Microsoft’s capex forecast lost $15bn. Read the accounting note.

Corrected September 13, 2026: the earlier version described an aggregate capital-spending estimate as AI-only spending and confused nominal and real Treasury yields. We removed the unsupported aggregate and historical comparisons, corrected the yield figures and rebuilt the analysis around identified sources. The original publication date is retained.

Microsoft’s calendar-2026 capital-spending forecast went from roughly US$190 billion in its April earnings call to about US$175 billion in its fiscal fourth-quarter call. The US$15 billion difference invites an obvious conclusion: less building, less money at risk. Microsoft’s explanation does not support that shortcut.

The company said in its fiscal fourth-quarter call that it was extending the estimated useful lives of data centres and office buildings from 15 to 25 years, effective in fiscal 2027. That changes how more future data-centre leases will be classified. Finance leases count in its reported capital expenditures; operating leases do not. Outside this effect, Microsoft said its calendar-2026 investment expectations were unchanged. These were building-life estimates, not a claim that GPUs would last 25 years.

The number and the obligation

The useful question is what changed in the underlying commitment. A smaller reported capex figure can come from cancelling capacity, paying less for equipment, changing delivery dates or changing which accounting category captures a lease. Those possibilities tell different stories about future costs and demand.

A hypothetical comparison makes the distinction plain. Two businesses might use similar buildings while one buys and the other rents. Comparing their upfront capital-spending figures alone would miss the rent the second business still has to pay. The businesses are not automatically equally risky; the contracts, duration and exit rights matter. The point is that one headline number cannot answer all those questions.

That is also why adding up company forecasts demands care. Readers need the same period, a defined group of companies and consistent treatment of leases. An aggregate for capital expenditure should not silently become an estimate for AI alone. Microsoft described its fourth-quarter hardware investment as serving both AI and non-AI infrastructure. The shared buildings and equipment make a clean label harder, not less necessary.

Strong earnings do not settle the investment case

There is evidence behind the market’s optimism. In its August 14 investment update, Bessemer Trust reported positive earnings growth in ten of eleven S&P 500 sectors. It said the index had returned 22% over the preceding twelve months while its forward price-to-earnings multiple fell from 22.4 to 20.2. Those are Bessemer’s reported figures for that period, not today’s market readings.

That supports a narrower argument: rising earnings helped support share returns. It does not establish the profitability of every new data centre. An existing business can earn substantial cash while a new project still depends on uncertain future utilisation, prices and costs. The investment case needs both sides of that ledger.

Financing comparisons need the same precision. On August 14, the U.S. Treasury’s ten-year nominal par yield was 4.68%; its ten-year real par yield was 2.41%. The real measure comes from inflation-protected securities. They are different benchmarks, not interchangeable descriptions of the same rate.

What would change the argument

For investors assessing this build-out, three comparisons are more useful than cheering or fearing a single spending total: capacity actually used against capacity installed; cash receipts against construction and contractual payments; and an asset’s expected earning life against its depreciation assumptions.

These are questions to investigate, not evidence that Microsoft has hidden a loss or that the AI boom must fail. A demand shortfall would matter differently from an accounting reclassification. So would a contract that leaves a customer paying for unused capacity. The documents have to establish which situation applies.

Our analysis is that the spending race is also a contest over measurement: what counts as investment, when the cost appears and whose obligation survives if expectations change. For the electricity side of that question, read our analysis of AI companies’ power-cost promises. For another distinction a large headline can blur, read our comparison of AI financing and company valuation.

The next time a forecast drops by billions, start with the footnote. It may change what the number means more than what gets built.