Microsoft moved $15B of capex off the reported line — in the same quarter free cash flow fell for the first time

Valuation — a reasonable P/E sitting on a 1.85% cash yield

The multiple depends entirely on which earnings number you use, and the trackers do not agree with each other.

The under-discussed number is that P/FCF of 54× against a P/E of 27×. The equity yields about 1.85% in free cash flow because capex is taking 63% of operating cash. On the P/E line the multiple is asking for roughly what management guided — ~14–15% NTM EPS growth, and the Q1 guide plus a ~45% Azure acceleration is consistent with that. The stretch is entirely on cash, not on earnings.

Analyst reaction was uniformly constructive and unusually dispersed: 56 analysts, mean target $562.73, median $550, high $870, low $400, with 40 Strong Buy / 13 Buy / 3 Hold / 0 Sell. Goldman went $610→$640; Wells Fargo reset to $650. I could not find a single post-print cut — but note that BofA ($500) and Phillip ($515) raised to levels barely above the current price, which is where the real disagreement lives.

AI-compute position — attacking the constraint that actually binds

Microsoft now reports capacity in gigawatts, not GPUs, and that unit change is the tell. It added ~1GW in the quarter, opened 31 datacenters (88 for the year), and says it is on track to roughly double total capacity in two years. Demand still exceeds supply, and management declined to say when that ends — or even to name whether the binding constraint is power, shells, GPUs, or land.

That reframes the silicon story. Maia 200 is an inference part, not a training part — TSMC N3, 10 PFLOPS FP4, 216GB HBM3e, and critically ~750W against Nvidia designs at 1,200W+. When your limit is a gigawatt rather than a purchase order, performance-per-watt is the capacity strategy. It is internal-only; enterprise customers cannot select Maia instances.

Caveat the skeptic pass insisted on: every custom-silicon performance claim in circulation this quarter — Microsoft's "30% better performance per dollar," Amazon's Trainium comparisons, Google's TPU numbers — is vendor-published and unrefereed. MLPerf Inference v6.0 (April 2026) drew 24 submitting organizations and included no Maia results and no current-generation TPU results. There is no neutral referee. Treat all of it as directional.

On competitive growth, the three clouds are not comparable without adjusting for base: Azure +43% off >$100B, AWS +37% to $42.2B, Google Cloud +82% to $24.8B. The Google figure is the one most often quoted without its asterisk — it includes the Wiz acquisition, closed March 2026, with no organic/inorganic split disclosed anywhere I could find. All three are now supply-constrained, which means the differentiator has moved from who has demand to who can energize watts fastest.

The OpenAI relationship is now genuinely two-sided. Microsoft holds ~27% as-converted at $135B carrying value, OpenAI committed $250B of Azure purchases — and Microsoft gave up right of first refusal on OpenAI's compute. OpenAI has since expanded AWS by ~$100B on top of an existing $38B. Products still ship first on Azure unless Microsoft cannot supply, but the exclusivity is gone.

Risk — six, isolated, not blurred

1. The reclassification question (the central risk). The calendar-2026 capex guide fell ~$190B → ~$175B because finance leases became operating leases, not because spending fell. Extending assumed asset life from 15 to 25 years makes a given lease term a smaller fraction of that life, which is the mechanism that moves the classification. Microsoft disclosed all of it on the call — but the headline number that reached most readers was "Microsoft cut capex."
2. The off-statement obligation. $329.1B of leases signed but not commenced, primarily datacenters, commencing FY27–FY33. That is roughly 3× Azure's annual revenue and it appears in none of the capex figures. Sources disagree on the prior-period base ($196.6B vs $92.7B) and I could not reconcile them against the filing.
3. Earnings quality. The FY26 OpenAI line swung $8.6B — from a $3.62B loss to a $4.96B gain — against ~$31.6B of total net income growth. The gain is a dilution gain from the OpenAI recapitalization: non-cash, non-operating, and reversible if OpenAI's valuation compresses. Separately, a $3.2B Anthropic gain sits inside the "clean" non-GAAP $4.74.
4. Depreciation, deferred rather than absorbed. FY26 depreciation was $34.3B, up from $22.0B and $15.2B. The useful-life extension pushes future depreciation out — and management guided FY27 operating margin "down less than a point." Real compression, pre-announced, small. Worth noting the extension covers buildings and shells, not servers or GPUs, so the classic "they stretched server lives" attack does not apply here.
5. Cloud margin compression. Microsoft Cloud gross margin fell 3 points to 65%. Amazon printed AWS margin expansion the following day. Same capex cycle, opposite margin direction — that contrast is the single most decision-relevant comparison in this report.
6. Related-party concentration. OpenAI accounts for somewhere in a 30–45% band of commercial RPO depending on which disclosure and quarter you read, and Microsoft does not disclose how much Azure revenue comes from OpenAI. The direction of travel is down as a share and up in absolute dollars.

Time horizon & position sizing (kept separate)

Horizon. This is not a next-quarter setup — the stock already moved ~25% in three sessions and the Q1 guide is in the price. The thesis resolves on a 3–5 year window: do the FY27–FY33 lease commencements convert to billed Azure revenue at a return above the depreciation that was just pushed out? Under two years you are trading sentiment on the capex headline and the mark-to-market swings on two private stakes. At five years you are underwriting Azure ROIC against a deferred depreciation schedule — the actual bet.

Sizing considerations (not a recommendation). MSFT now carries the same AI-capex-cycle beta as NVDA, AMZN and GOOGL, so it adds correlation rather than diversification to an AI-infrastructure basket. Reported earnings now swing on marks on two private companies — OpenAI and Anthropic — which is a new and lumpy source of non-operating volatility Microsoft does not remark routinely. And a ~1.85% free-cash-flow yield leaves less cushion than a 27× P/E implies.