Phase Space AI

Microsoft Trade Construction 2026-07-27

Microsoft [MSFT]

Microsoft Corporation [MSFT] — Trade Construction & Risk Management

Task 5 · investment-memo skill v1.4.0 · 2026-07-27 · spot $389.10

---

Updated 2026-07-29 — Criteria + two-horizon valuation. The numbered Gates are retired and replaced by named Criteria with types (BINDING / MEASURED) returning PASS / FAIL / INDETERMINATE. This memo no longer issues a position verdict — no Long, Short, Watchlist or Avoid. The expected-return-versus-cash-hurdle test is replaced by a reverse-DCF implied-path test plus a 12-month target, with sensitivity run over the exit multiple rather than over scenario probabilities. Momentum is entry timing only and vetoes nothing. Sections below that predate this update are retained as the historical record and are annotated where superseded.

Reading the retired "Gate N" numbering. Any Gate N below is historical. The mapping is: Gate 1 → Quality Criteria (BINDING) · Gate 1b → Short Mechanism Criteria (MEASURED) · Gate 2 / 2A / 2B / Path B → dissolved, absorbed by the Valuation Criteria · Gate 2C → Peer Spread Criteria (MEASURED) · Gate 3 → Catalyst Criteria (MEASURED) · Gate 4 / 4a → Valuation Criteria (BINDING) · Gate 5 → Liquidity Criteria (BINDING) · Gate 6 → Momentum Criteria (MEASURED, entry timing only) · (new) Downside Criteria (MEASURED). A Criteria returns PASS / FAIL / INDETERMINATE; a missing input is INDETERMINATE, never FAIL. Historical calibration items keep the old numbering on purpose — the record is the record.

0. Portfolio context — the portfolio-book contract, executed

portfolio_book.json read at 2026-07-27.

0.1 Correlation — computed from actual daily returns, and the result contradicts the prior

The brief anticipated that "a hyperscaler long is correlated with all of them." Computed from trailing daily returns, that is wrong, and the finding is worth more than the assumption it replaces.

Trailing 1-year daily-return correlation (252 sessions to 2026-07-27, Alpaca SIP):

Trailing 6-month (the live AI-capex-fear regime) — the divergence widens:

MSFT's mean correlation to the six-name AI-infrastructure book is +0.08 over one year and slightly NEGATIVE over six months. The within-book mean pairwise correlation is +0.35. MSFT–GOOGL is only +0.13.

Why, and it is not a statistical accident. The underwriting book (MU, NBIS, NET, SNDK, CIEN, AAOI) is composed of AI-capex receivers. Microsoft and Alphabet are AI-capex payers. When Alphabet raised capex guidance on 22 July, GOOGL fell 7.1% — and that same news is revenue for the memory, optics and neocloud names. They are on opposite sides of the same trade, which is precisely why the six-month correlations turn negative.

Portfolio implication, stated explicitly: a hyperscaler long would be a genuine diversifier against the existing watchlist cluster, not an addition to it. That is a real point in its favour — and, exactly as with ISRG, diversification cannot rescue a position whose standalone expected return is negative. It is a reason to prioritise the name if and when its E[R] clears, not a reason to own it at $389.10.

0.2 Capital competition — moved to the strategy layer

The cash-hurdle test is retired (references/criteria.md: "The cash hurdle, replaced by slot competition in the strategy"). Whether this analysis wins a slot against the book's alternatives is a question for the book, not for the memo, and the memo does not answer it. The portfolio facts above are recorded because they are inputs a strategy would need — not because a conclusion is drawn from them here.

---

1. Four conclusions, kept separate

1. Fundamental. Strong and improving. Revenue +17.0% (FY2026E), operating margin expanding to 46.6%, commercial RPO $627bn (+99%), Azure guided to +39–40% cc, clean accruals (−0.055), high gross profitability (0.313), headcount falling while capex rises 84%. There is no deterioration mechanism visible anywhere in the data.
2. Expectations. House FY2026E revenue and EPS land exactly on consensus ($329,632m vs $329,600m; $16.77 vs $16.77). House FY2027E EPS of $18.32 is 7.5% below the $19.80 mean but sits inside the published $17.89–$21.74 range, at the 11th percentile. ΔE = house − Street is approximately zero for FY2026 and modestly negative-but-inside-the-range for FY2027.
3. Valuation. DCF $357.11 (−8.2%), comps $376–$421, scenario-weighted $377.59 (−3.0%). Every method lands within ±8% of spot. Fair value.
4. Portfolio. Not the memo's call. What is recorded for a book to use: MSFT is a genuine diversifier against the AI-supplier cluster (mean one-year correlation +0.08, negative over six months), realised volatility is 27.6%, and a binary event sits two days out with front-week implied volatility at 2.6x realised — an execution consideration, not a verdict.

---

2. THE CRITERIA (2026-07-29 — the numbered "Gates" are retired)

Per ~/.claude/skills/investment-memo/references/criteria.md, every test is a named Criteria carrying a type, and returns PASS / FAIL / INDETERMINATE.

• BINDING — a long-only absolute-return book treats a failure as disqualifying.
• MEASURED — always scored and stored; informs timing, sizing or a future strategy. Never blocks.

A missing input is INDETERMINATE, never FAIL. nan > 0.50 silently evaluating to FAIL rejected two live names on 2026-07-29 (calibration item D1, 20 recorded instances). An absent field is not a rejection.

The memo scores every Criteria and blocks on none of them. The BINDING/MEASURED labels describe the long-only absolute-return strategy that consumes this analysis — not the analysis itself.

2.1 What changed versus the previous version of this section

3. ANALYSIS SUMMARY — no position verdict

This memo does not output "Long", "Short", "Watchlist" or "Avoid". The prior version of this section issued INVESTMENT DECISION: WATCHLIST (bias: long-if-converted). That line has been removed. The evidence it rested on is retained above; the conclusion is not the memo's to draw, because whether an analysis justifies a position depends on the book that would hold it — its existing exposures, its slot competition and its risk budget.

MSFT — Microsoft Corporation                          framework: Criteria + two-horizon valuation (2026-07-29)
Spot $393.44 (2026-07-28)                             Archetype: COMPOUNDER

BINDING    Quality ................ PASS
           Valuation .............. PASS   required 5.0% vs demonstrated 12.4%  ->  margin +7.5pp
                                           exit 25.1x EV/EBIT, GROWTH_MATCHED (n=32, growth 6.7-18.3%)
                                           implied compression: +5.2x / +26.4% EXPANSION from today's 19.9x
                                           sign flips at ~17.5x exit; PASSES at today's multiple unchanged (+2.5pp)
           Liquidity .............. PASS   Sep-2026 chain: 38 strikes with OI, median 2,905, total 179,274
                                           median quoted spread 3.1% of mid, size 121 bid / 107 ask

MEASURED   Momentum ............... bottom quintile (12-1 -28.1%, 12.2nd pctile, RSI-14 56.6) - TIMING ONLY
           Catalyst ............... FQ4 FY2026 earnings 29-Jul-2026; FY2026 10-K ~30-Jul-2026
           Downside ............... $252 (-35.9%) - 4-year server life + FY2027 capex >$190bn -> 16x on -14% EPS
                                           de-rating not impairment; no going-concern case; +$46.8bn net cash
           Consensus .............. house FY2027E EPS $18.32 vs Street $19.80 (11th pctile, inside range)
           Peer Spread ............ INDETERMINATE - no same-end-market peer multiple history assembled
           Short Mechanism ........ FAIL (no short mechanism: revenue +17.9% YoY, margin expanding)
           Sub-sector ............. LC software / hyperscale cloud; AI-capex PAYER (mean 1y corr +0.08 to the
                                           AI-supplier cluster, negative over 6 months)

12-MONTH TARGET   $710  (+80.5%)   own six-year median 13.97x EV/Sales on NTM revenue $375.2bn
                  floor $463 (+17.6%) at today's 9.06x, no re-rating
                  MSFT trades at the 3rd PERCENTILE of its own six-year EV/Sales range
                  Street $557.25 (56 analysts, +41.6%) implies 10.93x - inside the band.
                  The Street and this analysis differ on multiple reversion, not on revenue.
                  CAVEAT: full reversion to a six-year median inside 12 months is the UPPER END of
                  what that history supports, and is labelled as such rather than as a central case.

INVALIDATION      $252 (Downside Criteria realised)
CORRECTIONS       scan TTM revenue $311,898m -> $318,273m (+2.0%): the scan's "TTM" skips Q4 entirely.
                  scan valuation FAIL (-11.8pp) -> PASS (+7.5pp): the scan's solve embedded a 20.0%
                  terminal margin against Microsoft's actual 45.6%.
DISCLOSED GAPS    NTM revenue is corrected-TTM YoY held flat, NOT a consensus estimate (Alpha Vantage
                  quota exhausted). Peer Spread INDETERMINATE.

---

4. Thesis-invalidation triggers (checkable, dated)

For the long-if-converted bias: 1. 29-Jul-2026: FY2027 capex guided above $190bn → the payback arithmetic in Task 1 §4.4 breaks. 2. 29-Jul-2026: FQ4 Azure constant-currency growth prints below 35% against a 39–40% guide. 3. 29-Jul-2026: FY2027 Microsoft Cloud gross margin guided below 62% (FQ4 guide is ~64%). 4. Any quarter: commercial RPO ex-OpenAI growing below 20% y/y (it was +26% in FQ3-26). This is the cleanest single read on whether demand is broad or concentrated. 5. Any quarter: the disclosed 12-month RPO conversion percentage falling below 25% (10-Q R70). 6. FY2026 10-K (~30-Jul-2026): any shortening of the disclosed server/network useful life from 6 years. 7. Any quarter: PP&E sitting unpaid in accounts payable rising above ~$30bn (it was $22.6bn at Mar-2026, from $6.9bn a year earlier) — reported FCF would be increasingly overstated. 8. Q-M7, unresolved: any disclosure that quantifies Teams' position. Its 87% collapse in management attention is the one signal in this memo I could not explain.

Conditions that would materially change this analysis (not upgrade triggers — no verdict exists to upgrade): - Price route: a close at or below $310.00, at which the base case delivers +24.1%. - Variant route (price-independent): ALL THREE of — (a) Microsoft discloses a commercial Copilot seat count and ARPU for the first time (withheld for eleven quarters; sell-side previews name it as one of three things wanted on 29-Jul); (b) commercial RPO ex-OpenAI accelerates above 30% y/y for two consecutive quarters; (c) a weekly close back above the 200-day moving average (~$434) with a higher-high sequence, fixing Gate 6. These are falsifiable predictions, kept because they are checkable. They are no longer framed as conversion conditions, because there is no position verdict to convert.

---

5. Vehicle selection — two compared, with a recommendation

No catalyst-dated tactical options recommendation is made. Task 6 identifies a hard catalyst two days away, and the reference file says to produce a specific contract when a dated catalyst exists. It also says the expected move must exceed the implied move. At 86.5% implied against 33.3% realised it does not, and issuing a contract recommendation into that would be exactly the "buying overpriced pre-catalyst options" failure the skill's own edge hypothesis says this system exists to prevent. Declining to recommend a contract is the output.

---

6. Answering the seven required questions

---

7. The pre-flight verdict, scored

The Stage C.6 pre-flight, recorded before any modelling, said: variant unlikely; the honest prior is that Gate 2 will fail; spend the effort on unpublished scoping, the units build and the implied-penetration statement.

It was right. Gate 2 failed on both branches. The effort that produced value was: the 18-quarter mention-frequency series (which dated the OpenAI commitment from the word RPO alone, and corrected a seeded prior about power), the R-file pass (which found $22.6bn of unpaid capex and overturned four assumed disclosure gaps), the correlation computation (which overturned the assumption that hyperscalers correlate with the AI-supplier complex), and the units-based implied-penetration statement. The effort that produced nothing was the search for a variant — and fifteen minutes of honest pre-flight predicted that correctly.

The pre-flight was right that no variant existed, and wrong about what follows from that. Under the current framework the absence of a variant is not a finding against the name at all — no variant versus consensus is required, Path B is dissolved, and the Valuation Criteria now PASSES on the same evidence base that produced a Gate 4 FAIL. The label the old framework produced was a run of consecutive Watchlists; the diagnosis in CALIBRATION_WATCH.md was that the sourcing might be wrong rather than the gates. On this name the gates were wrong — specifically, a hardcoded 20% terminal margin against a 45.6% actual, and a range run over scenario probabilities rather than over the exit multiple. That correction is the substantive output of this update.