Four companies have now committed something in the region of a trillion dollars to artificial intelligence infrastructure. Combined capital spending is expected to reach about $760bn this year, against $413bn last year. Those are cash numbers, and the cash is already gone.
The expense is not. A data centre is a capital asset, which means it does not hit the income statement when it is bought. It arrives in slices across the years the equipment is in service. If a server is treated as having a six-year life, roughly a sixth of its cost lands on the profit and loss account each year for six years.
So the reported cost of the AI build-out is not a measure of what has been spent. It is a measure of how much of what has been spent has been allowed to arrive.
Where it has started to show
Meta's operating margin fell to 31 percent from 43 percent a year earlier. Its finance chief, Susan Li, attributed the growth in expenses to higher depreciation, data centre operating costs and third-party cloud spend. Free cash flow fell to $784m, from $8.55bn a year earlier.
That is what the first slices look like on a company that has been spending for a while. Nothing there is a surprise or an accident. It is the mechanism working exactly as it is supposed to, on a scale that makes it visible in the headline numbers rather than in the notes.
The estimate that sets the pace
The variable in all of this is useful life, and useful life is an estimate the company makes.
Microsoft is extending the estimated useful life of data centres and office buildings from fifteen years to twenty-five, effective from the start of its 2027 financial year. Its finance chief, Amy Hood, described the effect on that year's operating income as minimal.
Stretch the life and the annual slice gets thinner. The total does not change — the same cost is recognised, over a longer period — but the near-term earnings drag falls, and it falls in every year until the schedule catches up. Independent analysts have put the cumulative effect of revisions like this in the region of two hundred billion dollars of depreciation deferred through 2028. That is an estimate about estimates, and should be read as one.
None of which is improper. Useful-life assumptions are judgements, they are disclosed, and revising them as evidence accumulates is what the standard asks for. The question is a different one: whether the evidence supports twenty-five years for the assets doing the work.
Buildings last. The things inside them may not
A data centre is not one asset. It is a shell, power and cooling plant, and computing equipment, and those three age very differently.
The shell plausibly lasts twenty-five years. Substations and chillers last a long time. Accelerators are a harder case — they sit in a market where each generation is substantially more capable per watt than the one before, and where the buyer's real constraint has become the chip and the power to run it rather than the building around them. Equipment can remain functional and stop being economic, and depreciation schedules answer the first question rather than the second.
If the useful life is right, the earnings drag is being spread appropriately across the years the assets genuinely produce. If it is long, the drag is being deferred into years that will also carry the depreciation of whatever gets bought between now and then.
What to watch instead of the capex headline
The spending numbers are the ones that get reported, and they are the least informative part. Three others say more.
The first is the useful-life disclosure itself, and any revision to it. The second is the gap between operating cash flow and free cash flow, which is where the build shows up before the income statement absorbs it. The third is what the assets earn — the point at which infrastructure spending stops being an experiment is when the revenue attached to it can be named.
Depreciation is not a controversy. It is a clock. It was started by decisions already taken, it runs whether or not the demand arrives, and the only discretionary part is how fast it is allowed to tick. That discretion is now being used, in public, in the disclosures — which is where anyone who wants to understand the cost of this build should be reading, rather than in the announcements about how much is being committed next.
Even the more exotic proposals for escaping the constraint, putting compute in orbit among them, do not escape this part. Whatever the hardware is and wherever it sits, somebody has to decide how many years it lasts, and then live inside that decision.
The capital spending totals, Meta's margin and free cash flow figures and the comments attributed to Meta's Susan Li and Microsoft's Amy Hood are as reported in coverage of the companies' most recent results and in Yahoo Finance's 4 August 2026 analysis of the four-company data centre bill. Estimates of the cumulative effect of useful-life revisions through 2028 are from independent analysis and are not company figures. The analysis is our own.
Topics aicapital spendingdata centres
