiscal 2026 year. But here’s the thing: The worst of this transition is mostly over.

The company fired up 1.2 gigawatts of datacenter capacity in fiscal 2026, and in the first quarter of fiscal 2027, which ends in August, it will turn on another 1 gigawatt of capacity. And in fiscal 2027, it will start reaping the benefits of its general purpose CPU and GPU-accelerated AI hardware investments. Hilary Maxson, who took over as chief financial officer two months ago and who previously worked at Schneider Electric, which makes the power distribution equipment for massive datacenters among other things, said on the call going over the numbers that 12 percent of its revenue backlog of $638 billion would be recognized in fiscal 2027, which works out to an incremental $76.6 billion in revenues added to this fiscal year above and beyond the run rate for Oracle’s massive software business. Oracle’s entire revenues for fiscal 2027 were $67.4 billion, just to give you some perspective. Maxson added that over the following two years – in fiscal 2028 and fiscal 2029 – another 34 percent of that revenue backlog would be recognized, which works out to $219.6 billion. If you average it out, that is around $110 billion a year in incremental revenues.

This assumes that OpenAI, the big customer in that revenue backlog, can come up with the cash. Oracle is making that assumption, and that is why it is building five massive datacenters. The Stargate datacenter in Abilene, Texas is now running at 42 percent of capacity, and another 35 percent of capacity will be added in the next three months. This facility will be at full power before the end of the calendar year. Oracle’s datacenter in Shackleford, Texas, which was contracted in August last year, has 115 megawatts of power lit up and first deliveries of machinery will happened in the first half of calendar 2027. The Dona Ana, New Mexico datacenter, the deal for which was inked in September 2025, will have customers installing equipment in 1H 2027 as well, and this one will be powered by Bloom fuel cells. The Saline, Michigan datacenter, contracted in October 2025, will start taking delivery of customer gear in the second half of calendar 2027 as will the Port Washington, Wisconsin datacenter that was contracted in September 2027. These datacenters add up to 5.5 gigawatts of capacity.

Which is why Oracle has been floating debt and burning cash to build them, at an alarming rate compared to its cloud business, as you can see below:

The key thing to remember is that there are not enough GPUs in the world to satisfy the demand to generate tokens to do useful work. So even if Oracle can’t sell as much capacity as it has contracted to OpenAI, there is a long, long list of customers who will be happy to rent that GPU capacity. In fact, if OpenAI were to go bust, one could argue that Oracle might make more money selling those GPUs on the open market than it is getting from OpenAI. . . .

Still, it is plenty dramatic for Oracle to have had a measly $6.9 billion in capex for fiscal 2024, when the OCI cloud was just serving up database instances and running Oracle applications on top of them, and then see capex skyrocket to $54.75 billion in fiscal 2025 and then just utterly explode to $166.8 billion in fiscal 2026. Capex spending will cool off a bit to a mere $70 billion or so in fiscal 2027, which should have calmed Wall Street down. That is $291.6 billion in capex over three years, which is a lot, but Oracle will be making back at least $293.5 billion out of that revenue backlog. So it balances. Every GPU hour that Oracle sells above and beyond the OpenAI deal has essentially no cost to Oracle.

Clay Magouyrk, who is a co-chief executive officer at Oracle, said that Oracle’s GPU utilization across the entire fleet of machines on OCI is 97.5 percent, which seems pretty respectable. And even if as customer doesn’t renew for some reason or another, there is an existing OCI customer or a new one waiting in the wings that will snap up that GPU capacity in a heartbeat.

As SpaceX is showing, even vintage “Hopper” H100 and H200 GPUs from Nvidia are valuable commodities, and once Elon Musk decided to shitcan his current Grok AI model and just rent out capacity on the Colossus 1 and Colossus 2 supercomputers in Memphis, Tennessee, SpaceX could charge Google nearly $1 billion a month to rent time on 110,000 GPUs across those two machines. SpaceX is till going to be developing Grok, but it is going to start over because it did not scale well across Colossus.

With that, let’s dig into the Oracle financials.

In the quarter ended in May, Oracle had $19.18 billion in sales, up 20.6 percent. Operating income rose by 20 percent to $6.13 billion, and net income rose by 25.6 percent to $4.3 billion, representing 22.4 percent of revenues and down quite a bit as the costs of the AI buildout bite.

The company ended the quarter with $31.9 billion in cash and equivalents in the bank, and spent $55.7 billion in capital investments during Q4 F2026.

As you can see from the Hardware group sales, which were $924 million in Q4 F2026, some of Oracle’s customers are bringing their own hardware to the OCI datacenters and others are using capacity that Oracle builds for OCI but are not shown in the financials because they are not external sales. Oracle has very good operating profits on its hardware, which is what you expect from Larry Ellison. But those margins are trending down, just like margins are shrinking over time for the Cloud and Software groups. The Services group is seeing its operating income grow, however.

Of note in the financials for Q4 F2026 is the fact that the Cloud Infrastructure division had 93.2 percent growth to $5.79 billion. Of this, $4.8 billion came from renting CPUs and GPUs and $800 million came from rentals of Oracle databases. (We assume the other sliver of $200 million in revenues was for raw storage.) Rentals of CPUs and GPUs grew by 2.2X in the quarter.

Here is what the mix of Cloud group revenues looked like:

I expect for Cloud IaaS revenues to go exponential as OpenAI and other big customers come online and start renting CPUs and GPUs. The bulk of that AI revenue backlog will end up here. But there will be services and software components as well in these big deals. Oracle is being paid to build and manage the datacenters as well as the gear inside of it.