Microsoft: capex bear market or FCF inflection buy? Running the burger through the depreciation wedge

MSFT is down ~27% off its January 2026 high, to a 52-week low near $349 before bouncing to ~$390 — despite Azure growing 40%, the AI business hitting a $37B run rate (+123% YoY), and a fourth straight EPS beat. This isn't an earnings problem. It's a capex-return problem, and the market is currently pricing the pessimistic answer to one specific, checkable question: does capex plateau by FY27, letting free cash flow inflect, or does the AI arms race keep pushing capex higher than revenue can follow?

The trailing numbers already show the fork. Operating cash flow is $170.1B TTM, up 26% — the business itself is compounding fine. But capex is $97.2B TTM and running toward a ~$150-190B annualized pace (including finance leases), so free cash flow is stuck at $72.9B, essentially flat versus last year. Because roughly two-thirds of that capex is short-lived GPU/CPU hardware on a 3-5 year depreciable life, the resulting depreciation is about to grind through GAAP earnings for several more quarters regardless of what happens next — GAAP EPS will look "ugly" even in the bull case. The disagreement worth having isn't about GAAP; it's about whether the cash line turns before the market gets impatient.

The meat: Azure

Azure is the only driver that can actually break this thesis. It's currently growing 40% (39% CC), with AI services contributing roughly 12 of those points, up from 9 a year ago — meaning ~28 points of growth is still ordinary cloud migration, not an AI bubble metric. Azure is also supply-constrained, not demand-constrained: management has said growth would be higher with more capacity. The base case assumes an orderly deceleration — 40% now, ~32-35% by FY27-end, ~25% by FY29 — because a $250B+ run-rate business mathematically can't hold 40% forever. The tell to watch isn't the level, it's the slope: a smooth glide is maturation; an abrupt break below ~27-30% ahead of schedule, especially if it coincides with capacity going unused, would be the demand-side alarm that invalidates the whole thesis.

The wedge: capex growth (not the level — the rate of change)

This is the actual pivot variable. The FCF inflection isn't triggered by revenue accelerating — it's triggered by capex growth simply stopping. As long as OCF keeps compounding at ~15-25%/yr while capex growth decelerates toward flat, the gap between the two lines (i.e., FCF) widens fast and bends convexly upward. Q4 FY26 earnings (July 28, 2026) is the first real data point: the specific sentence to watch is how management frames the FY27 capex guide — "still ramping" keeps the thesis in bear territory; "moderating" or "plateauing" is the earliest confirmation of the bull/base case.

The bread: M365 Commercial cloud (funds the kitchen, isn't what's being built)

M365 Commercial cloud (+19%, 15% CC) is not what today's capex is for — that's Azure AI infrastructure. M365 runs on mature, cheap-to-serve software and carries the highest margin in the company. Its role here is different from Azure's: it's the demand/margin health signal and the high-margin cash that funds the buildout, so it governs FCF magnitude, not direction. If it decelerates, the FCF snap-back still happens (assuming Azure and capex behave), just with a shallower slope and a slightly worse blended margin as revenue mix tilts toward lower-margin Azure AI.

The garnish: Copilot (validates the bet, but it's hedged)

Copilot paid seats sit at roughly 20M, or ~4.4% penetration of the ~450M commercial M365 base — the lowest conversion of any major AI platform, even after seats jumped from 15M to 20M in a single quarter. Copilot's real job is validating that giving first-party AI priority claim on scarce GPUs was worth it. But because the compute fleet is fungible, a Copilot disappointment with Azure demand intact just means that capacity gets reallocated to paying Azure customers — a higher-ROIC outcome, and close to FCF-neutral. The failure mode that actually matters is Copilot and Azure demand softening together, which would mean the fungible-fleet hedge has nothing to reallocate into. Watch paid seats and M365 Commercial cloud growth (where Copilot revenue actually shows up); ignore total-user and DAU counts, which are engagement theater, not monetization.

Base case, in numbers

At a ~$385 entry, the base case assumes Azure glides to ~32% by FY27-end, capex growth decelerates to roughly 12%/yr, M365 Commercial cloud holds high-teens growth, and Copilot penetration reaches ~7% by FY29. That implies FY29 EPS near $24, a modest re-rate to ~24x (versus ~22x today, well below the ~30-32x five-year average), and a three-year target near $580 — roughly 15%/yr price return plus a ~1.5%/yr blended dividend-and-buyback yield, near 16-16.5%/yr total. Notably, that $580 three-year target sits only modestly above where Street consensus already expects the stock in twelve months (~$529-560), so the base case is giving the thesis two extra years to clear a bar it's nominally supposed to clear in one — that gap is the margin of safety.

What's already priced in

The ~27% drawdown has priced in real capex-return skepticism: forward P/E at ~21-22x is the cheapest since 2023, and the technical picture confirms the pessimism (below both the 50- and 200-day moving averages, a death cross, RSI in the low-to-mid 30s). What is not obviously priced in: the ~27% OpenAI equity stake worth roughly $135B, the OpenAI relationship's shift to a capped, non-exclusive structure through 2032 (de-risking the concentration story), and the mechanical fact that FCF doesn't need heroic revenue growth to inflect — just a capex plateau.

Risks worth tracking

The dominant risk is that the arms race doesn't stop: AWS and Google are building identical AI capacity, and if Azure's growth premium over them compresses at the same time capex keeps climbing, the entire capex-return bet is broken, not just delayed. Secondary risks: the EU's "Cloud Gatekeeper" designation under the Digital Markets Act creates a licensing/pricing overhang on Azure; component (DRAM/NAND) cost inflation could keep pressuring margin independent of the capex-growth story; and OpenAI, while now capped and non-exclusive, remains a large single-partner concentration inside the $627B RPO backlog.

Bottom line

This is a rerating story, not a broken business — the multiple compressed by roughly a third while earnings kept growing. The single checkable disagreement with the market is whether capex plateaus on schedule; everything else (Azure, M365, Copilot) is either confirming evidence or a magnitude modifier, not the trigger. At ~$385, I'm underwriting the bond-like base case (~15-16%/yr) with real upside optionality if the convex snap-back shows up, and I'm treating the July 28 FY27 capex framing as the first real checkpoint rather than a reason to wait longer.