Alphabet.
Segments, capex and value.

Purpose. This is a quality sample of company research. It states findings and their evidence.

Fair value. Our central estimate is $187 per share. The 21 Oct 2026 close of $331 is 76.7% above this value.^S14S15 Two practitioner checks give $167 to $239 (cross-check).

Risk measure. Scenarios use real company years. A FY2022–23 trough regime gives $102. The H1 2026 run-rate regime gives $283. The range is 97% of the central value. The price is 17.1% above the run-rate case. A stress with trough operations and permanently high capex gives $66; it is outside the weights.

The business. Alphabet is an exceptional franchise in a heavy investment phase. Services earned 41.9% operating margin over the last twelve months. Cloud revenue grew 73% in H1 2026 and carries a $513.9bn backlog.^S1S2 Capex reached $132.4bn in the twelve months to June; free cash flow fell to $53.3bn.^S2

Why the value moved from $104. The earlier note applied 12 × FY2025 EPS less 20%. It held no growth, no cash and no cost of capital. Our no-growth earning power is $102 per share, close to the old figure. Net cash and investments add $18. Growth in the central path adds $67 (section 16).

What changes the view. Three facts move the value most: the pre-depreciation margin Cloud earns at scale, the capex intensity after 2027, and the discount rate. One point off the WACC gives $226. A 55% Cloud margin gives $210. Terminal capex at 12% of revenue gives $201. Two thesis breaks sit outside the weights: AI search substitution of 20% gives $162; loss of the main Cloud customer (Anthropic by press reports, Unverified) with the related stake gives $150; both give $124 (thesis breakers). Watch paid clicks, Search & other growth and the Cloud backlog each quarter (leading indicators).

Review status: provisional. Two blind rounds ran. Round 1 closed one resolvable fact (the trough Cloud path sat below contracted backlog) and moved the Bear case; the central value did not move. Round 2 found no new resolvable fact, but the rule needs two quiet rounds in a row, and Q3 2026 results (projected 28 October) will test Cloud margin, capex and investment marks. Open items: Cloud backlog concentration and whether the unnamed investee is a Cloud customer; cash versus book tax. Version 1.1 adds a six-angle blind-spot pass; three filed liabilities entered the bridge.

Thesis intact. No operating-results filing followed the Q2 10-Q of 23 July. The capex guide rose, debt rose, and the discount rate rose. None of these changes the regime cases; each sits in a model input.

Channels: estimate, multiple, discount rate, flows, read-across. An event without a mechanism is sentiment and does not enter the value.

Two practitioner checks sit beside the cash-flow value. Each uses Alphabet’s own history and reported figures. The earlier flat rule is shown for the record.

Where they disagree. The P/E check is +27.7% from the DCF. Historical multiples were set when the 10-year Treasury yield was lower and capex was 10–15% of revenue.^S11S12 The owner-earnings check capitalises today’s NOPAT and ignores the capex build. The DCF charges that build in full. All three methods agree on direction: the price is above each value. At the price, the core P/E is 33.2×.

Alphabet Inc. reports three segments: Google Services, Google Cloud and Other Bets. AI research costs sit at group level.^S1 The security here is Class C stock (GOOG, Nasdaq), CIK 1652044.^S10

Language models write fluently but compute and recall probabilistically. This method confines them to reasoning. Data and arithmetic stay outside the model.

Company class. A classification step runs before data collection. Alphabet has three segments with different economics, so it takes the sum-of-parts route. Each segment is valued on its own cash flows. A single group ROIC and a group Faustmann ratio are not valid tests for this class; section 15 shows them for comparison only.

Two-phase review. Phase A closes resolvable facts from primary data before any judgement round: the statutory tax rate, the dated 10-year yield, a regression beta and a dated equity premium. Phase B runs blind rounds. A result is stable when two rounds in a row keep the conclusion, move the central value by less than 1% and find no new resolvable fact.

Review status: provisional. Two blind rounds ran. Round 1 closed one resolvable fact (the trough Cloud path sat below contracted backlog) and moved the Bear case; the central value did not move. Round 2 found no new resolvable fact, but the rule needs two quiet rounds in a row, and Q3 2026 results (projected 28 October) will test Cloud margin, capex and investment marks. Open items: Cloud backlog concentration and whether the unnamed investee is a Cloud customer; cash versus book tax. Version 1.1 adds a six-angle blind-spot pass; three filed liabilities entered the bridge.

Signal strength. Strong: audited and filed figures (levels 1–3). Medium: dated market data and practitioner datasets. Weak: press reports and trackers (section 18).

Weights. Scenario weights are the reader’s judgement. The calculator in the Summary shows how the weighted value moves. This makes the judgement visible.

Compared with typical AI research. Typical output states numbers from memory and one valuation. Here every figure carries a source and date, the arithmetic is reproducible, and the conclusion survived blind challenge.

v1.4 (22 Oct 2026). Re-laid out to the price-first contract: what the price implies comes first, then thesis breakers, fallback use, market test, entity status and a leading-indicator table built from six quarters of filings. Per-share values are rounded to whole dollars. Anthropic is named where press or its own release ties it to Cloud or the private stake. Values did not change.

v1.3. Market-size test, a rental route for stranded Cloud capacity and the identity of the stakes behind the Q2 gain (market test, fallback use, entity status). Model cases unchanged; rental adds a stress variant.

v1.2. Two thesis-break scenarios added outside the weights (thesis breakers). Bear, Base and Bull did not change.

v1.1: six reviewers worked separately from the same facts, without the conclusion: bull, bear, regulator and legal, competitor and AI, capital and credit, accounting. Each named facts missing from v1. Only filed evidence moved an input.

Inputs changed: the bridge now subtracts accrued fines ($17.356bn), the July DMA fines (about $1.0bn) and the booked backstop liability ($0.8bn); Bear also takes a 10% loss on the $51.4bn backstop exposure. Operating forecasts did not change. A capital-return claim of $28.3bn H1 2026 buybacks was checked and rejected: the 10-Q shows none.^S1

We solve the central model for the growth that makes value equal the $331 price, holding margins, capex rules and the discount rate.^S24

Implied revenue: about $2,032bn a year in 2035, nominal US dollars: Services $909bn, Cloud $1,123bn. In 2030 it requires $1,160bn a year. The Base path reaches $1,202bn in 2035; FY2025 revenue was $403bn.^S3

Implied growth: 17.6% a year from FY2025 to 2035, against 11.6% in the central path.

Implied discount rate: 7.1% at central operations, against our dated WACC of 9.8% (10-year Treasury 5.1% on 21 Oct 2026).^S12

Implied market share: Cloud at $1,123bn is about 52% of a 2035 cloud infrastructure market of $2,164bn, against 15% today; Services advertising is about 36% of 2035 global advertising of $2,196bn, against about 26% in 2025. Both 2035 markets extend the dated estimates below at 15% and 6% a year (calculated by us).

The share the price needs is the weak link: about a third of global advertising and about half of cloud infrastructure by 2035. No market estimate we found supports a cloud share near half. The full grids sit in section 14.

Each breaker starts from the Base path and changes one mechanism. They are stress tests, not weighted cases. All loss sizes are our assumptions.

Reading. The price is $144 above Base. S1 at 20% widens that gap by $26; S2 with the stake markdown by $37. We hold no probability for either. The gap between price and Base comes from growth and the discount rate (section 01), not from these risks.

If one large customer leaves (S2), stranded capacity can be rented out wholesale. We test the $122.8bn of assets not yet in service.^S1

• GPU rental index, 7 Oct 2026: median on-demand H100 rent $4.20 per GPU-hour across 16 providers; spot fell from $6.62 per GPU-hour (late 2023) to $2.83 per GPU-hour (Feb 2026, Unverified). ^S39
• Peer filing, Mar 2025: one buyer was 62% of a GPU cloud’s 2024 revenue, so rental demand itself is concentrated. ^S36

Assumptions, not company data: rent revenue equals 35% of capacity cost in 2027, falls 15% a year, earns a 60% cash margin, and lasts 6 years. Wholesale term rent sits below the on-demand median.

Fallback value per share: $6 ($68.2bn of present value). S2 with the stake markdown moves from $150 to $155; S1+S2 from $124 to $130. Rental recovers about 15% of the S2 loss, because rent per unit falls each year.

Does the market the price needs exist?

Thesis. AI spending grows into trillions a year by 2030, and the Cloud backlog of $513.9bn shows contracted demand.^S1S40

• Counter-argument: compute prices fall. H100 spot rent fell from $6.62 per GPU-hour to $2.83 per GPU-hour between late 2023 and February 2026 (Unverified), so revenue per unit of capacity falls.^S39
• Counter-argument: open models. Open models lower the price of AI services; enterprise return on AI spending is not in any filing we read.
• Counter-argument: power. Utility connection quotes of four to ten years limit capacity.^S42
• Counter-argument: history. Long capex cycles in telecom networks ended in overcapacity; we cite no figure for that history.

Market test, calculated by us. If global advertising grows 6% a year from $1.3tn (2026), it reaches about $2,196bn a year in 2035. Advertising is 86% of Services (FY2025), so Services at $909bn needs about 36% of global advertising, against about 26% in 2025. If cloud infrastructure grows 15% a year from $574bn, it reaches about $2,164bn a year in 2035; Cloud at $1,123bn needs about 52% of it, against 15% today. Both growth rates are our assumptions.

Conclusion. The markets themselves can plausibly reach the needed size. On present evidence the price needs a share gain the evidence does not show. Effect on value: none; this tests the price, not the model.

The 10-Q says the $99.0bn Q2 gain came “primarily” from SpaceX and “a private company”.^S1 SpaceX is not the private company: it listed on Nasdaq (SPCX) in June 2026 at $135 per share.^S38 The largest private stake sits in non-marketable securities, where $87.9bn was remeasured in Q2. The 10-Q does not name it; press reports name Anthropic (Unverified).^S1 Effect on value: none; the bridge already carries both lines.

For each thesis breaker: the metric, its last six reported values, the level that would change the view, and the next report. Values come from the 10-Q and 10-K; growth rates and margins are calculated by us from reported dollar figures. Q4 values are the annual figure less nine months. The company does not split traffic acquisition cost for Search in its filings; paid clicks and cost per click are the closest public proxy.

Advertising and search

Cloud

Market size

Capex and valuation inputs

GOOG is Alphabet Class C capital stock on Nasdaq. Class C carries no vote. Class A (GOOGL) and Class B share the same economics.^S1S10 The SEC register lists SIC 7370.^S10

The latest periodic filing is the Q2 2026 Form 10-Q, filed 23 July 2026. No Q3 2026 10-Q is on file on 22 October 2026.^S10 A report date of 28 October 2026 is projected by a calendar service and is not confirmed by the company.^S29S26

Services earns the profit. Cloud turned from a loss in FY2022 to 14% of Services plus Cloud income. Other Bets and group AI research cost $30.5bn a year.

Sources: FY2022 from the FY2023 10-K; FY2023–25 from the FY2025 10-K; TTM from the Q2 2026 10-Q, all reported by the company.^S4S3S1 TTM = FY2025 + H1 2026 − H1 2025, calculated by us.

SBC is a cost here. Stock-based compensation was $28.1bn over the TTM.^S2S3 Segment operating income already deducts it. The model does not add it back. The share count therefore adds no further dilution for future grants.

Depreciation by segment is not disclosed. The model needs a pre-depreciation margin per segment, because capex is modelled once for the group. We allocate depreciation 50% to Services, 45% to Cloud and 5% to group level. This split is our assumption. On it, the TTM pre-depreciation margin is 45.3% for Services and 46.0% for Cloud. A 30/65 split gives a central value of $198 (appendix F).

Customers. No customer or group of affiliated customers was above 10% of revenue in 2023, 2024 or 2025.^S3 Named Cloud customers behind the backlog are not disclosed in the 10-Q or 10-K; we searched both and the Q2 release (Unverified).^S1S2S3

Capex has run at two to five times depreciation since FY2022. The gap is growth capex. It must earn a return later, and it raises depreciation for years.

Capex and depreciation are reported in the cash flow statement.^S2S3S11 The 2026 guidance of $195–205bn comes from the Q2 call, reported by two press sources; the 10-Q says only that 2026 investment will “significantly increase”.^S17S18S1

Commitments behind the build. Purchase commitments were $811.0bn on 30 June 2026, of which $200.7bn short-term. Leases not yet commenced add $85.2bn. Revenue backlog was $519.5bn, with $513.9bn in Cloud; just over half is expected within 24 months.^S1

Maintenance versus growth. The company does not split capex. We use depreciation as the maintenance proxy. On that proxy, TTM maintenance capex was $25.2bn and growth capex $107.2bn, calculated by us. The proxy understates maintenance when server lives shorten.

Normalisation in the model. 2026 capex is the guidance midpoint. 2027 capex is a scenario input ($230bn / $250bn / $270bn). From 2028 the capex-to-revenue ratio fades in a straight line to a terminal ratio by 2032 (12% / 15% / 16%). Depreciation is a rate on opening net plant plus half of the year’s capex: 8.7% now, which reproduces 2026 depreciation of about $30bn, rising to 11.5% by 2029 as construction enters service. The terminal value then charges reinvestment as growth divided by return on new capital.

In the central path, depreciation rises from $30.2bn in 2026 to $119.7bn in 2030. The EBIT margin falls from 33.1% TTM to 25.9% in 2029, then recovers to 28.5% by 2035. Free cash flow stays negative until 2029.

Purchase-commitment loss rate. The company reports no write-down history on these commitments in the filings searched (10-K FY2023–25, 10-Q Q1–Q2 2026). We run no loss stress; the bear case cuts capex intensity instead (Unverified).^S1S3S4

Cash, bonds and equity stakes are added to the operating value. Debt and the tax on unrealised gains are subtracted.

Balances are reported by the company.^S1S2 Haircuts are our assumptions: non-marketable stakes 30% / 15% / 0%, marketable equity 20% / 0% / 0%. Most non-marketable value is a mark from observed transactions: $87.9bn was remeasured in Q2, cost was $47.6bn and cumulative net upward adjustments $76.6bn.^S1 The company names SpaceX and one private company as the source of the Q2 gains; the private company is not named (Unverified).^S1

Accrued fines are a current liability, $17.356bn at 30 June 2026.^S1 The DMA fines of €890m were announced on 23 July 2026; whether they are accrued in Q3 is Unverified.^S33 The credit backstops cover data-centre payment obligations of third parties: $43.8bn of credit derivatives and $7.6bn of guarantees, up to 15 years. Alphabet may take over the leases on default.^S1

Operating leases ($18.0bn) are not subtracted: lease cost sits inside operating margins. The $25bn August notes raised cash and debt by the same amount, so the bridge does not change.^S7 Funding commitments to unconsolidated entities of $21.9bn convert cash into stakes; they are not subtracted.^S1

Share count. Common shares were 12.230bn on 15 July 2026.^S1 The 19m preferred shares convert at 2.252–2.816 (Series A) or 2.274–2.842 (Series B) common shares each; below the initial price the maximum rate applies, so we add 53.8m shares.^S1S8 Unvested RSUs (294m) are excluded because their cost is inside future operating income; including them gives $183.

Other Bets. Bear: TTM losses of about $9bn continue five years (−$3 per share). Base: zero; losses offset by option value, a judgement. Bull: $63bn, half of the Waymo round value, because Alphabet is reported as majority holder. The February 2026 round valued Waymo at $126bn including new money; a pre-close report said about $110bn. We use the closed round and do not average.^S21S22 Alphabet’s stake is not disclosed (Unverified).

All rate inputs are dated and closed from data before any judgement round.

Terminal growth is 3.0% in every case, below the risk-free rate. The cases differ in operations, not in the discount rate. A global issuer with a US premium only is a simplification; no country premium is added.^S25

Each case copies a regime Alphabet has lived through. No case pushes every factor at once.

Cloud margins in Base use the TTM level, not the FY2024–25 level. Cloud is scaling, and its margin rose each year from FY2022 to TTM. The Base return on new capital, 24%, equals the measured incremental ROIC from FY2023 to TTM (section 15). Growth fades in a straight line from 2027 to 2035. All forecast inputs are our assumptions, not company guidance.

Plausibility. Tangible book is $45 per share. The Bear value is 2.3× tangible book, above the 0.5× floor that would need proof.

Central value $187 per share. Services carries 62% of it and Cloud 39%.

Cash flows sit at mid-year and are discounted to 21 October 2026. H2 2026 is the only 2026 flow; H1 is in the 30 June balance sheet. The segment split allocates capex and depreciation by the 50/45/5 assumption; the total does not depend on it. Group AI research cost in the terminal value is spread across segments by their contribution.

Central path, calculated by us. 2026 revenue includes reported H1; the 2026 EBIT margin and FCF are for H2 only.

At central margins, capex and WACC, the $331 price requires revenue to grow 17.6% a year from FY2025 to 2035. The central path has 11.6%.

That growth would take revenue from $403bn in FY2025 to about $2,032bn in 2035. Alternatively, the price fits central operations at a discount rate of 7.1%, below the dated cost of equity of 9.9%.^S24

Implied 10-year revenue CAGR: WACC × Services pre-depreciation margin

Central value per share: WACC × terminal growth

No cell of the value matrix reaches the price. The highest cell, WACC 8.8% and growth 4.0%, gives $254. The growth rate matters more than the margin: each margin step changes the implied CAGR by under half a point.

For a multi-segment group these are screens, not tests. They are shown for comparison with earlier notes. The value comes from the model.

NOPAT uses a 17% tax rate for comparability across years. Capital excludes investment stakes, so unrealised gains do not inflate the base.^S1S5 ROIC fell from 45.3% to 32.6% because capital grew faster than profit. Assets not yet in service were $122.8bn on 30 June 2026, up from $78.6bn in December; they earn nothing yet.^S1 On capital in service, ROIC is 48.5%. The incremental return of 23.8% is still well above the 9.8% WACC. The AI build has paid so far. The model’s question is whether the next $450bn earns the same.

Why the Faustmann ratio misleads here. At 6.5× it equals price to book, because net worth reduces to common equity. Book equity holds $76.6bn of unrealised gains on private stakes and $57.8bn of goodwill, about $22.7bn of it from Wiz.^S1S6 The ratio says nothing about the cash return on the operating assets.

The earlier note gave $104 from 12 × FY2025 EPS less 20%. It was a no-growth multiple with a haircut.

The old multiple of 12× equals an earnings yield of 8.3%. With the 20% cut, it is 9.6×, close to capitalising today’s profit at our WACC. So the two methods agree on earning power. They differ on three items. First, the old rule ignored $223bn of cash and stakes net of debt. Second, it gave zero value to growth; Cloud alone has a $514bn backlog. Third, FY2025 EPS included $1.7bn of other income gains but excluded the H1 2026 profit step; core TTM EPS is $10.^S2S3

The new method also charges what the old one hid: the capex build. In the central path, FCF is negative until 2029. Growth is worth $67 only after that cost.

These items are context, not evidence for the value. They do not enter the model.

Every financial table states its period and unit. Reported figures keep issuer provenance. The TTM cash flow reconciles to the Q3 2025 10-Q and the FY2025 release.^S27S28 Two quote publications confirm the closing price; their feed independence is Unverified. Forecast inputs are judgements.

Same-issuer filings, releases and syndicated copies do not give independent confirmation. Each filing figure was read twice: from the filing text and from the XBRL register.^S11

This report contains public company research only. Figures may differ by rounding.

A. Scenario assumptions

B. Annual forecasts, $bn

C. Valuation outputs

D. Base inputs and provenance

E. Sensitivities on the central value