Reading Hyperscaler Capex Like a Systems Engineer

Earnings season for the big cloud providers is a peculiar exercise. The analyst community spends the first fifteen minutes on revenue growth and margins, then the next forty-five asking management questions that management is contractually incentivized not to answer directly.

But the capex table is right there. Nobody's hiding it. And for anyone trying to understand where AI infrastructure is actually going — not where it's being marketed — those numbers are more useful than anything on the call.

This week's reports from the major hyperscalers deserve a close read. Not for the headline numbers. For what the infrastructure spend implies.

The Number That Matters

Capex in Q1 2026 across the three dominant cloud providers came in materially above analyst consensus. Which is not a surprise — this has been the trend for six quarters running. But the composition of that capex is what's interesting.

The split between general datacenter buildout and AI-specific compute infrastructure has been shifting. What started as a modest AI premium on top of existing capacity spending has become a structural line item. GPU compute, networking fabric (the interconnect costs for large-scale training and inference clusters are genuinely enormous — often underestimated by people who haven't priced out InfiniBand at scale), and the power infrastructure to run it all.

That power piece is not a rounding error. It's a constraint. And the capex reports are starting to reflect that the constraint is being taken seriously.

What the Guidance Tells You

Full-year capex guidance held or increased across the board. That matters more than the quarterly number for one reason: datacenter infrastructure has an 18-to-36-month lead time from commitment to production capacity.

When a hyperscaler guides for higher capex in 2026, they're not talking about something you'll see in deployed capacity this year. They're building for 2027 and 2028 demand assumptions they're making today. The people making those assumptions have access to their own internal demand signals, their enterprise customer pipeline, and their model training roadmaps.

The bet they're making, with tens of billions of dollars of committed spend, is that AI compute demand will be substantially higher two to three years from now than it is today.

That could be wrong. But it's worth noting that these same organizations were criticized for over-committing to cloud infrastructure in 2019-2021, and then the pandemic made them look prescient. They've since built entire teams around making this call well.

The GPU Dependency

The other signal buried in the footnotes: vendor concentration. A meaningful share of current capex is flowing to a very small number of GPU suppliers — specifically to the Blackwell architecture that NVIDIA has been ramping since late 2024.

This creates an interesting structural dynamic. The hyperscalers are simultaneously the biggest customers of the company whose hardware defines AI infrastructure capability, and competitors to each other, and increasingly competitors to NVIDIA itself as they build custom silicon. Azure has Maia. Google has TPUs. AWS has Trainium.

Custom silicon doesn't replace NVIDIA overnight — or possibly ever, depending on how you model the software ecosystem lock-in — but the capex story is more complex than "cloud companies buy GPUs." They're also hedging.

What This Means for Infrastructure Builders

The practical implication: capacity that seems unavailable or expensive today is getting substantially easier to access over the next 18 to 24 months, as the infrastructure being committed to now comes online.

For anyone designing systems today, that matters. Architectural decisions you make under current capacity constraints may not be the right decisions for the environment you'll be operating in by 2028. The right move isn't to architect for future abundance — you have to ship something now. But it's worth knowing that "this is too expensive to run at scale" may not be a permanent constraint.

The numbers in those capex tables are effectively a bet. A very large, very committed bet. And the people making it aren't doing it based on hope.

Watch the capex, not the keynote.

— Dustin