Chapter 3

Return on AI Capital

Microsoft's AI build-out is backed by real, contracted demand — a $627 billion backlog and an AI run-rate near $37 billion — but the return on it is not yet visible in reported profit. Capital spending of $64.6 billion in FY2025 ran at nearly three times the $22.0 billion depreciation charge, so the cost of this capital has only partly landed. And a six-year server life, set in 2022 just as a rival now shortens its own for AI hardware, flatters today's margin. The evidence supports patience, not proof.

The cash-conversion gap this report is built around lives almost entirely in this line item. Chapters on the numeric spine established that reported profit is cash-backed at the operating line; the open question is whether the cash Microsoft is now redirecting into infrastructure earns an adequate return, or whether reported margins are being held up by how slowly that infrastructure is being expensed.

The spend, and how fast it is expensed

Capital expenditure has more than tripled in four years — from $20.6 billion in FY2021 to $64.6 billion in FY2025 [1]. Depreciation, the charge that turns that spending into a cost of doing business, rose far more slowly, from $9.3 billion to $22.0 billion, and actually fell in FY2023 despite a larger asset base [2] [3].

Source: FY2025 Annual Report (Form 10-K), Consolidated Statements of Cash Flows [1] and Property and Equipment note [2].

The gap between the two lines is the point. When a company adds assets far faster than it depreciates the ones it already owns, its fleet is young and the depreciation charge is still climbing toward the run-rate of spending. Property and equipment, net of accumulated depreciation, reached $205.0 billion at June 2025, up from $135.6 billion a year earlier; accumulated depreciation itself grew by $17.2 billion in the year, to $93.7 billion [4]. Management guided to roughly $190 billion of capital expenditure in calendar 2026, and stated it remains "confident in the return on these investments given higher demand signals and increasing product usage" [5]. On that trajectory, the depreciation line has years of catch-up ahead of it, and the reported operating margin of 46% reflects an asset base still only partly charged.

The demand behind it is real, and increasingly long-dated

The build is not speculative. The commercial remaining performance obligation — contracted revenue not yet recognized — stood at $627 billion at March 2026, up 99% year over year, and the AI business reached a $37 billion annual revenue run-rate, up 123% [6]. Azure grew 40% in constant currency in the same quarter, with demand described as exceeding available capacity [6].

Source: Q3 FY2026 Earnings Call, CFO prepared remarks [6].

Two qualifications keep that backlog honest. First, its growth is concentrated: excluding Azure commitments from OpenAI, commercial RPO grew 26%, not 99% [6]. Much of the headline figure rests on a single counterparty that is itself a heavy user of Microsoft's own capital. Second, the backlog is getting longer-dated: Microsoft expected to recognize about 40% of total remaining performance obligation within twelve months as of June 2025 [7], but only about 25% as of March 2026, with a weighted-average duration near two and a half years [6]. The demand is contracted, but less of it converts to revenue soon.

The depreciation question — a six-year life against faster-aging hardware

Whether this capital earns its keep turns partly on an accounting estimate: how long the equipment lasts on the books. Microsoft has twice extended that estimate. In FY2021 it lengthened server life from three to four years [8]; effective FY2023 it extended server and network equipment from four to six years, which is why depreciation dipped that year even as the fleet grew [9]. Its FY2025 filing still depreciates computer equipment over "two to six years" [10].

That estimate is moving in the opposite direction from a direct peer. Amazon shortened the life of a subset of its servers and networking equipment from six years to five, effective January 2025, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning" [11]. Alphabet depreciates servers and network equipment over roughly six years [12].

Sources: Microsoft FY2025 10-K [10] and Q4 FY2022 call [9]; Amazon FY2025 10-K [11]; Alphabet FY2024 10-K [12].

The tension matters more now than a year ago because the composition of the spending has shifted toward exactly the assets that age fastest. In the June 2025 quarter, management said more than half of capital spending went to long-lived assets supporting monetization "over the next fifteen years and beyond," with the remainder on servers, "both CPUs and GPUs" [13]. By the March 2026 quarter the mix had flipped: "roughly two-thirds of our CapEx was for short-lived assets, primarily GPUs and CPUs," a proportion the company expected to hold through calendar 2026 [6]. GPUs depreciated over five years rather than six carry a 20% higher annual charge on the same asset — arithmetic that Amazon's disclosure applies to itself and Microsoft, so far, does not. Microsoft does not break out a GPU-specific life, so the effect cannot be sized precisely from the filings; the direction, however, is not in doubt, and the cloud gross-margin trend already shows it — Microsoft Cloud gross margin fell to 69% in FY2025 "driven by the impact of scaling our AI infrastructure" [14], and to 66% by the March 2026 quarter [6].

Management's answer, and what would change the read

Microsoft's CFO addressed the risk directly. As spending pivots toward short-lived GPUs and CPUs, she argued, those assets are "generally aligned with the duration of the contracts," so the depreciation on them tracks the revenue they earn; the long-lived portion is datacenter capacity leased over 15 to 20 years that she is "very confident" will be used [15]. It is a coherent defense: if a five-year GPU serves a contract of similar length, matching the two is exactly what accounting is meant to do, and the concern is not that the assets are mismarked but that the six-year disclosed life may be generous relative to how the newest, GPU-heavy cohort actually performs.

One further item sits alongside the capital question rather than inside it. Microsoft holds about 27% of OpenAI, accounted for under the equity method, with $13 billion of funding commitments of which $11.8 billion was funded by March 2026 [16]. That stake has swung reported results in both directions: OpenAI-related losses cut net income by about $2.0 billion ($0.28 of EPS) in the first nine months of FY2025, while the first nine months of FY2026 carried a one-time $4.5 billion gain ($0.60 of EPS) from OpenAI's recapitalization [17]. OpenAI is thus both the largest single driver of the RPO backlog and a volatile line below operating income — a reason to read the AI return through operating metrics, not headline EPS.

The measured read: the capital is backed by contracted, fast-growing demand, so this is an investment build rather than a speculative one, but the return is not yet demonstrable in the numbers, and two things make today's reported margin flatter than the likely steady state — a depreciation charge running well below the pace of spending, and a six-year server life that looks generous against a now-majority-GPU fleet and a peer moving the other way. What would change the read, in either direction: the short-lived asset mix normalizing rather than rising further; Microsoft Cloud gross margin stabilizing instead of stepping down each quarter (guided to about 64%); depreciation converging toward capex without margins giving way; and commercial RPO excluding OpenAI continuing to compound in the mid-20s. Those are the lines to watch in each quarterly filing.