Microsoft

Refresh delta (since 2026-07-22)

1 · The CY2026 capex guide moved $190B → ~$175B — and it is NOT a cut. Hood, verbatim:

"Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged. However, the shift from finance to operating leases adjusts our expectation to approximately $175 billion."

The mechanism is a useful-life extension on datacenters and office buildings, 15 → 25 years, effective at the start of FY27. Longer lives push new datacenter leases from finance to operating classification, and only finance leases sit inside capex. So ~$15B leaves the reported number while the spending intent does not move. The desk's big-four aggregate must not read this as a $15B reduction — it is a definitional boundary shift, and from FY27 an increasing share of Microsoft's datacenter spend will not appear in capex at all. [primary: FY26 Q4 transcript]

2 · Q4 capex $41B; the finance-lease share kept falling. $35.8B cash paid for PP&E + $5.6B finance leases = $41B, as stated. Lease share 13.7%, continuing the 32% → 18% → 15% decline the prior dossier tracked. Roughly two thirds of capex was short-lived assets — "primarily CPUs and GPUs." [primary]

3 · RPO — the prior dossier's central worry is falsified. It asked whether a third flat quarter would show the backlog to be "an OpenAI artifact," and set the bar at "$20B+ sequential says the ex-OpenAI base is still compounding."

"Commercial remaining performance obligation grew 84% to $678 billion. All sequential commercial RPO growth was driven by commitments from customers outside of frontier model companies. And RPO increased 25% when excluding [OpenAI]."

$633B → $678B, +$45B sequential — more than double the bar, and Microsoft attributes all of the sequential growth to non-frontier-model customers. Commercial bookings +18% ex-OpenAI, +10% (+11% cc) including it. The artifact hypothesis is dead on the company's own attribution. [primary]

4 · Azure accelerated, and the guide is above the prior track. Azure surpassed $100B annually, +41%. Q1 FY27 guide: ~45% growth in constant currency — against the +39–40% the prior dossier was tracking. The "modest acceleration in H2 CY2026" language was not merely reaffirmed; it was given a number, and the number is higher. [primary]

5 · Free cash flow $19.6B, and the FY27 guide is unusually thin. Q4 FCFFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. $19.6B, "reflecting higher capital expenditures." For FY27 the commitment is only that Microsoft expects "to remain free cash flow positive." For a business that returned $43B to shareholders in FY26, guiding to positive rather than to a level or a growth rate is a notable choice of words, and it is the disclosure the prior dossier flagged as worth watching in its own right.

6 · New, and not on the prior dossier's list: a $3.2B gain on Microsoft's investment in Anthropic.

"These include a $3.2 billion gain from our investment in Anthropic and lower-than-expected expenses related to the Voluntary Retirement Program, which were partially offset by severance expense and impairment charges in XBOX."

Part of a +27c discrete benefit to diluted EPS. Microsoft holds an Anthropic position material enough to move the quarter — a structural fact absent from both prior dossiers.

---

Phase B — Measure performance

Earnings result (FY2026 Q4, reported 2026-07-29)

All [primary: FY26 Q4 earnings call transcript, Microsoft IR, 2026-07-29].

Discrete items: +27c on diluted EPS versus the April guide — the $3.2B Anthropic gain plus lower-than-expected Voluntary Retirement Program expense, partly offset by severance and XBOX impairment charges.

Earnings calls (sentiment trend)

Against the three transcripts ingested last pass, the shift this quarter is from capacity-constrained to cost-engineered. Prior calls framed AI as a supply problem — can we energise enough capacity. This call spends its density on cost per unit of capability, and quantifies it:

• MAI-Code-1-Flash delivers "comparable quality to GPT-5.6 for the most common tasks, while operating at significantly lower cost"
• MAI-Voice-2-Flash and MAI-Image-2.5 cut GPU costs in Dynamics 365 and PowerPoint, the latter "up to 84% reduced GPU costs"
• GitHub Copilot: "higher code acceptance rates and 10% lower median token usage, while still having access to frontier capabilities from OpenAI and Anthropic"
• Cobalt: "30% better performance per dollar than the latest-generation hardware in our fleet"

This is a hyperscaler quantifying, in production, the substitution of smaller in-house models for frontier calls on cost grounds. It is first-party evidence for the specialization thesis from the party with the least incentive to make it — and it belongs in kb/models/wiki/concepts/open-weight-frontier-parity.md, which currently rests entirely on DeepSeek and Kimi.

Also notable: Microsoft now serves Anthropic and xAI models in Foundry (11,000+ model catalog), reports a 5× increase in customers building with models from multiple providers, and cites Levi Strauss running OpenAI and Anthropic. The multi-model default is being disclosed as a trend, not a concession.

Stock-price catalysts (updated)

The print supplied three re-rating inputs at once: the RPO re-acceleration (+84%), the Azure guide raise (~45% cc), and a capex headline that reads like a cut but is not. The third is the one most likely to be mispriced — a reader who takes $175B at face value against $190B sees a retreat from AI spend that management explicitly denied in the same breath.

Forward arithmetic (mechanics only — NOT a projection)

Per the prior dossier's gate, no EPS path is issued. The mechanics that changed:

• Q1 FY27 capex guide: "over $50 billion" — a floor, not a point, and explicitly "including the lease reclassification impact from the useful life update."
• FY27 operating margins "down less than a point"; effective tax rate ~20%; FCF "positive."
• Intelligent Cloud Q1 guide $40.95–41.25B, +33–34%; Azure ~45% cc.
• The comparability break is now dated. From FY27, the useful-life change moves datacenter leases out of capex. Any FY26→FY27 capex series is measuring two different things, and a growth rate across that boundary is meaningless without the adjustment. This is the single most important thing for the capex dataset to carry.

---

Phase C — Judge people & books (partial re-run)

Forensic red flags (re-run on the capex/lease axis)

The useful-life extension is the item to watch, and it is disclosed, not hidden. Extending datacenter lives 15 → 25 years reduces depreciation per period and flatters operating margin — while simultaneously moving lease spend off the capex line. Both effects run the same direction: reported profitability improves and reported capital intensity falls, without the underlying spend changing. Management stated the change plainly and quantified the capex consequence, which is the right disclosure posture. The risk is not concealment; it is that downstream readers propagate the new numbers into old series.

Two of the prior dossier's open items remain open and now have a place to be resolved: the $3.0B nine-month gap between Hood's stated finance-lease figures and the XBRL series, and the ~$10B gap between derivable and management-definition CY2025 capex. The FY2026 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes. (filed 2026-07-30) is on EDGAR and is where the lease note and the useful-life disclosure will settle both. Not read this pass — the transcript was the priority.

The FY2026 10-K — read this pass after all (filed 2026-07-29)

The 10-K was pulled onto the shelf and read. It closes two of the prior dossier's three named checks, and produces the most consequential number in this refresh.

$329.1 billion of leases not yet commenced. Verbatim:

"As of June 30, 2026, we had additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion, with some arrangements subject to certain contractual conditions being met. These leases will commence between fiscal year 2027 and fiscal year 2033 with lease terms of 1 year to 20 years."

Set that beside the capex restatement and the two facts fuse into one:

• The useful-life extension moves new datacenter leases from finance to operating classification.
• Only finance leases sit inside capex.
• $329.1B of datacenter leases are queued to commence from FY2027 — exactly when the new classification takes effect.

So Microsoft's reported capex is becoming a progressively worse proxy for its datacenter commitment, and the 10-K quantifies how much is queued behind the line. This is not an accusation — every element is disclosed, and management stated the capex consequence on the call. It is a measurement warning: anyone tracking hyperscaler AI spend through the capex line is about to under-count Microsoft by an amount that dwarfs the $15B restatement.

Check (a) — server/network useful lives: NO lengthening. The 10-K's stated lives are "software developed or acquired for internal use, three years; servers and network equipment, two to six years; buildings and improvements, five to 15 years." Servers and network equipment are unchanged. The extension is buildings-only.

That distinction matters and cuts in Microsoft's favour. The classic hyperscaler accounting concern is lengthening server lives to flatter margins on the fastest-depreciating asset. That is not what happened here — the change is on buildings, which genuinely do last 25 years, and the fastest-depreciating class was left alone. The FY2026 10-K still shows the old 15-year buildings figure because the change is effective FY27, which conveniently pins the baseline.

Contractual obligations, for scale [primary: FY2026 10-K]:

Check (c) — Item 1A on customer concentration and backlog cancellability: not read this pass.

---

Lenses carried forward unchanged

Lens 1 (overview) · 2 (supply chain) · 3 (moats) · 4 (segments — Q4 splits not yet extracted) · 7 (comps) · 9 (management — no changes this quarter) · 12 / 13 (bull-bear, devil's advocate — withheld under the Socratic gate) · 14 (management questions). See the previous dossier.

---

What the next refresh should do

1. Item 1A — the one 10-K check still unread: newly quantified customer-concentration or backlog-cancellability language.
2. Resolve the two lease-definition gaps against the 10-K's detailed lease note (the note is now on the shelf; the reconciliation was not attempted this pass).
3. Extract Q4 segment splits into our figures and the balance-sheet block into our figures — neither was updated this pass, so the workbook still opens on FY26 Q3 data.
4. Watch whether the AI voluntary disclosures continue. Copilot paid seats went 20M → >30M with net adds "more than doubling quarter-over-quarter." The quarter these stop being given is the quarter to pay attention.

---