The AI Capex Cycle, and Why It Is Your Real Counterparty

Four companies' capital budgets underwrite almost all demand for AI data centers. Here are their actual capex figures from their own SEC filings, what share of their cash flow they represent, and what it means for a building you have bought a unit of.

Every AI data center investment rests on an assumption that somebody will rent the building. In practice that somebody is one of a very small number of companies, and the rent comes out of a capital budget that those companies set each quarter and can revise. That budget, not the concrete, is your real counterparty.

This page puts numbers on it, taken from the companies' own annual reports filed with the SEC. No forecasts, no analyst estimates: only what each company reported it actually spent.

What the four largest buyers actually spent

Sources, each the company's most recent Form 10-K as filed: Amazon, Microsoft, Alphabet, Meta. Figures are as tagged in each filing's cash flow statement.

The growth rate is the part that matters

The same four companies, in their immediately preceding fiscal years, spent $237.4B between them: Amazon $83.0B, Microsoft $64.6B, Alphabet $52.5B, Meta $37.3B. Two years before that the combined figure was about $154B.

So the aggregate roughly doubled in two years, then rose about 72% again in one. There is no asset class in which that rate of spending growth continues indefinitely, and nobody serious claims it will. The investment question is not whether the growth rate falls — it is what happens to a specific building when it does.

What this means for a single building

A diversified listed REIT and a single private project are exposed to the same cycle in completely different ways.

A REIT holds hundreds of facilities on staggered leases to a mix of tenants. A slowdown in new commitments shows up slowly: as weaker leasing on new development, then as softer renewal spreads years later. The existing rent keeps arriving under contract. You can also sell your shares the same morning you change your mind.

A single unleased development project is exposed to the cycle at exactly one point: the moment it goes looking for its first tenant. If that moment lands in a quarter when the buyers are extending the life of existing kit instead of commissioning new halls, the building does not get a worse lease — it may get no lease, for as long as it takes. Your capital sits in a finished shell earning nothing. See the risks page on why failure to lease is first on the list, and how data center leases work for what a real commitment looks like.

That asymmetry is the whole argument of REITs vs private deals, and the capex cycle is the clearest reason it exists.

Why demand being real does not make a deal good

The spending above is real, audited and enormous. It is also the single most common argument used to sell retail investors a data center deal, and it is a non-sequitur. Strong demand for a category tells you nothing about whether a particular building, in a particular location, at a particular price, with a particular sponsor and fee load, will earn you money.

Three specific ways the argument fails:

1. Capital follows the same story you heard. Published vacancy in the primary North American markets has been at record lows, which is precisely why so much new capacity is being built. Supply responds. The deal you are shown was priced in a tight market and will be leased into whatever market exists when it is finished.
2. Hyperscalers do not rent just any building. They have specific requirements on power, redundancy, fibre, security and, increasingly, cooling topology. "There is huge demand for AI data centers" is not the same sentence as "there is demand for this data center".
3. The rent is not the return. Even a well-leased building delivers to you only what survives fees, sponsor units and timing. See why the headline is not your return.

What to actually watch, and where to find it

You do not need a subscription to track this. The four companies report capital expenditure every quarter in their Form 10-Q cash flow statements, filed on EDGAR and free. Two things are worth reading each quarter:

• The capex line itself, and its direction. One soft quarter is noise; two consecutive quarters of flat or falling capex at several of these companies would be the first hard evidence of a turn.
• The language around it in management's discussion. The phrases that matter are commitments to "continue to invest", and any shift towards "optimising" or "moderating the pace" of capital spending. These companies telegraph direction before it appears in the numbers.

For a private deal, the question that uses this information is narrow and answerable: "which specific tenants has this project approached, and what stage is each conversation at?" A sponsor pointing at industry capex totals instead of naming a pipeline is quoting the weather rather than the forecast for your house. The full list is in questions to ask a sponsor.

The honest summary

Four companies' discretionary budgets, currently running at about two thirds of their combined operating cash flow, underwrite the demand side of this entire asset class. That is an unusually concentrated dependency, and it is the reason a ten-year building financed against a one-quarter budget decision deserves more scepticism than the headline demand numbers invite.

It does not mean the category is a bad place to put money. It means the diversification and liquidity you give up in a single private deal are being given up against a cycle, not against a certainty — and that you should be paid visibly for that, in price or in retired risk, before you accept it.

What to read next: REITs vs private deals for the structural comparison, and the risks for the specific ways this goes wrong.

Nothing here is investment advice or a recommendation of any security. See how this site is funded.