Document · 206 blocks

Google’s two internal directions

ALPHABET INC. · DUE DILIGENCE

Google’s two internal directions

Verticalising the ad business from CTR to the whole pipeline, and monetising information directly. Evidence for and against, decomposed probability estimates, competitive analysis, and the cash flow model.

INSIDE

Written 23 August 2026 Data through Q2 2026 — reported 22 July 2026, 10-Q filed 23 July 2026 Method Ten parallel research agents, then an adversarial verification pass that found and corrected 21 errors in the first draft Status Analysis, not investment advice. Figures described as estimates are estimates.

CONFIDENCE LABELS — these are load-bearing. DISC disclosed by Alphabet in a filing or on an earnings call · DER derived by arithmetic from disclosed figures · PRESS company blog or press release · COURT court finding · REG regulator statement · 3P third-party measurement · EST my estimate or model output · GAP searched for, no public data exists

CONTENTS

1. Summary

Two corrections come out of the research. Both matter more than the yes/no answer.

Correction 1: Google reaches the payment step and deliberately takes nothing from it

Direction 1 as stated was: discovery → clickthrough → selection → payment. The first three are happening and are documented. The fourth needs a distinction that the word "payment" hides.

Google is present at the payment step. Google Pay is the payment handler inside UCP checkout, Google added BNPL through Affirm and Klarna in May 2026, and Google launched "Buy with Google Pay" on connected TV in Q2 2026 PRESS/DISC. Google reaches the payment step. What Google does not do is take any economics from it.

The precise position, from Google's own documentation:

So Google Pay is a credential store and an authorisation interface, not a rail Google charges for. The distinction that matters for the thesis is between the payment interface, which Google has, and the payment economics, which Google does not have and has repeatedly declined to build.

Two facts show the declining is deliberate rather than incidental. Google Plex — Google-branded checking and savings accounts with Citi and ten other banks, roughly 400,000 people on the waitlist — was cancelled in October 2021 before launch, and Google's stated replacement strategy was "digital enablement for banks… rather than us serving as the provider" 3P/PRESS. And the Google Pay app and peer-to-peer payments were shut down in the US in June 2024 and folded into Google Wallet 3P. Both moves take Google out of the regulated perimeter: holding balances and moving money requires state money-transmitter licensing and BSA/AML compliance, while pass-through card tokenization requires neither. It also earns neither. India is the clearest illustration of the trade: Google Pay there is enormous by transaction volume and earns nothing, because UPI merchant discount rates are regulated to zero 3P.

Nothing in regulation prevents Google from being merchant of record. Google holds an e-money institution licence in Lithuania and could operate as a payment institution in the EU 3P. The constraint is cost, liability and channel conflict, and Google's revealed preference across three separate retreats is to stay outside the perimeter. A facilitator earns basis points. A merchant of record earns margin. Google has chosen facilitator every time.

The record on transaction fees follows the same pattern. Google has reached the point of charging for a transaction repeatedly and has withdrawn from it every time, most recently in February 2025.

The record since 2020 is nine withdrawals from transaction fees and zero additions:

The 2025 hotel decision is the clearest evidence. Google had built a real take rate — commission on completed stays, billed 45 days after checkout, not billed on cancellations — and held an estimated 64–80% of hotel metasearch traffic with it 3P. Google closed the product and returned to charging per click. What survives is CPC%: a percentage of room price, charged per click. Google indexes its price to transaction value. Google does not index its revenue to transaction occurrence.

The current agentic commerce stack repeats the same design. Under the Universal Commerce Protocol the merchant remains seller of record, and Google Pay passes an encrypted payload to the merchant's own payment service provider rather than settling funds [PRESS, primary: Google Pay developer documentation]. No fee schedule has been published for UCP, Universal Cart, AP2 or agentic checkout, anywhere GAP. AP2 was donated to the FIDO Alliance on 28 April 2026 PRESS.

The better-supported version of direction 1 is narrower: Google is verticalising the measurement and allocation layer, not the transaction layer. Google is taking the purchase data, and Google is taking the decision about where the advertiser's budget goes. Google is not taking the money.

That version is shipping, with an enforcement date. From 15 June 2026 the Google Ads API stopped accepting new adopters of offline conversion imports and enhanced conversions for leads; new callers receive CUSTOMER_NOT_ALLOWLISTED_FOR_THIS_FEATURE, and migration to the Google Data Manager API is mandatory [PRESS, primary: Google Ads Developer Blog, 15 May 2026]. That one API now consolidates offline conversions, enhanced conversions, Customer Match, store sales and events. Google required every advertiser's purchase data — CRM records, in-store sales, offline conversions — to move through one pipeline Google owns.

Correction 2: the information-monetisation direction is working in enterprise, not in consumer

Direction 2 was framed around consumer subscriptions — Google AI Pro, YouTube Premium. That business exists and is growing, but the amounts are small and the pricing power in AI specifically is negative.

The business that is monetising organised information at scale is Google Cloud: $77,617m of TTM revenue, a 35.6% operating margin in Q2 2026 against 9.4% in Q1 2024, and a $513,900m Cloud revenue backlog with "just over 50%" expected to be recognised within 24 months DISC.

Cloud also produced +$28.4bn of the +$74.5bn increase in Alphabet's TTM revenue — 38% of all the growth DER. Consumer AI subscriptions at 5% conversion of 1bn Gemini monthly users at $15/month net produce $9bn a year EST. Cloud produced $24,768m in one quarter.

The answer in one paragraph

Direction 1 is right about the mechanism and wrong about where the money is collected. Google is likely to own the discovery, selection and measurement layers of commerce; Google is unlikely to charge a transaction fee, because it has declined to do so nine times in six years and because a fee substitutes for an ad dollar that already carries no cost of goods. Direction 2 is right that Google is monetising information directly, but the growth is in enterprise cloud and in YouTube Premium price increases, not in consumer AI subscriptions, where Google is cutting prices into a market in which 3% of people pay. Neither direction, executed fully, is large enough to determine Alphabet's cash flow. The dominant cash flow variables are capex rising from $195–205bn in 2026 and a depreciation charge that roughly triples to quadruples by 2029.

2. Direction 1 — verticalising from CTR to the whole pipeline

2.1 What Google has built, stage by stage

Look at the last column. Google charges at one point only — the click — and it charges there in three places. Everything else in the funnel Google built and gave away, including the checkout and the payment handler.

Note the payment row specifically. Google Pay is genuinely present and genuinely used. It is the authorisation interface and the credential store. It is not a settlement rail Google charges for, and Alphabet has never disclosed a dollar of revenue from it GAP.

2.2 The scale of what is live is small, and Alphabet said so

On the Q2 2026 call Philipp Schindler's own evidence for UCP adoption was: "Merchants are rapidly adopting UCP with Target and Steve Madden now live, while new members have joined." DISC

Two named live merchants, seven months after the January 2026 launch. That is the company's best available example.

Nothing else is disclosed. No AP2 transaction count, no UCP order count, no Universal Cart GMV, no Google Pay volume, no AI Mode monetisation rate GAP on all six.

2.3 The base rate: 18 priced attempts, 2 durable

Across 2002–2026 the research identified 36 distinct Google moves down the funnel [EST, compiled from company sources, press archives and killedbygoogle]. Restricting to the ones where Google inserted itself into a third party's transaction and priced it, there are 18:

Base rate of durable success at intermediating another party's transaction: 2 of 18, about 11% EST. If you require a disclosed, material take rate, the count is closer to zero, because Local Services Ads revenue has never been broken out GAP.

The four most instructive cases:

Google Checkout (2006–2013). Google had the demand side, the merchant relationships, and priced below PayPal — free for 19 months, then 2.0% + $0.20, plus free processing up to 10x a merchant's AdWords spend. Merchants adopted it as a discount on advertising. When Google removed the AdWords subsidy in May 2009 and matched PayPal's tiers, adoption stopped. Adoption tracked the subsidy, not the product 3P.

Buy on Google (2018–2023). Google charged a commission, cut it to zero in July 2020, and closed the checkout anyway in September 2023. Zero price did not produce adoption — about 4,500–8,000 active retailers at peak 3P. Google's own description: "a small feature that a very limited number of merchants used." The replacement, sending the shopper to the merchant's own checkout, tested 1–10% better on GMV than Google owning the transaction PRESS. Google converted worse than the merchants did.

Google Compare (2015–2016). The one attempt where the economics were clearly good: insurance and mortgage referral fees are worth far more per lead than generic search ads. Google built it twice and bought BeatThatQuote for £37.7m. It closed 13 months after the US launch. The structural reason is that Google was competing with its own highest-spending advertisers in its highest-CPC verticals. Taking a referral fee meant reducing the auction.

Hotel Ads commission (2020–2025). Described above. This is the strongest single piece of evidence against the payment leg of direction 1, because Google had the take rate, had the market share, and gave it up voluntarily.

2.4 The rule the record supports

Google succeeds at moving down the funnel when it changes the pricing unit of its advertising business. Google fails when it tries to become the counterparty.

Every durable success fits this rule. Play billing works because Google owns the operating system and the store, so there is no incumbent to displace. YouTube works because Google owns the inventory. Local Services Ads works because the unit is the lead, not the transaction: Google never touches the money, never becomes counterparty, and never assumes fulfilment risk. $25–80 per lead 3P is an advertising price with better attribution, not a take rate.

Performance Max fits the same rule. It takes allocation control, not money. It changes what Google decides, not what Google charges.

2.5 What is different this time

Three things are different from 2018–2023. They are the reason the estimate is not simply the historical base rate.

1. Google owns the standard its competitors will transact through. On 24 April 2026, Amazon, Meta, Microsoft, Salesforce and Stripe joined the UCP Tech Council, alongside founders Google, Shopify, Etsy, Target and Wayfair [PRESS, verified directly by fetching the announcement]. OpenAI is not a member. The council "aligns UCP's technical direction, reviewing contribution proposals and stewarding the open-source protocol."

This converts a disintermediation risk into optionality. If agentic commerce becomes large, the transaction runs through Google's protocol whether or not Gemini is the assistant that started it.

2. The competition tried the take-rate version and it failed publicly. OpenAI launched Instant Checkout with Stripe in September 2025 and shut it down on 24 March 2026, stating it "did not offer the level of flexibility that we aspire to provide" DISC, OpenAI. Fewer than 30 merchants ever went live. Walmart measured in-chat conversion at one-third the rate of click-out to Walmart.com [3P, Walmart EVP Daniel Danker]. Etsy saw no large volume. Instacart said agents drove "a very small percentage" of orders. OpenAI's 4% Shopify merchant fee took effect 26 January 2026, two months before the product was deprecated 3P.

The merchant objection is about data, not fees. A specialty retail executive listed what in-chat checkout lacked: "real-time inventory tracking, coupons, promotions, customer data collection, loyalty program integration and store pick-up" 3P. Circana's Marshal Cohen: "Why would I give someone else control of my customer base?"

3. The measurement layer is being verticalised by requirement, not by persuasion. The Data Manager migration is the first time Google has set a hard cutover date on advertiser purchase data. This part of direction 1 is happening.

2.6 Competitive position

Where Google is ahead:

Where Google is behind:

The market is small and decelerating. Juniper Research puts AI agents' influence at about $8bn of US retail spend in 2026 against roughly $1.3trn of US e-commerce — under 1% 3P. Adobe's AI-referred retail traffic growth fell from +393% year on year in Q1 2026 to +62% in July 2026 3P. Only 10% of AI-assisted shoppers complete the purchase on the AI platform; 69% go to the retailer's own site 3P, Publicis/eMarketer.

2.7 Probability estimate, decomposed

A single number for "will direction 1 succeed" is not useful, because the direction bundles five separate claims with different odds.

Combining these matters, and the report should not hide the arithmetic.

Direction 1 as originally stated — pipeline through to payment, monetised — requires 1a and 1b and 1c and 1e. On independence that is 0.75 × 0.65 × 0.30 × 0.15 = 2.2%. The legs are positively correlated: if Google owns discovery and selection, hosting checkout and charging for it both become more likely. Under near-total correlation the answer is capped by the weakest leg, 15%. The honest range is 2–15%, central estimate 5–8% EST.

Direction 1 as Google is executing it — pipeline ingested informationally, monetised through advertising — requires 1a and 1b and 1d. On independence that is 0.75 × 0.65 × 0.85 = 41%. Under near-total correlation it approaches 65%. The honest range is 41–65% EST.

The difference between those two ranges is the main conclusion of this section.

3. Direction 2 — monetising information

3.1 The complete subscription inventory

Growing (ranked by estimated revenue contribution):

Caveat on the largest line. From 30 June 2026 Play subscriptions fall from 15% to 10% plus a separable 5% billing fee, and one-time purchases fall from 30% to 20–25% plus 5%, with external web links carrying no billing fee DISC. The largest item in the growth table is growing in dollars while its rate compresses. See section 4.1.

Stalled:

Killed, sold or absorbed — roughly 25–30 products, of which about 20 were paid subscriptions EST:

Stadia Pro (2019–2023), Pixel Pass (2021–2023, 22 months), Google Play Music (2011–2020), YouTube Originals (2016–2022), YouTube Red, Google Fiber TV (2012–2020), Nest Secure monitoring (2017–2020), Google Photos Print (5 months, 2020), Google One VPN (2020–2024, "people simply weren't using it"), Dark Web Report (2023–January 2026), Google Domains (sold 2023), Google Optimize 360 (2023), Jamboard (2024), Google Currents (2023), Cloud IoT Core (2023), Postini, Google Search Appliance, Google Site Search, Google Play Newsstand, Google One Pass, Tables.

3.2 The pattern in what Google kills

There is a rule, and it is not a user threshold.

Google kills a subscription when it has to buy the thing it is selling, or when the subscription depends on dedicated hardware.

Content licensing killed Play Music, YouTube Originals, Stadia, Fiber TV and Newsstand. Hardware dependency killed Stadia, Pixel Pass, Jamboard and Nest Secure.

Google keeps subscriptions that run on a surface it already owns and does not pay for: Search, Android, Chrome, Gmail, Drive, YouTube's creator supply.

This has a direct implication for the AI subscription. AI inference is a cost Google owns — its own TPUs, its own data centres, no Nvidia margin and no cloud margin in between. SemiAnalysis estimates Google's all-in cost per Ironwood TPU chip is about 44% below a GB200 server internally, and about 52% lower per effective PFLOP than a GB300 NVL72 [3P/EST — analyst model, not disclosure]. By the kill rule, AI subscriptions are structurally the keepable kind. That is a point in favour of direction 2.

3.3 Google is getting better at running subscriptions

Kills of paid subscriptions cluster in 2020–2024. Since 2025 the same products are absorbed rather than terminated: Nest Aware became Google Home Premium (October 2025), Fitbit Premium became Google Health Premium (May 2026), and both were then bundled into AI Pro at $19.99 alongside YouTube Premium Lite PRESS.

Google now has one consumer subscription base product — Google One and the AI plans — with enough perceived value to absorb orphan services instead of shutting them. That is a real organisational improvement.

3.4 The pricing power split is the main finding on direction 2

Pricing power confirmed:

All prices DISC/PRESS from Google's own pricing pages.

The YouTube increase was applied to a base of 125m+ subscribers six months ago with no reported churn event PRESS.

Pricing power absent:

All prices DISC/PRESS from Google's own pricing pages.

The distribution is being bought, not earned. AI Pro free for 18 months to Reliance Jio customers in India (announced 30–31 October 2025; Jio has about 505m subscribers; redemptions never disclosed GAP). AI Pro free for 12 months to US students and AI Plus free in 140+ markets (19 August 2026, second consecutive year, payment method required and auto-charges at the end). Free years with Pixel and Chromebook Plus.

One indicator against: Google cut the Pixel AI Pro giveaway from 12 months to 6 months with the Pixel 11 Pro in August 2026 3P. That is either a margin problem or an indication that the giveaway was not converting.

OpenAI made ChatGPT Go free for a year to all users in India on 27 October 2025, four days before the Google–Jio announcement. Both companies are buying the same market.

3.5 The 350m subscription number is not a revenue metric

Alphabet has never defined "paid subscription." The 350m figure (29 April 2026, not restated in Q2 2026) mixes $1.99 storage, $8.99 Premium Lite, $82.99 YouTube TV, and 18-month-free Jio users. It has no ARPU attached.

Google tracks the distinction internally — its Q1 2026 disclosure noted YouTube Music and Premium had their largest quarterly increase in "non-trial" subscribers DISC — and does not publish it for the group total.

The Jio cohort signed in October and November 2025 on 18-month terms does not begin converting until roughly April–May 2027 EST. A material share of net additions between Q4 2025 and Q2 2026 is plausibly free-period.

3.6 The intent-transformation claim: mixed, and the two best studies conflict

All figures are year-on-year change DISC, 10-Qs. Q4 rows are DER because Alphabet never publishes Q4 separately; they are full-year less nine months and are rounding-sensitive by ±2–3pp.

The inflection is Q4 2025, not Q1 2026. That coincides with the global AI Mode expansion in October 2025. Search and other revenue grew 16.7% in Q4 2025 and 16.8% in Q2 2026 DER from filed totals.

One definitional caveat that must travel with this series: the 10-Q defines paid clicks as clicks on advertisements on Google search properties "and other Google owned and operated properties including Gmail, Google Maps, and Google Play" DISC. It is not a clean search-query proxy.

Supporting the intent-transformation claim: Tinuiti's Q2 2026 benchmark decomposes Google's blended +13% clicks and +1% CPC into text ads (clicks +8%, CPC +3%) and Shopping (clicks +19%, CPC −1%) 3P. Falling blended CPC is substantially a mix shift into Shopping, which sits closer to the transaction. Brand keyword CPC fell 1–9% while non-brand CPC rose 4%, meaning advertisers are paying up for demand capture rather than defensive brand terms. Semrush found the intent mix of AI-Overview-covered queries went from 91% informational in January 2025 to 57% informational, 18% commercial, 14% transactional in October 2025 3P. SE Ranking found ad coverage in AI Mode rises with keyword value: 24.3% on sub-$2 CPC terms and 53.6% on $10+ terms 3P.

Against it: Ahrefs, on 146m SERPs, found 99.9% of AI-Overview-triggering keywords are informational in intent, with AI Overviews appearing on only 4.3% of commercial and 2.1% of transactional queries 3P. Semrush's later study (600k keywords, November 2025 to April 2026) found commercial-intent AIO coverage up 71% but transactional-intent AIO appearance down 5% — the cleanest single data point against the claim. Pichai disclosed that AI Mode queries are 3x longer than traditional searches, and long conversational queries are the signature of informational volume. Ads in AI Mode sit below the answer, cover 29.45% of commercial queries, and only 11.53% of advertiser domains appear as cited sources 3P.

These studies measure different things — Semrush measures the composition of AIO-covered queries, Ahrefs measures the conditional trigger rate by intent — so both can be correct. The synthesis they jointly support: AI is expanding Google into the commercial-research middle of the funnel, not into the transactional bottom.

One softening in the language worth noting. Schindler moved from "we see monetization at approximately the same rate" (Q3 2025, and note this was scoped to AI Overviews, not AI Mode) to "we continue to be encouraged with monetization performance on queries that show AI Overviews" (Q2 2026) DISC. Seven quarters of qualitative language and never a number GAP.

Schindler did give one forward signal worth tracking, in Q1 2026, on the roughly 20% ad coverage figure: "with the ability of AI to better understand intent and a lot of other vectors around it, I think there is upside in that coverage number" DISC. No magnitude was given.

3.7 Competitive position

Google is third in enterprise API, third in consumer paid, and not at the frontier. Its advantages are distribution (1bn Gemini MAU, about 3bn Workspace users, 91.4% search share) and cost structure.

Reach is bundled rather than chosen. About 100m of Gemini's 1bn users are on iOS EST; the remainder are predominantly Android, where Gemini is pre-installed, though this does not mean every non-iOS user is bundled. YouGov brand preference (February–July 2026): ChatGPT 33.7%, Gemini 18.2%. Gen Z prefers ChatGPT over Gemini 3.1 to 1 (44.4% against 14.3%); Gen X is near parity 3P. Google's preference share is strongest in the older cohorts.

Google reports monthly actives while OpenAI reports weekly. A product with 1bn WAU has materially more than 1bn MAU. Assume the weekly number is unflattering until Google publishes it.

3.8 Probability estimate, decomposed

Direction 2 as originally stated — consumer subscriptions become a major direct-monetisation line — is essentially sub-claim 2b. 20–30%, central 22% EST.

Direction 2 restated to include enterprise — Google monetises organised intelligence directly at scale — is already true today and is about 65% likely to keep compounding, with a large caveat: Q2 2026 was the first quarter containing outright TPU hardware sales, at an undisclosed split, with the "significant majority" of that revenue landing in 2027 DISC. Cloud's +81.8% growth and 35.6% margin are not representative run rates.

4. The constraints that bind

4.1 Regulation attacks both directions, and the UK is ahead of the EU on the AI-specific question

On direction 1:

On direction 2:

The Play Store precedent is the closest comparison. When a regulator forces a Google fee open, the fee does not go to zero; it resets to roughly a third to a half of the prior rate. That is the base case for any forced-open commerce take rate, and it points downward.

4.2 Organisational risk

No commerce initiative at Google has survived a sponsor departure. Bill Ready, who ran Commerce through the free-listings and zero-commission decisions, left for Pinterest in 2022. Commerce now sits under Vidhya Srinivasan in the ads organisation; no appointment date is publicly documented GAP.

One further item I could not independently verify: two research agents reported that on 5 August 2026 Jeff Dean left Google after 27 years and Demis Hassabis stepped aside as DeepMind CEO, with the stock down about 4%. The agents cite CNBC and Reuters URLs. This session's web search budget was exhausted and both URLs returned 403 to direct fetch. Treat as unverified [3P, single research pass, not re-confirmed]. If accurate it is material to the model-capability leg of both directions.

5. How this converts into cash flow

5.1 The base

TTM to Q2 2026:

Prior full years, for the comparison used later DISC: FY2023 FCF $69,495m; FY2024 $72,764m; FY2025 $73,266m (operating cash flow $164,713m less capex $91,447m).

Balance sheet at 30 June 2026 DISC: cash and marketable securities $242,474m; total debt face value $101,085m against $49,085m at 31 December 2025; net cash including leases +$121,683m.

Two off-balance-sheet items moved sharply: purchase commitments went from $232,700m at Q1 2026 to $707,000m at Q2 2026 (Note 10), and leases signed but not commenced went from $58,500m to $85,200m DISC. Confirm the scope of the purchase-commitment figure before treating it as a fourfold increase — the wording changed from "fixed or minimum guaranteed" to "fixed or guaranteed" and the energy-agreement horizon extended from 2047 to 2054.

Correction to a figure in the existing project file: the $40bn ATM is authorised but entirely undrawn — "As of June 30, 2026, we have not sold any shares under the ATM Program" DISC. The H1 2026 equity raise was $20.5bn public offering plus $10.0bn Berkshire private placement plus $19.0bn of 6.25% mandatory convertible preferred, totalling $49.5bn. The preferred was not in the prior baseline. It auto-converts around 15 May 2029 into 42.8m–53.6m additional shares DER and costs about $1,190m a year in dividends.

5.2 The dominant variable is depreciation

Capex has run $32.3bn (2023) → $52.5bn (2024) → $91.4bn (2025) → $195–205bn guided for 2026 [3P, transcript-sourced; the 10-Q says only "significantly increase" with no range]. Depreciation of property and equipment has run $11.9bn → $15.3bn → $21.1bn → $25.2bn TTM DISC.

At 30 June 2026 gross PP&E was $427,124m, of which $122,814m is not yet in service — 28.8% of the asset base is not yet depreciating DISC/DER. In-service PP&E of $304,310m against $25,237m of TTM depreciation implies a blended in-service life of 12.1 years DER.

The forward path depends on how much of the new capex is short-life compute. Two cases:

Both assume a one-year lag from spend to in-service, consistent with the 28.8% not-yet-in-service disclosure EST.

All projected figures EST. Actual FY2025 depreciation of property and equipment was $21.1bn, which the slow case reproduces and the fast case overshoots by $2.1bn. Both cases depreciate faster than the 12.1-year blended in-service life implied by the disclosures; the fast case materially so. The slow case is the better central estimate.

Depreciation roughly triples to quadruples between now and 2029. That is a $63–83bn annual operating-income effect against the TTM charge of $25.2bn. It is non-cash, so it does not reduce free cash flow directly, but it does reduce reported operating margin, and it is why reported margins can compress while the business grows.

Separately, if Alphabet shortened the server life from six years to five, the cost is $4–7bn pre-tax in year one EST, rising past $10bn as the 2026 and 2027 vintages enter service. At an 18% tax rate on 12,151m shares that is $0.27–$0.47 of EPS in year one, about $0.67 at +$10bn DER. No cash effect. The policy language is unchanged in both 2026 10-Qs, and the annual report flags that "expected technology advancements… could change the period over which we expect to benefit from the asset" DISC.

5.3 Scenarios

All scenario figures are EST. Base year is FY2025 actual revenue of $402,837m, not the TTM figure. Method: apply a revenue growth rate and an EBITDA margin, subtract the modelled depreciation above (slow case unless stated), then convert to cash flow as operating income × 0.82 (an 18% cash tax rate) plus D&A plus stock-based compensation at 6.31% of revenue plus a calibrated residual of $9,936m for interest income, working capital and deferred taxes.

That formula reproduces Alphabet's actual TTM operating cash flow exactly by construction. Without the residual it under-predicts TTM by $9,936m, which is why the residual is there. Two known biases: the residual is held flat to 2029 even in the scenarios where free cash flow is negative for three years and the cash balance, and therefore the interest income inside that residual, would fall — this is mildly optimistic in the base and bear cases. And 2026 is largely fixed by H1 actuals: H1 2026 revenue was already +23.1% year on year DER, so the 2026 spread across scenarios is only 1.7% and the scenarios genuinely separate from 2027 onward.

Bull — both directions land. Search holds mid-teens growth, AI subscriptions reach about $15bn, Cloud compounds. ($bn)

Base — the advertising-side verticalisation works; the payment leg and consumer subscriptions do not. ($bn)

Bear — neither direction works and AI inventory dilutes CPC. ($bn)

Columns may not foot exactly because every figure is rounded to the nearest $1bn. On fast-case depreciation, bear 2029 operating margin falls to 14.0% and FCF to −$38bn; bull 2029 operating margin falls to 25.5% and FCF rises to $58bn EST.

Two assumptions to challenge. First, the bull case assumes EBITDA margin expands from 39.5% to 40.2% while Cloud — at a 35.6% operating margin, below the group — is the fastest-growing segment. Mix alone pushes the blended margin down, so the bull margin path is optimistic. Second, in the bull case free cash flow does not return to the FY2025 level of $73.3bn until after 2029.

5.4 The hurdle rate on one capex vintage

Take the 2026 vintage alone at $200bn, six-year life, 18% effective tax, 9% cost of capital. The capital recovery factor is 0.2229, so the vintage must produce DER:

At 8% cost of capital it is $76bn; at 10% it is $81bn. For the 2027 vintage at $230bn it is $90bn.

Alphabet's total revenue increase over the last twelve months was $74.5bn DER. One year's capex vintage requires roughly one year's worth of total-company revenue growth, permanently, to earn its cost of capital.

Now size the two directions against that. All figures EST, on a corrected 2026E Search and other base of $264,728m (H1 2026 actual of $123,670m grossed up by FY2025's H2/H1 Search ratio of 1.1406; cross-checked as FY2025 Search of $224,532m times the H1 2026 growth rate of 1.179 — same answer).

Probabilities in brackets refer to the numbered sub-claims in sections 2.7 and 3.8. The 60% on the coverage row is not one of them — it is a standalone EST of the probability that ad coverage reaches about 24% at all, based only on Schindler's statement that there is "upside in that coverage number," and it is the largest single contributor to the expected value.

Three caveats on this table, all material:

Cloud is deliberately excluded from that table, because it is not one of the two directions under test. That exclusion makes the capex look worse than the disclosures support. Cloud grew +$28.4bn TTM, 38% of Alphabet's entire revenue increase, at +81.8% in Q2 2026, and it is the segment the data-centre capex directly serves DER/DISC.

5.5 What this means

These two directions are not the main determinant of Alphabet's cash flow. They determine whether Search margins and growth hold.

They matter because they determine whether Search revenue growth stays above about 13%, and Search revenue growth at that level is what funds the capex. They do not matter as a new revenue line, because neither one is large enough. Fully delivered, they add about $75bn of incremental annual revenue — approximately one 2026 capex vintage's $78bn hurdle. Probability-weighted using this report's own estimates, they add about $28bn, roughly 36% of one vintage's hurdle.

The actual cash flow drivers, in order:

One accounting note that applies to every valuation figure quoted anywhere. At 30 June 2026, initial cost of $47,642m plus $85,732m of cumulative upward adjustments less $9,115m of cumulative downward adjustments and impairments gave a carrying value of $124,259m on non-marketable equity securities held under the ASC 321 measurement alternative DISC. Q2 2026 alone booked $99,031m of net gains on equity securities: $77,354m unrealised on measurement-alternative holdings, $21,399m unrealised on marketable and other equity securities, and $278m realised on securities sold, "primarily related to unrealized gains… from SpaceX and a private company" DISC. Alphabet does not disclose the SpaceX stake size or cost basis anywhere GAP. Consensus 2026 EPS of $20.58 is arithmetically a GAAP blend embedding $135,946m of H1 marks. Do not compute a price-to-earnings ratio off it. The comparable operating figure is roughly $10.91 [EST, from the prior project record].

There is a reconciliation problem in these figures that should be checked against the 10-Q directly: cumulative upward adjustments rose $41,247m in H1 2026 ($44,485m to $85,732m), against $135,946m of total H1 gains on equity securities, of which the Q2 measurement-alternative portion alone was $77,354m. Those numbers do not reconcile without dispositions, reclassifications, or a large marketable-securities component that is not visible in the summary. See section 7 item 9.

6. The signal dashboard

Ranked by how much a change would move the answer.

Tier 1 — direction-deciding

Tier 2 — direction-confirming

Tier 3 — hard dates

7. Open items — verify before relying on these

8. Bottom line

On direction 1. The mechanism is right; the point of collection is wrong. Google is ingesting the pipeline — discovery, selection, and above all the purchase data through the mandatory Data Manager migration — and Google now owns the protocol its competitors will transact through. Google is unlikely to charge for the transaction. Google Pay is present at the payment step — handler, BNPL, connected-TV checkout — but Google is not merchant of record, does not settle funds, charges merchants nothing, and has never disclosed a dollar of payment revenue. Against a fee: nine withdrawals in six years, two retreats out of the regulated payments perimeter (Plex 2021, Google Pay US 2024), a protocol donated to a standards body, and an advertising dollar that already carries no cost of goods and would be substituted roughly one-for-one by any fee.

Pipeline through payment, monetised: 3–15%, central 5–8%. Pipeline ingested and monetised through advertising: 41–65%.

Three things constrain even the higher figure, and they are not in the numbers above. Amazon holds the stronger position at the point of purchase — 350m+ customers on Alexa for Shopping, +40% spend per order, $19.8bn of quarterly ad revenue growing 26%, 79.7% of US retail media, against Google's approximately 0% measured share after twenty years, and retailer-native assistants taking 54.1% of AI-driven retail e-commerce through 2030. DMA Article 5(2) breaks the data join in the EEA, which attacks the 85% leg directly. And Google is not at the model frontier — 56 against 63 and 61, with the Pro tier stuck on a February "Preview" for six months — which matters because model quality is what makes an agent reliable enough to be trusted with a payment instrument.

The regulatory constraints are not in either range either, and two of them carry superlatives elsewhere in this report. Judge Brinkema's ad tech remedies ruling has been pending since 21 November 2025 and could issue any day; it is the largest unpriced binary in the file. The EC's €890m DMA decision does not mention AI Mode or AI Overviews, so whether the self-preferencing prohibition reaches an AI answer surface is untested in the EU — the largest open regulatory question for direction 1. DMA Article 6(11) forces Google to license query and click data to "AI chatbots offering search functionalities" from January 2027, which attacks sub-claim 1a. DMA Article 6(7) opens 11 Android hooks, including shopping-list automation, to rival assistants by 1 August 2027, which attacks 1b and 1c. The CMA's fair-ranking requirement comes into force around December 2026 and expressly binds generative AI features — the nearest-dated constraint of any of them. And the EC's Article 102 investigation into publisher content in AI Overviews, if sustained, means the input cost of AI answers stops being zero.

On direction 2. Right about the direction, wrong about which business is delivering it. Consumer AI subscriptions face a market where 3% of people pay, Google is cutting prices 20% and 37.5% while rivals hold at $200, and Google's consumer app-store spend is 27x behind ChatGPT's. The subscription pricing power Google has demonstrated is in YouTube Premium (+14.3% and +17.4%, no churn event) and Workspace (+16.7% with the AI add-on removed), not in AI.

One further item the subscriber-based case has to absorb: the Reliance Jio cohort signed on 18-month free terms in October and November 2025 does not begin converting until roughly April–May 2027, so a material share of the net additions between Q4 2025 and Q2 2026 is plausibly free-period, and Alphabet does not publish the trial split for the group total.

Consumer subscription version: 20–30%, central 22%. Including enterprise: already true, about 65% to keep compounding — with the caveat that Cloud's 35.6% margin and +81.8% growth are contaminated by undisclosed TPU hardware sales whose "significant majority" lands in 2027.

On cash flow. Neither direction is the main determinant. Fully delivered they add about $75bn of incremental annual revenue, which is approximately the $78bn hurdle on one 2026 capex vintage — and that hurdle recurs every year, at $90bn for the 2027 vintage. Probability-weighted using this report's own estimates, they add about $28bn, roughly 36% of one vintage's hurdle. In the bull scenario free cash flow does not return to the FY2025 level of $73.3bn until after 2029. In the bear scenario it is −$42bn in 2029. The outcome is decided by 2027 capex, whether Search growth holds above about 13%, and a depreciation charge that roughly triples to quadruples by 2029.

And a warning that applies to every reported profit figure in the meantime. Q2 2026 booked $99,031m of net gains on equity securities, almost all unrealised marks on private holdings under the ASC 321 measurement alternative. Consensus 2026 EPS of $20.58 embeds $135,946m of H1 marks. Do not compute a price-to-earnings ratio off it — the comparable operating figure is roughly $10.91 — and note that the disclosed movement in cumulative upward adjustments does not currently reconcile to the disclosed gains (section 7 item 9).

On the share count, which the two directions do not address at all. Zero buybacks in H1 2026, $49.5bn of equity raised, 42.8m–53.6m more shares arriving from the mandatory convertible in 2029, $1.19bn a year of preferred dividends, and a basic share count rising since Q3 2025. Even if both directions work, per-share outcomes are worse than company-level outcomes for the first time in years.

The single most useful thing to watch is whether Google ever publishes a fee schedule for UCP. That one document, or its continued absence, resolves most of direction 1.