Hyperscalers Spend 102% of Cloud Revenue on AI

UBS ran the numbers on how much Amazon, Alphabet and Microsoft are pouring into AI, and landed on a figure that reads like a typo. In 2026 the three of them will spend roughly 102% of their cloud revenue on capital expenditure. Every dollar cloud brings in, and then a little extra on top, goes straight back out into data centres and chips.

That is not a company burning through its wages to keep the lights on. These businesses are vast and diversified, and cloud is only one slice of them. The ratio is a measure of how hard the biggest technology firms are flooring the accelerator on infrastructure they have not yet been paid to build. (Not investment advice.)

Table of Contents

The numbers keep getting bigger

UBS reckons total hyperscaler capex will run to about $492bn in 2025, then $1.009tn in 2026, $1.447tn in 2027 and $1.619tn in 2028. Add up 2026 to 2028 and you get roughly $4.1tn, more than three times the $1.292tn these companies spent across the previous six years combined.

The 102% figure eases a little from here, to around 99% of cloud revenue in 2027 and 94% in 2028. Do not mistake that for restraint. The percentage dips while the absolute spend keeps climbing, which is how you end up laying out more than $1.6tn in a single year even as the growth rate cools.

Who is writing the cheques

Across 2026 to 2028, UBS pencils in cumulative capex of about $938bn for Alphabet, $683bn for Meta, $672bn for Microsoft and $628bn for Amazon. Behind the household names come the newer spenders: roughly $335bn for SpaceX, $276bn for Oracle, $130bn for CoreWeave and $93bn for Nebius. The buildout has spread well past the handful of firms people picture when they think about AI.

For chipmakers, networking suppliers, power companies and data-centre builders, all of that becomes revenue somewhere down the chain. That is the bull case in one line: the spending is somebody else’s sales.

Where it could go wrong

A data centre does not earn a return simply because it is stuffed with expensive GPUs. The capacity has to stay full, customers have to keep paying, and AI services have to generate enough cash to cover depreciation, electricity and financing. Build too far ahead of demand and depreciation starts rising faster than revenue, which chews through margins and free cash flow.

There is a financing wrinkle too. Analysts at 24/7 Wall St reckon Alphabet, Meta and Microsoft are carrying around $3tn of AI-related obligations that do not always sit in plain view on the balance sheet. When spending is this large and this early, the gap between visionary and reckless comes down to whether the demand actually turns up.

Nobody knows yet, and UBS says as much. The bank stays bullish overall, arguing the real winners will not be the biggest spenders but the companies that convert the buildout into recurring revenue and strong returns on capital. Follow the money by all means. Favour the firms turning the spend into profit rather than the ones merely financing it. (Not investment advice.)

Did you know: at $1.619tn, projected hyperscaler capex for 2028 alone is in the same league as the entire annual economic output of a country like Spain.

Sources

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →