This post resurfaces and expands a point in my recent post Follow the CAPEX: The Clown Car Race Checkered Flag in advance of Oracle’s June 11th earnings report.
To hear Oracle tell it, they are the best-est and everything is always going super-fantastic. Earnings announcements are opportunities for others to bask in Oracle’s greatness.
The database vampire wants us to believe that, despite their almost decade-late-start, they’re a player in cloud infrastructure services. Oracle is so awesome, in fact, their only problem is coping with seemingly infinite demand. They wax Sagan-esque (“billions and billions”) about all the cloud business they fail to capture.
But, have no fear, they’re building data centers like nobody else. They tell us “Oracle has been building data centers at a record level.”
Except they’re not building data centers at a record level. Not even by Oracle’s own recent standards, never mind hypercloud standards.
The ability to serve cloud customers is a function of CAPEX spend. No servers, no services.
Oracle practically brags they have neither customer demand forecasting nor a systematic process to turn demand and capital into sellable cloud services.
Manufacturing cloud capacity efficiently and at hyperscale is an extremely complex, multi-year discipline (and very different than dropping a few racks into a co-lo facility and calling it a region, which is Oracle’s experience as a boutique provider). Their outright decline in CAPEX spending over the last year suggests Oracle is still a long way from attaining this muscle.
Contrast Oracle’s recent rhetoric and actual CAPEX spending:
| FY23 Q1 (Sept 2023) | |
| “Because we have far more demand than we can supply, our biggest challenge is building data centers as quickly as possible.” | CAPEX down 24% Reiterate FY24 CAPEX guidance of ~$8.3B |
| FY2024 Q2 (Dec 2023) | |
| “We have to build 100 additional cloud data centers because there are billions of dollars more in contracted demand than we currently can supply. Cloud Infrastructure demand is huge and growing at an unprecedented rate.” | CAPEX down 56% Lowest CAPEX spend in 8 quarters Reduce FY24 CAPEX guidance to ~$8B |
| FY2024 Q3 (Feb 2024) | |
| “We are working as quickly as we can to get the cloud capacity built out given the enormity of our backlog and pipeline.” “Oracle has been building data centers at a record level” | CAPEX down 36% Claim spent more than reported in 10-Q Reduce FY24 CAPEX guidance to $7.0-7.5B FY25 CAPEX guidance of $10B |
Oracle’s CAPEX peaked five quarters ago and declined 40% in first three quarters of FY24 relative to the prior year. This is not the spending of a company building data centers “at a record level”. This is a company that does want to grow its cloud capacity but doesn’t know how to convert capital into capacity, even at their relatively modest scale.
While Oracle can’t even keep up with Oracle of the recent past, the contrast with the hyperclouds is even more embarrassing (I’ve omitted AWS because they don’t provide quarterly CAPEX numbers, and AWS is only roughly half of Amazon’s overall reported CAPEX):
Oracle falls further and further behind with every passing quarter. Google and Microsoft start from a much higher level of spend, and their AI-driven CAPEX ramp of the last year is very visible. Oracle, meanwhile, continues to be the x-axis (I may never tire of that joke).
Learning in Public
Over-the-top bluster has long been one of Oracle’s core corporate values.
That ethos perhaps explains the company’s history of making big cloud infrastructure claims that fall ridiculously short. With incidents like “I don’t need as many data centers” and “We never thought it would go as high as it really did” (an astonishing 20% growth!), Oracle publicly and repeatedly demonstrates an incomplete grasp of what it means to be hyperscale.
Building cloud infrastructure capacity is a complex, multi-year supply chain discipline spanning multiple domains including hardware procurement, provisioning, electricity, permitting, and construction. Various inputs, notably compute and electricity, are scarce and presently face multi-year lead times.
Oracle is competing against other cloud infrastructure providers who are more experienced, more sophisticated, and operate at much higher scale. Those competitors built their capacity deployment muscle through a series of hard lessons when stakes were lower. And they’re higher on the waiting lists for components and power.
Oracle has started acknowledging they are struggling to build data centers:
“So, there is a tremendous amount of demand, the data centers take longer to build, and we would like that said, we are getting very good at building them quickly and getting the building the power and the communication links in, we’re doing faster than we have ever happened in the past.”
“It’s very different than it used to be. So, we’re able to bring additional capacity online very quickly if we have that the electric power and the communication lines. So, is the long pole in the tent is actually building the structure, connecting the electricity, connecting the communication lines.”
Transitioning from buying components at the local Best Buy to building hyperscale infrastructure is a big leap. It is a leap that requires a vastly more sophisticated process to convert capital into capacity. There is no sign of that process at Oracle.
Oracle has more learning ahead of them. They will discover that data center buildings are not a relevant metric. They will be disappointed to learn that building a really big data center isn’t the solution (measuring data centers in terms of how many 747s they will hold is ironic, given the 747 has also given way to smaller jets…). They will eventually move beyond costly whipsaw spending and focus ruthlessly on the capital efficiency of their CAPEX spending. And they might even figure out how to forecast demand so they can match supply and demand. We’re going to get to watch this learning unfold in public over the next few years.
FY 2024 Q4 Expectations
Oracle’s Q3 statement was “our Q4 CapEx should be considerably higher”.
To hit the (reduced) $7.0 to 7.5B guidance for the fiscal year, (non-LLM) arithmetic suggests we should expect Q4 CAPEX of $2.93B to $3.43B (CAPEX over the first three quarters as reported to the SEC was $4.068B, with a peak quarterly spend of $1.67B).
But there is some weirdness. In Q3, Oracle reported $1.67B in CAPEX to the SEC, but bizarrely claimed in the scripted portion of their conference call that they actually spent $2.1B in the quarter (analyzing securities fraud is outside our scope here – see Matt Levine for that):
Now while we spent $2.1 billion on CapEx this quarter, the $1.7 billion in the cash-flow statements is slightly lower just due to the timing of payments. So the $2.1 billion is actually what we spent and will pay for.
And I would include for Q3 the one we just are announcing. I would add in the amount we haven’t paid yet as the CapEx number for this quarter. Okay? And then I guess that would be and then Larry gets the second question. But anyway, so $2.1 billion for this quarter and you’ve got Q1 and Q2 and I’m going to be somewhere between $7 billion and $7.5 billion for the full year, which is actually a little bit lower than I thought. But we were able to do pretty well. You know-how we spend very carefully.
(Those last two platitudinous sentences don’t help their case).
If we somehow give them credit for the magical $400M (as opposed to assuming it is just Q4 spend), they only need to spend $2.5B to $3B to hit guidance (and, more importantly, make it seem like they’re building “at record levels”!).
Is this just random confusion (in the noble cause of trying to make their CAPEX spend seem larger)? Do they have the finance and accounting controls in place to know how much they’re actual spending? Is this an attempt to lower expectations to cover for a continued inability to build capacity in Q4? Does Oracle actually want to play (i.e. spend) at hyperscale levels? (they are ~$150 billion in cumulative CAPEX spend behind the hyperclouds). We’ll see Tuesday.