"Microsoft Earnings, Microsoft vs. Meta, The Efficiency Payoff" — @BenThompson

Why this is in the vault

Direct follow-on to yesterday's Meta piece, read as a paired contrast: same author, same week, opposite market reaction (Microsoft shares +16% on earnings, biggest single-day market-cap gain in history, vs. Meta's post-earnings slide). Also flags a hyperscaler capex-accounting change (depreciation lifespan 15→25 years) that changes how headline capex numbers should be read going forward — relevant to the RDCO capex-tracking loop.

The core argument

Three linked pieces.

Microsoft earnings. Azure grew 43% (fastest since early 2022), crossed $100B revenue for the fiscal year, and Hood guided next quarter to ~45%. Capex was $41B for the quarter (below the $42B estimate) after Microsoft changed the assumed useful life of its owned data centers/office buildings from 15 to 25 years. Under ASC 842, that shift pushes newly-leased data centers below the 75% "major part of remaining life" threshold, so future leases get booked as operating leases rather than finance leases — meaning they no longer count toward reported CapEx, even though the cash liability is still real. Thompson: projected CapEx dropped from $190B to $175B on paper, but "Microsoft isn't on the hook for" less.

Microsoft vs. Meta. Thompson contrasts the two earnings reactions along three axes: (1) clarity of strategy — Nadella's opening remarks were four sentences, centered on keeping the model "harness" separate from the model itself so enterprises "don't outsource their core IP" (this is model middleware: stateless, swappable models behind a proprietary context/memory layer), versus Meta's harder-to-articulate frontier-model bet; (2) cost of frontier — Microsoft has stayed out of frontier-model training spend and rides inference demand instead (lowest CapEx-to-free-cash-flow ratio of any hyperscaler), while Meta is paying both to rent third-party compute now and build its own frontier buildout; (3) tangibility of application — Azure, the middleware layer, and Microsoft's own first-party apps are concrete demand sinks, where Meta's monetization story is comparatively abstract. Thompson's verdict: Microsoft's strategic clarity is easier when "the alternative is death" (an existential competitive threat from Anthropic/OpenAI replacing its enterprise position), while Meta's threat is longer-horizon and diffuse.

The efficiency payoff. Hood's framing: in a supply-constrained compute environment, efficiency gains (CPU/GPU fleet utilization, faster provisioning lead times) get "quickly monetized in quarter" rather than sitting as theoretical future upside. Thompson's read: efficiency is normally underinvested because payoff feels distant, but scarcity makes the payoff immediate — and closes admiring Microsoft's "grind" culture over Meta's discomfort-inducing enterprise pivot, while noting he still thinks Meta's underlying strategic position may be better.

Mapping against Ray Data Co

The capex-accounting mechanic is the concrete, actionable piece: investing-edgar-watch pulls hyperscaler capex straight from 10-Q filings on a QoQ/YoY delta basis, and this Update is a live example of how a pure accounting-policy change (useful-life reclassification triggering ASC 842 finance-vs-operating-lease treatment) can move a headline CapEx number by $15B without any change in actual committed spend. Any future capex-anchor pulse needs to flag depreciation-policy footnotes, not just diff the top-line number — Microsoft's move this quarter is exactly the kind of silent denominator shift that would make a QoQ comparison read as "deceleration" when the real liability didn't shrink. Worth a note in the next investing-edgar-watch run to check whether Amazon/Google make similar useful-life changes, which would suggest an industry-wide reporting-standard response to capex-fatigue investor pressure rather than a Microsoft-specific move.

Secondarily, Nadella's "harness kept separate from the model" architecture is the same defensibility shape already logged against RDCO's own agent stack in the 2026-05-06 Microsoft/Apple note (thin-harness/fat-skills, proprietary context as the moat) — this Update adds the efficiency argument as a reason that shape compounds: a harness-centric operator captures efficiency gains (faster provisioning, better utilization) that a frontier-model-centric operator like Meta can't, because Meta's spend is dominated by training rather than swappable inference. No new decision here, just reinforcement of a standing thesis.

Related

• [[2026-08-03-stratechery-meta-earnings-timing-financial-tail]] — prior-day Meta earnings piece; this Update explicitly reads Microsoft's reaction against it as a paired contrast
• [[2026-05-06-stratechery-microsoft-apple-earnings]] — prior Microsoft earnings Update; same author's harness/middleware thesis and "seats plus consumption" framing, directly extended here via Nadella's model-middleware argument

Sponsorship

No paid sponsor in this issue — single-author Stratechery Update. Only self-promo: an intro plug for the Sharp China podcast episode and the standard closing cross-promo block for the Stratechery Plus podcast stable (Sharp Tech, Dithering, Sharp China, Greatest Of All Talk, Asianometry). No bias implication beyond ordinary in-house cross-promotion.