Remaining CAPEX Obligation

Rather than further update the last post, we’ll start anew.

In our preparation for this week’s Oracle earnings, we wrote:

Absolute bestest most glorious company ever in entire universe of all time prepare awesome-sauce victory lap. With tomorrow’s earnings announcement the database vampire will permit us to bask in their greatness and invite us to contemplate why other companies even exist. Things are always so fabulous that the Oxford English Dictionary is considering replacing the word “superlative” with “Oracle”.

Despite a mediocre quarter, Oracle stock price took that awesome-sauce victory lap (and then some). Oracle missed revenue and earnings expectations, reported slower growth than each of the much larger Amazon, Google, Meta and Microsoft, free cash flow went further negative, and CAPEX shrank to $8.5 billion from the previous quarter’s $9.2 billion.

But it was a “bombastically scripted earnings call” for the ages. Oracle stock was up 36% today, maybe an all-time one day increase in valuation record. Kudos to Oracle for giving the market what it wanted to hear and to Larry for becoming the world’s richest man (doubtlessly a long coveted title).

The stock rocketed based on Oracle announcing a $317 billion increase in their Remaining Performance Obligation (RPO), which is a non-GAAP metric of booked but not yet delivered revenue (SEC reported Deferred Revenue, which is the stricter portion of RPO, was $12.1 billion, up 29% from the previous quarter).

My argument remains that Oracle is hyping future cloud revenue (that may or may not materialize), but neglecting the costs required to deliver and book that revenue. The iron law of cloud computing is no infrastructure, no revenue. And Oracle is a relative financial pipsqueak whose capital allocation abilities are already stressed.

This post offers guidance for people furiously updating their Excel models that detail every parameter of Oracle’s next half-decade.

Who’s Paying?

A day after the earnings call focused on that $317 billion increase in RPO, we learned Open AI was responsible for $300 billion of that increase (how much that actually translates into SEC reported Deferred Revenue is a question for next quarter).

Oracle stock is now a one-way bet on whether Open AI can raise hundreds of billions in new capital. Stargate “has struggled to get out of the gate” and Open AI is passing the hat for single digit billions instead of the desired trillions. Open AI’s projected losses are growing.

Oracle tried to paint the RPO explosion as broad based, citing xAI, Meta, NVIDIA and AMD as new customers. xAi has no revenue, is itself struggling to raise money, and not many customers are going to be dumb enough to take a business dependency on Elon for AI. Meta is in the midst of a $600 billion AI mulligan (maybe – neither Zuck nor the CFO are sure actually how much they plan to spend). But Meta likely promised every well-compensated recent hire their own GPU clusters, and Oracle is a source. NVIDIA and AMD are merely recycling revenue from Oracle (but the SEC, like Congress, is on recess). And why don’t they talk about their biggest current customer? (ByteDance).

How Much CAPEX?

If you are adding the new RPO numbers to your Oracle models, you also have to add the infrastructure costs, aka the Remaining CAPEX Obligation. Again, no infrastructure, no revenue.

Let’s return to the AWS history, as it is a great proxy, but step through it in more detail than the last post. We have ten years of data that lets us directly compare annual CAPEX to revenue.

We see AWS-specific CAPEX as a percentage of annual AWS revenue dropping from 60%, bottoming at 27% in 2023, and ramping back above 50% as the AI frenzy hits (they’re above 70% for the first half of 2025).

Now lets look at same numbers for Oracle:

You quickly notice Oracle is far less efficient than AWS, with CAPEX/revenue ratios in the triple digits ranging between 100% and 208% (the FY26 numbers are Oracle’s revenue and CAPEX guidance for the current fiscal year).

This is consistent with Oracle being a noob when it comes to large scale data centers, as we have chronicled. But it also shows the absurdity of Oracle’s claims to be “asset pretty light” – they are vastly less efficient than AWS despite bragging about not owning data centers (“I know some of our competitors like to own buildings”). Despite giving up margin, Oracle is not even remotely seeing any advantage from not owning data centers (I guess it could always be worse).

Now, as you update your five-year model for Oracle, you have to decide what CAPEX numbers go with that nearly geometric sequence of revenue increases. Some things to keep in mind as you pick parameters:

• Why is Oracle so much worse than AWS in CAPEX efficiency? Can they ever catch up? How long does it take to build manufacturing machinery that efficiently and effectively turns dollars into data centers at scale? What is the YOLO cost premium?
• Oracle doesn’t own the land or the data centers. They have taken a dependency, losing control and margin as the owners of those facilities presumably want to be paid (a Crusoe prospectus will be fascinating). And they will be paying not just for data center structures, but also behind-the-meter power. But the numbers above suggest they’re not getting any benefit from not owning the data centers. Why?
• GPUs are both pricier and depreciate faster than old school compute and storage. Oracle seems to want to only own shorter-lived, faster depreciating assets. What is the useful life of their infrastructure? Will NVIDIA slow delivery of new GPU generations?
• Maybe Open AI has agreed to pay for infrastructure up front? (with what money?)
• Maybe Oracle has magic beans? (admittedly previous magic beans claims just haven’t come to pass. The phrase “bombastically scripted earnings calls” again comes to mind).
• What could go wrong with Oracle’s plans? Are there ways to control strategic dependencies beyond yelling at contractors? Could the market possibly shift in any way over the next five years to Oracle’s detriment?

So let’s throw out some numbers for the required CAPEX spend as a percentage of revenue (Oracle’s average over the last three years is 170%, which would be a level of CAPEX that exceeds all the new-found RPO). Pick your preferred numbers and enter them into your model:

	Revenue Forecast ($B)	CAPEX Spend ($B)
		50%	100%	150%	200%
FY27	$32	$16	$32	$48	$64
FY28	$73	$37	$73	$110	$146
FY29	$114	$57	$114	$171	$228
FY30	$144	$72	$144	$216	$288
	Total CAPEX	$182	$363	$545	$726

How Do They Pay?

The last step for your updated model is to figure out where the money to pay for RCO is coming from. We detailed Oracle’s financial constraints in the previous post. Free cash flow is already negative. Oracle carries a fair amount of debt (though that debt/equity ratio looks better today!). And they allocate a lot of capital to Larry through both dividends and buybacks. What does Oracle’s capital allocation look like for the next five years that is consistent with your model?