Let's define the term before anyone gets to opine about it. Capital expenditure — capex — is money a company spends to acquire or build long-lived assets it expects to use for years, as opposed to the operating expenses it incurs running the business day to day. "Data center capex" is just that idea pointed at one category of asset: the land, buildings, power and cooling, servers, AI accelerators, and networking gear that constitute a data center. Accounting treats it differently from a salary or an electricity bill precisely because the asset lasts: the cash goes out now, and the cost is spread across the asset's useful life. On the cash flow statement, this spending shows up in the investing section, typically on a line called additions to — or purchases of — property and equipment.

That is not an abstraction; you can read the figure directly. In its fiscal 2025 annual report, Microsoft reports, in the cash-flow-statement line for additions to property and equipment, $64,551 million for the year — against $44,477 million and $28,107 million in the two prior fiscal years. Three numbers, one trajectory: the company's capital spending on property and equipment roughly doubled across two years. The 10-K is explicit about what is driving the line and that it expects more of it.

Additions to property and equipment will continue, including new facilities, datacenters, and computer systems for research and development, sales and marketing, support, and administrative staff.— Microsoft Corporation, Form 10-K (FY2025), source

Steelman the spending first, because it deserves it. A company that believes demand for AI services is real and growing has a rational reason to build capacity ahead of that demand: data centers take years to plan, power, and fill, so waiting until demand is proven means arriving late. The same Microsoft filing notes that its cloud business "benefits from three economies of scale," beginning with "datacenters that deploy computational resources at significantly lower cost per unit than smaller ones." Read straight, that is the bull case for capex in its own words — bigger buildouts can lower unit costs, so front-loaded spending is an argument about efficiency, not just optimism.

Why the number became the whole story

Here is the part the headlines circle. Capex is a spending decision made in the present against revenue expected in the future. The cash leaves now; the return — if it comes — arrives later, as the capacity gets used and monetized. That timing mismatch is the entire reason data center capex has become the most-scrutinized figure in AI: it is a large, disclosed, present-tense commitment whose payoff is, by construction, not yet on the books. A doubling of the property-and-equipment line is unambiguous in the filing. Whether the revenue that justifies it materializes on the expected schedule is the open question, and no single quarter's filing settles it.

This is also why capex is not the same as cost on the income statement. The cash outflow lands in the investing section of the cash flow statement in the year it is spent, but it is recognized as expense only gradually, through depreciation, over the asset's life. So a year of very large capex does not hit reported profit all at once — it builds an asset base that will depreciate for years afterward, raising the fixed costs the business must earn against going forward. The spending is visible immediately; its drag on margins is spread out. That accounting reality is part of why two observers can look at the same capex number and tell different stories about what it means for profitability.

It also helps to separate capex from two figures it is easily confused with. Capex is not the same as operating expense: a researcher's salary or a quarter's electricity bill is consumed in the period and expensed immediately, whereas a server or a building is capitalized and depreciated over years. And capex is not the same as free cash flow, the cash a business generates after its capital spending — in fact capex is subtracted to get there, so a quarter of very large data center capex mechanically pressures free cash flow even when operating cash flow is strong. That is why a rising capex line draws scrutiny from cash-flow-focused readers specifically: every additional dollar of buildout is a dollar that does not fall to free cash flow this period, on the promise of revenue in a later one. The filing discloses the buildout cleanly; whether the promise is kept is what later filings will or won't show.

How to read a capex figure without getting spun

A few distinctions keep the term honest. Capex disclosed as cash spent (additions to property and equipment) is not the same as capex committed — many filings separately list purchase commitments and construction commitments that represent future obligations not yet paid; Microsoft's 10-K lists construction commitments and purchase commitments among its contractual obligations, distinct from the cash-flow capex line. "Capex" also blends categories: a data center buildout mixes long-lived real estate with faster-depreciating accelerators, which age on very different schedules. And a rising capex line is a fact about spending, not a verdict about returns — the filing tells you what went out the door, not what came back.

Strip it to the load-bearing parts. Data center capex is the capital a company spends to build the physical capacity for AI, disclosed as additions to property and equipment in the cash flow statement, and recognized as cost only over time through depreciation. The reason it dominates the AI conversation is structural: the spending is large, disclosed, and certain, while the revenue meant to justify it is future and uncertain. The filings give you one half of that equation cleanly — Microsoft's $64,551 million is not in dispute. The other half, the return, is exactly the thing the documents cannot yet show, which is why the capex number is where the AI debate keeps landing.