AI's Capex Bill Moved to the Credit Desk

Three of the largest companies on earth will spend every dollar their cloud businesses earn this year, and then some, on infrastructure. That is not a growth strategy. That is a solvency question deferred.

The number that should be the headline of 2026 is not a valuation. It is a ratio. UBS estimates that Amazon, Alphabet and Microsoft will collectively spend about 102 percent of their cloud revenue on capital expenditure this year, easing to roughly 99 percent in 2027 and 94 percent in 2028.

A hundred and two percent. Every dollar of cloud income recycled into the buildout, plus a bit more from somewhere else.

This is an argument, not a report. My position is that the interesting risk in AI infrastructure has quietly moved from the equity story to the credit story, and almost nobody is covering it that way.

The shape of the spend

Across the five largest US cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta and Oracle — committed 2026 capital expenditure runs somewhere between $660 billion and $690 billion, close to double 2025. Combined quarterly capex hit $129.8 billion in the first quarter, up about 80 percent year over year. Some analyses put the full-year guidance total above $700 billion.

Set that against what the AI ecosystem actually sells. The gap between infrastructure spending and AI revenue was estimated at roughly $600 billion annually when the calculation was first widely circulated, and capex has since accelerated faster than revenue projections have.

Two responses to that gap are available and both are partly right.

The bullish one: this is what building a general-purpose utility looks like. Railways, electrification and fibre all ran enormous negative cash flow for years before the demand caught the supply, and the companies doing this have the strongest balance sheets in corporate history. Revenue lags capex by eighteen to thirty-six months by construction; complaining about the lag is complaining about arithmetic.

The bearish one: revenue that lags is still revenue that has to arrive, and depreciation on this asset base starts hitting income statements on a schedule that does not care whether it did.

Why the financing structure is the story

Here is what changed in 2026, and why I think it matters more than the capex number itself.

For most of the last decade, hyperscaler capex was funded from operating cash flow. It was self-limiting: you could only build as fast as the business threw off money, which meant the spending was automatically disciplined by the thing it was supposed to serve. That link has broken. The platforms are increasingly turning to debt markets to bridge the distance between capex ambition and internal cash generation.

Leverage on a strong balance sheet is not a crisis. But it changes who is exposed and how quickly. Equity investors in Microsoft can reprice a growth story slowly and vote with a multiple. Credit markets reprice differently — faster, and with covenants. A financing structure built on the assumption that AI revenue arrives on schedule transmits any disappointment much more sharply than an all-equity one would.

It also changes the reversibility. A company funding capex from cash flow can slow down in a quarter. A company that has issued debt against a buildout, signed multi-year power contracts and pre-committed to chip allocations has made a set of promises that are extremely difficult to unwind at speed.

That is the part that rhymes with 2001. Not the valuations — the telecom analogy always gets misused as a valuation argument. What actually broke in 2001 was that the capacity was financed with debt on the assumption of demand growth that arrived years later than the coupon schedule. The fibre was eventually used. The companies that laid it mostly were not around to bill for it.

The accounting is doing work too

There is a second-order problem that compounds the first, and we have written about it: the depreciation assumptions inside AI earnings are quietly flattering reported profitability. Extend the useful life of a GPU fleet and this year's earnings improve without anything changing in the business.

That is legal, disclosed, and defensible on its own terms. It is also a lever that gets pulled hardest exactly when the underlying economics are most strained, which makes reported operating income a worse signal in 2026 than it was in 2022. If you want to know how the buildout is really going, free cash flow is a more honest number than earnings per share, and it is the one that has deteriorated most.

Meanwhile the demand side has its own measurement problem. AI's productivity gains have been persistently hard to find in customer financial statements. Enterprises are spending; whether they are getting returns that justify renewal at higher volumes is genuinely unresolved. The capex assumes they will.

What would make me wrong

I want to be specific, because a bearish argument that cannot be falsified is just a mood.

If inference demand from agentic workloads compounds the way the current cohort of application companies suggests — and revenue at companies selling agent outcomes has been growing fast enough to make the case — then the capacity gets absorbed and the ratio normalises on schedule. UBS's own path back to 94 percent by 2028 assumes something like this.

If the capacity turns out to be shorter-lived than the depreciation schedules imply, that is bad for reported earnings but neutral-to-good for the demand story, because it means the fleet needs continuous replacement and the buildout is a recurring business rather than a one-time overbuild.

And if the platforms are right that they are buying a durable position in a compute market that will be permanently supply-constrained, then Amazon tripling down on Nvidia and its peers' equivalent commitments look like land grabs rather than overreach. Being early to a constrained input is how you win those.

None of these possibilities are unreasonable. What I object to is the framing that treats the capex number as evidence of confidence rather than as a claim requiring evidence.

The zoom-out

Record spending is not, by itself, information. Capital allocation only tells you something when you know what it is funding and how it is funded.

What is being funded is a bet that AI demand is a utility, not a product cycle. What is funding it is, increasingly, borrowed money against revenue that has not arrived. Both of those can be true and the bet can still pay. But the combination changes the failure mode: it converts a disappointing quarter from a stock story into a refinancing story, and those move faster and further than anyone models in advance.

Spending your entire revenue on capacity is a statement of conviction. Borrowing to spend more than your entire revenue is a statement of urgency. It is worth being clear which one you are watching.

Frequently Asked Questions

How much are hyperscalers spending on AI capex in 2026?

Microsoft, Alphabet, Amazon, Meta and Oracle have collectively committed roughly $660–690 billion in 2026 capital expenditure, close to double 2025 levels, with some analyses of company guidance putting the total above $700 billion. Combined quarterly capex reached about $129.8 billion in the first quarter of 2026.

What is the AI capex-to-revenue gap?

It is the difference between what hyperscalers spend building AI infrastructure and what the AI ecosystem earns in sales. Estimates put the annual gap at roughly $600 billion, and it has widened through 2026 as capital spending accelerated faster than revenue forecasts were revised upward.

Are hyperscalers spending more than they earn on cloud?

Per UBS estimates, Amazon, Alphabet and Microsoft will spend about 102 percent of their combined cloud revenue on capital expenditure in 2026 — slightly more than the segment generates. That ratio is projected to ease to roughly 99 percent in 2027 and 94 percent in 2028 as revenue catches up.

Is the AI buildout comparable to the 2001 telecom bust?

Partially. The relevant parallel is not valuation but financing: telecom capacity was debt-funded against demand that arrived years after the interest payments. The fibre was eventually used, but many builders did not survive to bill for it. Today's builders have far stronger balance sheets, which materially changes the risk.

Editor's note — sources: UBS estimates via Yahoo Finance; Futurum Group on 2026 AI capex; TMT Finance; Forbes on the capex-to-revenue gap. This piece is analysis; the estimates cited belong to their sources and the interpretation is ours. Nothing here is investment advice.