Aya Research
DEEP RESEARCH
中EN

Alphabet: It Invented the Transformer, Lost All Eight Authors, Got Cornered by Its Own Creation — and Staged the Fastest Comeback in Big Tech History

Panoramic research report · longitudinal history + cross-sectional rivalry + synthesis Subject: Alphabet Inc. (NASDAQ: GOOGL) Report date: 2026-07-05 · Data cutoff: Q1 2026 results (reported 2026-04-29) + news flow through 2026-07-05 (Q2 2026 reports late July) Sources: company IR releases and SEC filings, CNBC, Bloomberg, TechCrunch, Synergy Research, SemiAnalysis, court records (DOJ v. Google), Pew, Similarweb, Yahoo Finance For information and research purposes only. Not investment advice.

Before we start

In 1999, Larry Page and Sergey Brin tried to sell Google to the Excite portal for $1 million. Excite's investor negotiated them down to $750,000 — and the CEO still said no, because Page insisted Excite replace all of its search technology with Google's. Today Alphabet is worth $4.39 trillion, roughly 5.8 million times the price at which its founders were once willing to walk away.

That anecdote usually gets told as comedy. This report tells it as method, because the same pattern — the outside world systematically underpricing what Google built, including Google itself underpricing it — recurs at every hinge of the story. The company acquired Android for $50 million, YouTube for $1.65 billion and DoubleClick for $3.1 billion while being ridiculed for overpaying each time. It invented the Transformer architecture in 2017 and let all eight authors of the paper walk out the door to found and staff its future competitors. It built the TPU in 2015 and treated it for a decade as internal plumbing. Then, in late 2022, a chatbot built on Google's own invention triggered a "code red" inside the company, a botched demo erased $100 billion of market value in a day, and the world spent eighteen months writing Google's obituary.

The obituary was mispriced too. Between the Bard fiasco and this summer, Alphabet shipped three generations of Gemini (the third of which broke every public benchmark record), grew its cloud business 63% with a $460 billion backlog, watched a court decline to break it up, crossed $3 trillion and then $4 trillion of market value, returned the best annual stock performance it has recorded since 2009 (+65% in 2025), signed Apple as a $1 billion-a-year customer for Siri, and began selling the TPU — the decade-old internal plumbing — to Anthropic and possibly Meta in deals measured in the tens of billions. The stock that was the AI trade's designated victim in 2023 enters July 2026 as the Mag 7's third-largest company, four turns cheaper than Apple, with the industry's only vertically integrated AI stack: chips, cloud, models, distribution, and the world's default question box.

This report asks the question the round trip leaves behind: what, exactly, is the market still charging Alphabet for — and what is it now, perhaps prematurely, giving it credit for? Part One is the longitudinal history, which doubles as a study of the most expensive organizational-inertia case ever documented and its cure. Part Two maps the six-front war (search, models, cloud, silicon, video, courtrooms). Part Three synthesizes: the snapshot, the two numbers that disagree ($460 billion of backlog versus 68% zero-click searches), scenarios, and the falsifiable claims.

For readers of the series: the SK hynix report tracked Samsung supplying 60% of the HBM in Google's TPUs; the Meta report used Alphabet as the control group — the hyperscaler whose capex the market rewards because receipts attach. This report is the control group examined in its own right, and the fourth side of the AI value chain (chip → memory → landlord → tenant) viewed from the one company that occupies three of the four sides simultaneously.

Part One · Longitudinal: the underpriced machine (1996–2026)

1. $750,000, refused (1996–2004)

Google began in 1996 as BackRub, a Stanford research project ranking pages by their inbound links — PageRank, the insight that the web's structure was itself data. Sun co-founder Andy Bechtolsheim wrote the famous $100,000 check in August 1998 before the company legally existed; Sequoia and Kleiner Perkins, bitter rivals, split a $25 million round the following June — itself a signal of how obviously valuable the technology was to everyone except the incumbents who could have owned it. Excite passed at $750,000 — and the stated reason is the detail worth keeping: Page insisted Excite replace all of its search technology with Google's, and the CEO balked at the organizational surgery. The technology was affordable; the self-displacement wasn't. Every incumbent in this report's later chapters — including, in 2023, Google itself — fails or nearly fails on exactly that clause. Yahoo not only passed but, in 2000, hired Google as its search provider, personally financing its brand and traffic — the incumbent funding its own replacement, an error Part Two will show Google spent two decades terrified of repeating, and then nearly repeated.

AdWords launched in 2000 with 350 advertisers; the 2002 shift to auction-priced cost-per-click created the greatest business model in the history of commerce — advertising sold at the exact moment of declared intent, priced by the advertisers' own competition. Pause on why it is the greatest, because the entire 2026 search debate assumes the reader knows: a search ad is the only advertising format in which the consumer announces what they want, the instant they want it, and rival sellers bid in a real-time auction for the introduction. Every other ad model — feeds, TV, display — pays to interrupt attention and infer intent probabilistically; search harvests declared intent at zero inference cost. This is why search monetizes at multiples of feed advertising per unit of attention, why Meta needed a $10 billion AI rebuild to approximate what Google gets typed into a box for free, and why the question "will people stop typing into the box" is worth four trillion dollars.

The 2004 IPO (a populist Dutch auction, priced at $85 against a hoped-for $108–135, mocked as a flop — and, in retrospect, the last time Google's equity was ever cheap) valued the company at $23 billion. Every subsequent valuation argument about Alphabet, including today's, is a version of the same dispute the auction embodied: how much is a money machine worth if you cannot see its moving parts?

2. Three "overpriced" acquisitions (2005–2007)

Three purchases in thirty months built the empire's outer walls, each condemned as folly in real time. Android, 2005, ~$50 million: a dying startup handing out free phone software; it became the operating system of five of every six smartphones on earth and the distribution moat for everything else Google makes — the company's own M&A chief later called it the "best deal ever." YouTube, 2006, $1.65 billion: Mark Cuban said only a "moron" would buy it, that copyright suits would bury it; it is now the largest television distributor in the United States by watch share (13.5%, Nielsen) and would be worth several hundred billion dollars standalone. DoubleClick, 2007, $3.1 billion: privacy groups revolted, the FTC cleared it 4–1; it built the ad-tech stack that monetizes the open web — and, eighteen years later, produced the antitrust verdict that may yet force the only breakup Alphabet actually faces. The best asset and the biggest legal liability turned out to be the same purchase, bought at what was then the company's largest-ever price.

The pattern worth extracting: Google's existential acquisitions were all distribution purchases, made while critics were valuing them as products. Android was not phone software; it was the right to be the default. YouTube was not a video site; it was the second-largest search box on earth. The company has always understood — constitutionally, since the Yahoo deal of 2000 — that in information businesses, distribution is destiny. Hold that thought for the Gemini-in-Siri deal of 2026.

There is a governance note that separates this company from its Mag 7 conviction-peer, Meta, and it matters for how the history reads. Alphabet shares the dual-class founder-control structure — Page and Brin remain unoutvotable — but the founders delegated operations (to Schmidt, then Pichai) and reserved themselves for interventions: they reportedly returned to hands-on involvement exactly twice, for the 2015 reorganization and the 2023 code-red winter. Where Meta's history is one man's conviction expressed continuously, Alphabet's is professional management punctuated by founder overrides — a structure that produced fewer catastrophic bets (no metaverse equivalent) and slower responses (the Transformer decade), the two faces of the same coin. Investors pricing the two companies' "founder risk" are pricing different instruments: Meta's is always live; Alphabet's is a contingent trigger that history says fires roughly once a decade, and last fired three years ago.

3. Alphabet, Other Bets, and the cost of optionality (2015)

The 2015 reorganization into Alphabet — Google as one subsidiary among bets — was sold as management hygiene ("cleaner and more accountable," in Page's words). It also created the world's most transparent ledger of what long-horizon optionality costs: Other Bets has accumulated roughly $40 billion of operating losses since 2013, running at about $8 billion a year now, with Waymo the largest line. The market spent a decade treating that ledger as founder indulgence. It reads differently in 2026, when Waymo operates 500,000+ paid rides a week at a $126 billion externally validated valuation (Section 11), and the "indulgence" turns out to include the DeepMind acquisition ($400–650 million, 2014) — the purchase that, with Google Brain, gave Alphabet the research bench that eventually built Gemini. As with Meta's Reality Labs and SK hynix's HBM decade, the option book only looks like waste until the cycle that exercises it arrives.

4. The laboratory that armed its competitors (2011–2022)

The deepest irony in modern technology fits in one paragraph. Google Brain (2011) more or less created the industrial deep-learning era; DeepMind's AlphaGo (2016) gave it a global television moment; and in June 2017 eight Google researchers published "Attention Is All You Need," introducing the Transformer — the architecture underneath every frontier AI system that exists today, ChatGPT most famously. The paper's authors listed themselves in random order to signal equal contribution. By August 2023, all eight had left Google — for Character.AI, Cohere, Adept, Essential AI, NEAR, OpenAI and Sakana — a 100% attrition rate on the most valuable eight-person cohort ever employed by one company. Google later paid $2.7 billion (the 2024 Character.AI licensing deal) substantially to bring one of them, Noam Shazeer, back to co-lead Gemini. In June 2026, Shazeer left again — for OpenAI.

Why did the inventor fail to productize its invention? The internal answer, extensively reported: the money machine. A conversational answer engine threatened the click-auction economics of Section 1 — the innovator's dilemma in its purest recorded form, where the disruptive technology was invented inside the incumbent, published openly, and left on the table for anyone whose revenue it didn't cannibalize. OpenAI was, functionally, the commercialization department Google refused to build. The company that spent twenty years terrified of repeating Yahoo's mistake — funding your own replacement — repeated it with a paper instead of a contract.

Yet the same decade produced, almost invisibly, the asset that would fund the redemption. In 2013, Jeff Dean's team ran a projection: if speech recognition rolled out to every Android user at then-current serving costs, Google would need to double its global data-center footprint. The response was to design an inference chip in fifteen months — the TPU v1, deployed quietly in 2015, disclosed only at I/O 2016. It ran AlphaGo. It trained the Transformer's descendants. And because it was treated as plumbing rather than product, it compounded through six generations without ever being submitted to the market's judgment — no quarterly unit disclosures, no analyst day, no competitive positioning. The organizational trait that squandered the Transformer (research divorced from commercialization) is the same trait that protected the TPU (infrastructure divorced from scrutiny). Google's failures and its moats came from one culture, and 2023's crisis forced the culture to finally sort one from the other: Brain and DeepMind were merged that April under Demis Hassabis, research was chained to shipping, and the plumbing was promoted to weapon.

5. Code red: the two years the machine seized (2022–2023)

ChatGPT launched November 30, 2022. Within weeks Google declared an internal "code red"; within ten weeks it had rushed out Bard, whose launch materials contained a factual error about the James Webb telescope — and on February 8, 2023, Alphabet fell 7.7%, roughly $100 billion, in a day. Employees flamed leadership on internal boards ("rushed, botched, un-Googley"). The obituaries wrote themselves: the search monopoly was a Blockbuster, the AI era belonged to the startup, the inventor of the Transformer would be disrupted by it. For about eighteen months, Alphabet traded as the Mag 7's designated structural loser — the cheapest multiple in the group for much of 2023–24, a distinction since inherited by Meta.

The obituary rested on a real vulnerability and a false premise. The real vulnerability: search economics genuinely are exposed to conversational substitution (Part Two quantifies how much). The false premise: that Google lacked the assets to compete — when it in fact held more accumulated AI research, more compute, more data and more distribution than any entity on earth, arranged behind an organizational dam. Code red broke the dam. What poured through over the next three years is the subject of the rest of this report.

For the series' pattern library, file the 2023 trough alongside Meta's $88 and hynix's ₩73,000: three cases in three years of the market pricing organizational or cyclical distress as terminal decline at companies holding decisive but temporarily illegible assets. The Alphabet case is the purest, because nothing about the asset base changed between the obituary and the coronation — no acquisition, no pivot, no new invention. The entire round trip from "designated loser" to "$4 trillion" was a change in organization and disclosure: the same models, chips and researchers, rearranged and shipped. Markets price what companies show, not what they hold; the gap between the two is where this series keeps finding its subjects — and, as Section 16 will argue, Alphabet is the rare case where that gap has now closed from both directions.

6. The comeback: Gemini 3, $4 trillion, and the Apple reversal (2024–2026)

Date the reversal precisely. Gemini 1.0 (December 2023) stabilized the story; 1.5's million-token context (February 2024) gave it a technical edge (and, for disclosure, is the class of capability this research pipeline runs on); 2.0 (December 2024) and 2.5 Pro (March 2025) reached parity. Gemini 3 (November 18, 2025) ended the parity debate: first model past 1500 Elo on LMArena, a 37.4 on Humanity's Last Exam against GPT-5 Pro's 31.6 record, 91.9% GPQA Diamond — the first time since ChatGPT's launch that the frontier's clear leader, by public benchmark, was Google. The app followed the model: 450 million monthly users in July 2025, 650 million by October, 750 million by February 2026, 900 million by May — while AI Overviews reached two billion monthly users inside search itself. The cadence matters as much as the scores: five major releases in twenty-four months, each landing within weeks of announced dates, from an organization that had needed a public humiliation to ship at all. Institutional metabolism, once proven changed, is the hardest competitive variable to reverse — and the one the 2023 obituaries most confidently declared unfixable.

The adoption machinery behind the model deserves separate credit, because it is where the 2023 organization would have failed. Gemini was threaded through every surface the company owns within months of each release: two billion monthly users of AI Overviews inside search, 750→900 million app MAU, Workspace, Android, and — the metric that best measures actual work done — token throughput of 3.2 quadrillion per month across products, up 7× in a year, with the direct API alone processing 16 billion tokens a minute. Token counts are the cloud era's equivalent of query counts: unspinnable, cumulative, and directly billable. Meanwhile the pricing weapon stayed sheathed until it mattered: Gemini 3.1 Pro's $2/$12 per-million-token rates — roughly 40% of GPT-5.5's and Claude's list — are only sustainable because of the silicon economics of Section 10, and they converted the model from a benchmark story into a unit-economics story. The 2017 Google published its weapon; the 2026 Google prices it.

The financial market repriced in steps: $3 trillion in September 2025 (days after the antitrust remedy ruling removed the breakup tail — Section 12), $4 trillion in January 2026, +65% for calendar 2025, the best year since 2009. And in January 2026 came the moment that inverted a decade of narrative: Apple agreed to pay Google roughly $1 billion a year to license a 1.2-trillion-parameter Gemini model to power the new Siri, running on Apple's own Private Cloud Compute. For twenty years Google paid Apple ~$20 billion annually to be the iPhone's default search box — the single largest customer-acquisition expense in business history, and the exhibit at the center of the DOJ case. Now money flows the other way for AI: the company with the most valuable distribution on earth chose, after evaluating everyone, to buy its intelligence layer from the company the market had declared AI's biggest loser thirty months earlier. It is the single cleanest external audit of the comeback — and it should be read alongside the fact that Apple pays because its own models weren't good enough, a sentence nobody in 2023 would have predicted ending that way.

Part Two · Cross-sectional: the six-front war, July 2026

7. Search: the fortress generates record cash while the ground shifts under it

Search & Other revenue grew 19% in Q1 2026 to $60.4 billion — accelerating from 17% the prior quarter, at a scale (~$240 billion annualized) where acceleration should be arithmetically implausible. Pichai attributes it to AI: query volume at all-time highs, AI Overviews monetizing "at approximately the same rate" as traditional results (the company's A/B-tested claim, now repeated for four straight quarters), paid-click share in commercial verticals actually rising as AI answers absorb informational queries and leave commercial ones concentrated. Google's own disclosure: AI Overviews serve two billion users monthly; total "search-like" queries across the ecosystem are growing, not shrinking, because the pie of asked questions expanded 26% once asking became conversational.

The independent data says something more uncomfortable. Zero-click searches hit 68% of US queries in early 2026 (SparkToro/Similarweb), up from 60% in 2024; where an AI Overview appears, organic click-through falls by more than half (Pew: 8% vs 15%) and the links inside the Overview are clicked about 1% of the time; the third-party publisher web — the Network segment — is the canary already singing, down 4% year over year in the same quarter Search grew 19%. And AI Mode, the full conversational interface, sends outbound clicks at one-tenth the traditional rate; it is only 0.34% of queries today, which is either reassurance or a fuse depending on your priors.

Both datasets are true, and the reconciliation is the entire search debate: Google is successfully converting the monetizable core of search (commercial intent) to AI formats while the unmonetized periphery (informational queries, publisher traffic) is cannibalized first and fastest. Revenue lives in the core; the ecosystem that feeds the core lives in the periphery. The bull case is that intent is the fortress and the fortress holds — twenty years of advertiser tooling, measurement and habit do not migrate to a chatbot that lacks all three. The bear case is that the periphery is the moat — the open web Google indexes, the publishers it pays with traffic, the habit loop of the question box — and that its erosion shows up in the core with a lag measured in years, unbilled until it arrives. The honest position: the reported numbers have voted bull for eight consecutive quarters, and the structural data keeps voting bear, and the lag between them is the trade.

The next battlefield is already staked out, and it is transactional rather than informational: agentic commerce, where an AI assistant completes the purchase instead of handing off a click. Here Google has moved with un-2023-like speed — the AP2 agent-payments protocol (September 2025, with Mastercard, PayPal and Amex among 60+ partners) and the UCP checkout protocol with Shopify (March 2026) now anchor the standards layer, while OpenAI's rival Instant Checkout effort was shut down and refolded into ChatGPT Apps in March. Owning the payment rails of agent commerce is the twenty-year echo of owning AdWords: if buying migrates from clicking to delegating, Google intends to tax the delegation. It is early, contested, and the most direct answer to the deepest bear case — because if the assistant completes the transaction, the "zero-click" problem inverts into a take-rate opportunity. Search's terminal value question may ultimately be decided not by where questions are asked but by whose protocol moves the money.

8. Models: from fudged demo to frontier lead — with a caveat named depth

The Gemini 3 lead is real but narrow: the top of LMArena compresses within ~55 Elo points (Claude Opus 4.8, GPT-5.5 Pro, Gemini 3.1 Pro), and leadership rotates with each release. Where Google's position is structurally strong rather than cyclically strong is price and integration: Gemini 3.1 Pro's API pricing ($2/$12 per million tokens) undercuts closest peers by roughly 60%, a direct expression of the TPU cost advantage (Section 10) — Google is the only frontier lab whose training and serving costs are not paying NVIDIA's margin. Token throughput tells the adoption story: 3.2 quadrillion tokens processed monthly across products (7× in a year), 16 billion tokens per minute through the direct API, 375 cloud customers each consuming a trillion-plus tokens a year.

The Apple-Gemini deal deserves its economics spelled out, because the $1 billion price tag understates it in both directions. Understated for Google: the contract puts Gemini inside the intent stream of a billion iPhones — the exact distribution position ($20 billion a year of default payments) that built search's mobile decade, now re-established for the AI decade at negative cost. Understated as risk: Apple licensed a model, not a marriage — Private Cloud Compute keeps the data on Apple's side, the contract reprices annually under the antitrust remedy's 12-month term limits, and Apple's own silicon team has every incentive to make the dependency temporary. The deal is best read as the AI era's version of the 2000 Yahoo-Google contract, with Google cast, this time, in its own old role: the supplier whose product is good enough that the distributor cannot avoid strengthening it. Whether Apple eventually plays Yahoo's part in that story — funding its own replacement's rise — or Microsoft's OpenAI part — riding the supplier while building leverage — is one of the decade's better questions, and Google, having lived the first script, is presumably writing contract terms against it.

The caveat is engagement depth, the same metric that haunts Meta AI. Gemini's 900 million MAU is a number achieved substantially through Android integration and search adjacency; ChatGPT's 900 million weekly actives and ~2.5 billion daily prompts represent chosen, habitual use — the brand-name verb of the category. In the enterprise API market, Anthropic reportedly wins ~70% of head-to-head evaluations, and OpenRouter's developer-flow data shows a fragmented market in which no one — including Google — holds more than a quarter. Google has converted its research bench into frontier capability and cost leadership; it has not yet converted either into the default consumer relationship (OpenAI's) or the default enterprise relationship (contested, with Anthropic strongest in high-stakes work). Its compensating asset is the one it always had: it does not need users to choose Gemini, only to keep using Google — the assistant arrives through the side door of two billion existing habit loops. Whether side-door adoption produces durable preference is the open question; the Meta AI experience (1.2 billion reached, 40 million engaged) is the cautionary counterexample.

9. Cloud: the receipt machine

Google Cloud is the cleanest business in the company right now and the reason the market pays Alphabet's capex bill without complaint. The numbers compound on every axis: revenue $20.0 billion in Q1 (+63%, the fastest of the big three against AWS's +19% and Azure's +40%), operating margin 33% (from 21% a year earlier — a 12-point march in four quarters), and a backlog of ~$462 billion that nearly doubled in one quarter, roughly half recognizable within 24 months. GCP remains third in share (13% vs AWS 30%, Azure 25%), but in the AI-workload segment specifically it is the insurgent taking share with price (TPU economics) and models (Gemini exclusivity) as the wedge.

The margin trajectory deserves as much attention as the growth rate, because it answers the "can clouds make money on AI" question the 2024 bears asked: 20.7% → 23.7% → ~30% → 33% in four consecutive quarters, while absorbing the depreciation of the very capex wave financing it. Operating leverage of that steepness at $80 billion run-rate scale means the fixed costs are already largely bought — each incremental AI workload lands on infrastructure whose cost the P&L has begun digesting. The enterprise strategy sharpening the wedge: the $32 billion Wiz acquisition (closed with Q1's bond proceeds) bolts the fastest-growing name in cloud security onto GCP's compliance story — the classic objection to third-place clouds — and Gemini Enterprise seats grew 40% sequentially. GCP is no longer selling "cheaper compute"; it is selling the only place to rent TPUs, the house models, and now the security layer, a bundle neither AWS (no frontier model) nor Azure (no silicon) can copy in full.

One name dominates that backlog, and it deserves its own risk paragraph: Anthropic's $200 billion five-year commitment is 40%+ of the RPO — paired with Google's planned investment of up to $40 billion into Anthropic itself ($10 billion immediate, $30 billion in milestones), on top of its existing equity stake (whose revaluation contributed to the $36.9 billion unrealized gain distorting Q1 net income). Read charitably: Google captured the fastest-growing AI lab as anchor tenant, equity holder and silicon customer simultaneously — triple-dipping on Anthropic's rise. Read skeptically: a meaningful fraction of Alphabet's "receipt" is circular — Google invests in Anthropic, Anthropic commits the money back as cloud spend, the commitment inflates the backlog that justifies the capex that builds the TPUs that Anthropic consumes. The circularity is not improper (vendor financing is as old as computing, and Anthropic's ~$30 billion run-rate revenue is real third-party money), but concentration is concentration: the marquee receipt has one signature on it, and that signature also anchors a large fraction of Broadcom's AI backlog (per our SK hynix research) and a slice of Microsoft's. The AI economy's revenue certificates cross-reference each other; Alphabet holds more of them than anyone, and audits them least publicly.

History offers Google a specific caution here, because it has played the other role in this exact transaction. In 2000, Yahoo paid Google to supply search, believing it was buying a component; it was funding a successor. In 2026, Google supplies Anthropic with compute, silicon and capital, believing it is monetizing a customer. Anthropic, meanwhile, accumulates frontier capability on the cheapest infrastructure available, keeps its models portable across clouds, and wins the enterprise evaluations Gemini loses. The relationship is genuinely symbiotic today — Google gets the receipt, Anthropic gets the TPU discount — and nothing obliges it to stay symmetric. The 2000 lesson is not "don't supply your competitor"; Google's own rise proves suppliers can't tell. The lesson is that whoever owns the customer relationship at the end owns the economics, and in the Anthropic account, that question — whose customer is the enterprise, the model's or the cloud's? — is precisely what the ~70% enterprise win rate leaves open.

10. Silicon: the decade-old side project that became the second-biggest AI chip business on earth

The TPU story is the purest expression of this report's thesis — an underpriced internal asset, revalued by circumstance. Built from 2013–2015 (15 months from kickoff to deployment) because Jeff Dean calculated that voice recognition alone would force Google to double its data centers, the TPU spent a decade as invisible internal plumbing through six generations. Ironwood (v7, GA November 2025) ended the invisibility: 9,216-chip pods, 42.5 FP8 exaflops, 192GB of HBM per chip, and — per SemiAnalysis — throughput competitive with NVIDIA's GB300 at a 20–50% lower total cost per effective FLOP for large buyers, because Google buys dies from Broadcom and builds its own racks, skipping NVIDIA's system-level margin entirely.

The supply chain behind the scale-up connects this report to the rest of the series. Broadcom designs and assembles Ironwood and holds the contract through 2031 (Google-driven custom silicon is the largest block of Broadcom's exploding AI revenue); MediaTek was brought in for lower-cost inference variants targeting two million units in 2027 — a deliberate second-sourcing of Google's own supply chain; TSMC fabs everything; and at the HBM layer Google queues like every other buyer — Samsung holds 60%+ of TPU HBM3E supply with SK hynix the balance, which is why our memory reports keep meeting this company from the other side of the counter. When TrendForce reports HBM demand exploding "on ASIC growth," this section is what that sentence means: the TPU program is now large enough to be a first-order driver of the memory supercycle it buys into.

Then it went on sale. The October 2025 Anthropic deal — up to one million TPUs, over a gigawatt, "tens of billions" of dollars, structured as ~400,000 chips sold outright via Broadcom racks and ~600,000 rented through GCP — was the first at-scale external validation that the TPU is a merchant product. The reported Meta negotiations (rent in 2026, buy for 2027 deployment) would be the second, and more symbolically violent: the tenant of our previous report diversifying away from NVIDIA using the silicon of NVIDIA's only structural rival. Sell-side flow-through puts Google-driven custom silicon at ~$21 billion of Broadcom's 2026 revenue, doubling in 2027; DIGITIMES has Google at 46% of the entire ASIC accelerator market this year, ~3.3 million chips. The honest caveats: external tooling still trails CUDA badly (the TorchTPU effort is new), third-party benchmarks of rented TPU economics are far less flattering than internal-cost analyses (Artificial Analysis measured NVIDIA ahead on tokens-per-dollar at list rental prices), and NVIDIA's data-center revenue is still growing 50% — the TPU is eroding NVIDIA's pricing power (OpenAI reportedly extracted ~30% fleet discounts merely by threatening to switch) faster than its share. But note what Alphabet now is: the only company that designs frontier models, the silicon they run on, and the cloud that rents both — the vertical integration NVIDIA, OpenAI and Microsoft each possess only two-thirds of. In the framework of our memory-complex reports: Alphabet is the only demand-side player that has internalized its own supply chain above the HBM layer. (At the HBM layer, it queues with everyone else — Samsung supplies 60%+ of TPU HBM3E, SK hynix the balance.)

11. YouTube and Waymo: the two franchises the sum-of-parts always forgets

YouTube is quietly the largest television network in America: 13.5% of all TV watch time (Nielsen, March 2026), #1 among all media distributors, with TikTok structurally locked out of the living-room screen where CTV ad dollars migrate. Ad revenue of $9.9 billion in Q1 (+11%) understates the franchise: subscriptions (Music/Premium had their best non-trial quarter since 2018; NFL Sunday Ticket anchors sports) push total YouTube-attributable revenue well past $50 billion annualized — bigger than Netflix — inside a segment line most models still treat as "other advertising."

Strategically, YouTube is also Alphabet's quiet answer to two of its scariest questions. Against the attention-migration threat (TikTok's 95 daily minutes vs YouTube's 85, per Sensor Tower), YouTube is the only Western property that fights short-form on mobile and owns the big screen — Shorts absorbs the TikTok behavior while CTV, where TikTok has no seat, collects television's ad-budget migration; the Sunday Ticket losses (Morgan Stanley estimated $1.2 billion a season) are the tuition for that living-room position. Against the content-supply threat from Section 14: YouTube is the largest proprietary, self-refreshing, rights-cleared training and grounding corpus in existence — 500 hours uploaded a minute that no rival lab can license, crawl, or replicate. In a world where the open web's supply function is breaking, the walled garden of user-generated video may prove the single most valuable data asset in AI, and it sits inside a segment line the sum-of-parts crowd values as a mid-size ad network.

Waymo graduated in 2026 from science project to the only at-scale autonomous ride business in the West: 500,000+ paid rides weekly (10× in two years), eleven US cities with 11+ more announced plus London and Tokyo, a $16 billion round in February at a $126 billion post-money valuation — externally priced, majority-outside capital. The unit economics remain honestly unproven (a ~$5 billion annual Other Bets drag, fleet costs of ~$1.40/mile against Tesla's theoretical $0.81), and Tesla's Miami launch keeps the narrative contest alive. But the operational gap is not a narrative: Waymo runs ~3,600 driverless vehicles doing half a million weekly rides; Tesla's Austin deployment, per city filings, runs roughly 20 unsupervised cars doing low thousands. Three orders of magnitude, in Waymo's favor, priced by the market at one-tenth of Tesla's robotaxi-flavored market cap. One of those two prices is wrong, and our Tesla report (forthcoming in this series) takes up the other side. For Alphabet, the investment point is simpler: at $126 billion externally validated, Waymo alone is worth ~3% of the company — roughly the entire Other Bets ledger's historical cost — and the market's sum-of-parts habitually values it at zero.

The economics are on a knowable path even if unproven: roughly 15 million rides in 2025 at ~$18 average produced ~$270 million of revenue against multi-billion costs; 2026 tracks toward 40–45 million rides and $700–800 million; the current fleet grows ~265–300 vehicles a month; and the next-generation platform (a ~$32,000 vehicle base against today's ~$125,000 Jaguars) shortens third-party payback estimates from a year to a few months per vehicle. The wildcard is that fleet cost is now the constraint, not demand or regulation — the same shape as early AWS, where the business was capex-gated long before it was demand-gated. Alphabet has seen that movie from the inside, which is presumably why it let outside capital price the round rather than consolidating the spend: Waymo is being groomed as a separable receipt, the Other Bets ledger's first graduation. A 2027 IPO would convert twelve years of "indulgence" into the sum-of-parts line item the market has always refused to book — and would be worth watching for what it says about how Alphabet plans to crystallize its other invisible assets.

12. The courtrooms: the breakup that didn't happen, and the one still loaded

Alphabet's legal decade concentrated into eighteen decisive months. The search case: Judge Mehta ruled Google a monopolist in August 2024, and in September 2025 delivered remedies that spared everything that mattered — no Chrome divestiture, no Android divestiture, default payments permitted (non-exclusive, 12-month terms, with data-sharing duties to "qualified competitors" under a technical committee through 2031). The market read it correctly as acquittal-in-effect: Alphabet crossed $3 trillion within days. Both sides' appeals (DOJ wants the breakup back; Google wants the data-sharing gone) keep a tail alive, but remedies execute meanwhile, and the Apple default deal survived — transformed, ironically, into a two-way relationship by the Gemini-Siri contract.

The search remedy's fine print is worth two sentences beyond the headline, because it created a standing institution rather than a one-time judgment: a five-person Technical Committee (fully staffed May 2026) supervises data-sharing with "qualified competitors" through 2031 — search index and interaction data licensed at marginal cost, five-year terms, declining usage caps. It is the most intrusive behavioral remedy ever applied to a US tech company, and its practical effect is the live experiment: whether access to Google's data, without Google's infrastructure and habit loop, lets a Perplexity or an OpenAI actually convert. Early returns say the moat was never just the data — which, if it holds, quietly validates the highest-conviction bull argument about the fortress.

The ad-tech case is the live one. Judge Brinkema found Google illegally monopolized publisher ad serving and the exchange (the DoubleClick estate) in April 2025; the DOJ seeks divestiture of AdX and DFP; the remedy ruling — expected by her own March 31, 2026 deadline — is now overdue, with her closing-argument skepticism about forced sales (no obvious buyer; any buyer needs its own antitrust review) the main tea leaf. A divestiture order would be the first structural break of a US tech giant since AT&T; even so, the assets at risk generate a single-digit share of revenue and their loss might, in the darkest reading for publishers, simply strand more budget in Google's owned-and-operated properties. Europe adds friction, not fracture: the €4.1 billion Android fine final as of July 2, 2026 (opening civil-damages season), a record DMA fine over search self-preferencing expected before August, cumulative EU penalties above €11 billion. The pattern across every venue: fines Alphabet can pay from three weeks of free cash flow, conduct rules that constrain the edges, and — so far — no court willing to actually dismember the machine. The tail risk is real but has a name and a date: Brinkema's overdue ruling is the single document that could still change the corporate structure.

Step back from the dockets and note what the whole legal decade proved about the moat, since plaintiffs' discovery is the most adversarial audit a business model ever receives. Two US courts examined the machine for a combined seven years and concluded, in effect, that its dominance owes more to compounding advantages — data, defaults, habit, integration — than to any single removable practice; even the remedies that survived (data sharing, default-term limits) attack the inputs to the moat rather than the moat, on the theory that competitors given the inputs can rebuild it. Eight months into the data-sharing regime, none visibly has. For an investor, the litigation record thus reads as an involuntary fairness opinion: the strongest-resourced adversaries on earth, with subpoena power, could not locate a load-bearing wall to remove. What they could do — and did — is meter the tolls at the edges, which is margin friction, not thesis damage.

Part Three · Synthesis: the two numbers that disagree

13. Where it stands: the July 2026 snapshot

The segment engine, Q1 2026 (year over year):

Segment Revenue Growth Note
Search & Other $60.4B +19% accelerating, 2nd straight quarter
YouTube ads $9.9B +11% ex-subscriptions
Network $7.0B -4% the canary segment
Subscriptions/Platforms/Devices $12.4B +19% Premium, One, Pixel
Cloud $20.0B +63% margin 33%, backlog ~$462B
Other Bets $0.4B — op. loss $2.1B
Total $109.9B +22% op. margin 36.1% (+2pts)

Read the table as a machine with one decaying part: every line accelerating or compounding except Network — the third-party publisher web — which shrinks as AI answers starve the sites it monetizes. Alphabet is, on its own income statement, simultaneously the chief beneficiary and the chief chronicler of the open web's decline; the -4% is the smallest line in the table and the most informative, because it is the only place the 68% zero-click number currently touches reported revenue. Its decay rate is this report's equivalent of the memory complex's spot prints — the leading indicator hiding inside the lagging document.

Full-year 2025: revenue $402.8 billion (+15%), operating profit $129 billion, EPS $10.81 (+34%), record operating cash flow of $164.7 billion. Q1 2026 EPS of $5.11 (+82%) requires the asterisk the sell side keeps forgetting: $36.9 billion of unrealized gains on non-marketable equity (the Anthropic revaluation class) sits in other income; strip it and core operating growth is the honest +30%. The 2026 consensus: revenue ~$486 billion, EPS ~$14.2 (upgraded from $11.6 in six months — the fastest estimate migration in the mega-cap group). Capex: $91–93 billion actual in 2025, guided to $180–190 billion for 2026 (raised again in April, ~60% servers), with 2027 flagged "significantly higher" and the CEO describing the company as "compute constrained." The balance sheet took its first real leverage in company history to fund the program and the Wiz ($32 billion) and Intersect Power ($5.9 billion) closings: $31 billion of new bonds in Q1, long-term debt now $77.5 billion against $126.8 billion of cash. Shareholder returns continue at maintenance levels: $70 billion buyback authorization, dividend up 5% to $0.22.

Valuation, against the family (July 3–4 data): $4.39 trillion market cap — now third, having passed Microsoft and Amazon during the 2025 run — at 27.5× trailing and 24.7× forward. On 2026 consensus the PEG sits near 0.8. Alphabet has traveled from the Mag 7's cheapest member (2023–24, the AI-loser discount) to mid-pack premium (a four-turn premium to Microsoft on forward earnings) in roughly seven quarters. The re-rating is the comeback, monetized; it also means the margin of safety that defined the 2023–25 entry is spent. What remains is a fair price for a machine whose two most important numbers point in opposite directions. One technical footnote for anyone modeling from headlines: the reported Q1 EPS of $5.11 (+82%) is the most distorted print in the Mag 7 this season — strip the $36.9 billion unrealized private-equity mark and underlying EPS grew roughly 30%, in line with operating profit. Sell-side screens quoting Alphabet's "82% EPS growth" are quoting Anthropic's valuation round, not Google's operations; the irony of the receipt machine's best quarter being partly an accounting echo of its biggest customer's fundraise is left as an exhibit for Section 14.

14. The two numbers: $462 billion versus 68%

Every long thesis and every short thesis on Alphabet in 2026 reduces to which of two numbers you weight.

$462 billion — the cloud backlog. Contracted, signed, half of it recognizable within 24 months, growing ~100% quarter over quarter. It is the largest receipt in the AI economy, the thing Meta conspicuously lacks, the reason the market rewards Alphabet's $190 billion capex, and the hard evidence that the AI transition is adding a second engine to the company rather than merely defending the first. Weight this number and Alphabet is the best-positioned company on earth: the only full-stack AI vertical, with the cheapest silicon, a frontier model, and a receipts drawer its peers would kill for.

(Apply this series' standard audit to the receipt before weighting it: backlog is a promise, not cash — CoreWeave taught that lesson from the landlord side. Roughly 40% is one customer, Anthropic, whose commitment Google partly finances; the recognition schedule beyond 24 months is management's estimate; and cloud contracts, unlike hynix's HBM contracts, typically carry consumption flexibility inside committed floors. It is the best receipt in the AI economy and it is softer than the market's shorthand treats it. Both clauses matter.)

68% — the share of US searches that end without a click, up eight points in two years. It is the leading indicator of the question this company has never had to answer: what is intent worth when the answer arrives without a visit? Behind it stand the -4% Network segment, the 1% click-through inside AI Overviews, the publisher ecosystem visibly starving, and the awkward fact that the fastest-growing interface (AI Mode) monetizes an order of magnitude worse than the one it replaces. Weight this number and the cloud triumph is a second engine bolted on while the first engine — still ~55% of revenue and a much larger share of profit — burns its own fuel supply.

There is a third number lurking behind the two, and it may matter more on a decade view: the content supply. Google's answers — classical and AI alike — are synthesized from an open web whose economics Google's own success is dismantling. Publishers losing 60% of organic clicks do not keep publishing at the old rate; the 1% click-through inside AI Overviews is, from the supply side, a starvation ration. If the open web's production function breaks, every answer engine — Google's, OpenAI's, everyone's — inherits a corpus that stops refreshing, and the differential advantage shifts to whoever owns proprietary refreshing corpora: YouTube uploads, Maps data, Gmail-adjacent signals, licensed archives. Alphabet owns more such corpora than any rival, which converts even the darkest open-web scenario into relative advantage. But "we'll starve slower than the ecosystem we built" is a grim moat, and the licensing economy now forming (every AI lab signing content deals) is effectively the industry repricing, in cash, the traffic subsidy Google used to pay in clicks. Watch those licensing costs: they are the open web presenting its invoice.

The reconciliation the bulls offer: search revenue accelerates because the periphery is unmonetized — losing free traffic costs nothing while AI formats concentrate commercial intent. The reconciliation the bears offer: every platform transition in this report's Part One (portals→search, desktop→mobile, feed→short-video) looked exactly like this in its revenue lines for years before it didn't, and 0.34% AI Mode share is where 100% of the growth in interface is happening. Our read, consistent with the series' method of preferring falsifiable statements: the disagreement is empirically live, the reported numbers have sided with the bulls for eight straight quarters, and the correct posture is not to pick a side but to identify the crossover indicators — Network segment decay rate, AI Mode query share, paid-click share in commercial verticals — and watch them with the same discipline our memory dashboard watches NAND prints. The 2026 twist is that Alphabet, uniquely, gets paid on both sides of its own disruption: every query that migrates to anyone's AI still runs on rented compute, and Alphabet sells the compute, the chips and the model API. The fortress may erode; the company is building the siege engines too, and charging the besiegers rent.

15. Three scenarios for the year ahead

Anchors: ~$486B consensus revenue, ~$14.2 EPS, $180–190B capex, 24.7× forward at $360.

Note the scenarios' cross-exposure to the rest of the series before choosing one, because Alphabet sits on more sides of the AI market than any other subject we cover. Scenario A is directly bullish for Broadcom and the memory complex (TPU volumes are HBM demand — our SK hynix report's ASIC tier growing), mildly bearish for NVIDIA (pricing power erosion), and neutral-to-bearish for the neoclouds (GCP absorbs AI workloads they wanted). Scenario C's search-crossover version is, perversely, not bearish for the AI suppliers at all — queries migrating to AI interfaces consume more compute, not less, wherever they land. The only Alphabet scenario that damages the whole chain is a generalized capex retreat, which is Meta's Scenario C and hynix's air pocket wearing a different ticker. The four-sided market keeps reducing to one wager: the October disclosures remain the synchronized reveal.

Scenario A — the full-stack thesis compounds. Cloud sustains 50%+ growth as backlog converts; TPU external sales expand beyond Anthropic (a confirmed Meta purchase agreement is the headline to watch); Gemini holds frontier parity through the GPT-6 generation; search revenue keeps accelerating through the format shift. EPS runs toward the high end of upgraded estimates and the multiple holds or expands modestly — Alphabet challenges NVIDIA and Apple for the top of the index. This is the trajectory of the last three quarters, extrapolated; its enemy is the law of large numbers and the fact that it is now, for the first time since 2022, the consensus scenario, priced accordingly.

Scenario B — receipts decelerate, fortress holds (modal). Cloud growth glides toward 40% as comps steepen; the Anthropic concentration caps backlog optics; search grows low-teens; the ad-tech remedy lands as conduct rules rather than divestiture. EPS grows mid-teens, the multiple drifts back toward 22×, the stock marks time in a $320–400 band while the 2025 re-rating digests. Most large-cap years look like this; after a +65% year, reversion to it is the base rate. What distinguishes Alphabet's muddle from Meta's (whose Scenario B is a waiting room with a depreciation clock): Alphabet's B is stable — cash flow funds the capex with room to spare, the dividend and buyback continue, the option book (Waymo, quantum, Isomorphic) keeps ripening on external money, and no 2027 mechanism forces a resolution. Investors can be paid to wait here in a way Meta's structure does not allow, which is what a four-turn multiple premium buys.

Scenario C — the crossover arrives early. Some combination of: AI Mode share inflects upward while its monetization gap persists; a Brinkema divestiture order forces the first structural break and reopens every other case; the Anthropic circularity draws formal scrutiny just as cloud comps steepen; ChatGPT's commerce stack starts visibly routing transactions around search. Search multiple compression does what it did to IBM and Oracle in their transitions — the earnings hold but the market re-prices the terminal value — and the stock gives back 25–35% while remaining "cheap" the entire way down. The tell that C is arriving will not be in Alphabet's own P&L first; it will be in the Network segment's decay steepening past ~-10% and in third-party click data, the same way SNDK's spot prints lead the memory complex.

The IBM comparison deserves precision, because it is the bears' best historical template and it half-fits. IBM in 1990 also led its industry's research, owned its stack, and generated enormous cash from a franchise (mainframes) whose growth — not existence — had ended; the stock spent a decade compressing while earnings held, because terminal-value repricing needs no earnings miss. The half that doesn't fit: IBM's successor technologies were sold by others, while Alphabet owns the leading candidates for its own succession (Gemini, Cloud, the TPU, YouTube, Waymo). A fair reading of the 2025 re-rating is that the market examined the IBM template and rejected it on exactly this ground. Scenario C is the case where the rejection was premature — where owning your successors matters less than the margin difference between the franchise they replace and the businesses they are.

16. The valuation paradox, Alphabet edition

Place the whole family on the table first (July 3–4 data): NVIDIA $4.72T at 29.8×/15.3×; Apple $4.53T at 37.3×/32.1×; Alphabet $4.39T at 27.5×/24.7×; Microsoft $2.90T at 23.3×/20.2×; Amazon $2.61T at 31.6×/24.5×; Meta $1.48T at 21.2×/15.8×; Tesla $1.48T at 357.7×/154.5×. Alphabet's trailing-to-forward ratio of 1.1× is the narrowest in the group — the market expects almost no earnings growth deceleration and prices almost no cyclicality. For a company whose profit pool is majority advertising (a cyclical) plus cloud (a growth asset with steepening comps), that flatness is itself a statement: the market has reclassified Alphabet from "ad cyclical with science projects" to "diversified compute utility." Reclassifications of that kind are where multi-year returns hide — and where they get taken back.

The series' recurring paradox — record earnings, disbelieving multiple — appears here in inverted form: Alphabet's multiple has already re-rated (21× → 27.5× trailing in eighteen months) while the structural question that justified the old discount remains empirically unresolved. The market has, in effect, paid out the comeback in full and moved the search-disruption risk from the price into the tail. Compare the family: Meta carries a five-turn discount for an unproven capex program with an accelerating core; Alphabet carries a premium on a proven program with a core whose ten-year durability is the single most contested question in technology. By the logic of our Meta report — that allocation discounts are removable and structural discounts are not — the market has these two priced on the wrong sides of each other, unless one believes the search fortress is now safer than the feed fortress. That is a defensible belief (intent monetizes better than attention; the Apple-Gemini deal armored the default position; the courts blinked), but it is a belief, and eight quarters of good prints have quietly promoted it from thesis to assumption. The prudent read of Alphabet at 24.7× forward is not "expensive" or "cheap" — it is fully narrated: the first company in this series whose price contains no visible mispricing to exploit, only a bet that the narration continues. In a series about finding the gap between story and number, Alphabet in July 2026 is notable for having closed its gap — which is itself information, and a warning: the reports in this series that found the most alpha (Sandisk at $4B, hynix at 5.5× forward, Meta at 15.8×) all began where the narration had failed. Here it has succeeded, and succeeded recently.

17. What would have to be true, the risk matrix, and five signals

For Scenario A to pay from $360, most of the following must hold:

  1. Cloud revenue growth stays above ~45% while margins hold 30%+ (the receipt machine is the multiple's foundation now);
  2. Search & Other growth stays double-digit through the AI-format transition;
  3. The Brinkema remedy avoids divestiture (or a divestiture proves as contained as the market hopes);
  4. TPU external sales produce a second anchor customer beyond Anthropic;
  5. No enforcement or disclosure event reframes the Anthropic backlog concentration.

Risk matrix (probability × severity):

Two items deliberately absent from the matrix, with reasons: quantum (the Willow "verifiable advantage" milestone of October 2025 is scientifically real and commercially undated — an option on an option, unpriceable at this range) and Isomorphic Labs (the $2.1 billion externally funded drug-design spinout; same logic, pharma timelines). They are mentioned because a full-stack accounting of Alphabet's option book — Waymo, DeepMind's science arms, quantum — is the strongest version of the sum-of-parts bull case, and because the Waymo precedent (twelve years of "waste," then a $126 billion external mark) counsels against zeroing any of them. But options that cannot fire within the report's horizon do not belong in its risk matrix; they belong in this paragraph, on the record, for the 2028 edition to grade.

Five signals, in firing order:

  1. Q2 earnings (late July): cloud growth vs the 63% bar, backlog progression ex-Anthropic if disclosed, Network segment decay rate — the first two read Scenario A, the third reads C.
  2. The Brinkema remedy ruling (overdue): divestiture or conduct rules — the one court document left that can change the structure.
  3. Meta-TPU confirmation or lapse: a signed purchase agreement validates the merchant-silicon thesis; a quiet death re-contains the TPU inside GCP.
  4. AI Mode query share and monetization disclosures: the crossover dial; any quarter it exceeds ~2% of queries without a monetization update is a C-scenario data point.
  5. October capex season: Alphabet's own 2027 frame plus the tenants' — the same synchronized disclosure event every report in this series now watches, because all four sides of the AI market publish their mutual dependency in the same two weeks.

Signal design notes, as throughout the series: ordered by arrival, chosen for independence — the earnings print reads the receipts, the court ruling reads the structure, the Meta-TPU item reads the silicon thesis, the AI Mode data reads the fortress, October reads the whole chain. Signals 1 and 5 are calendar-fixed; 2 is overdue and could land any Friday; 3 and 4 are event-driven. The asymmetric one is Signal 2: it is the only binary in the set, the only one with structural (rather than narrative) consequences, and the market's current price treats its benign outcome as roughly settled — which makes the malign branch the cheapest tail hedge in the file.

The bottom line. Alphabet is the only company that occupies three sides of the AI economy at once — it designs the silicon, rents the compute, and ships the frontier model — while defending the largest legacy franchise ever exposed to a platform transition. The comeback from code red to $4 trillion is the most impressive execution story in the Mag 7, and the market has paid it in full and promptly. What remains at $360 is a fair-value bet on two propositions: that the search fortress converts rather than erodes, and that the receipt machine's growth outruns its own comps and its one enormous, circular, brilliant anchor customer. The company's history — $750,000 refused, $50 million for Android, eight authors underpriced out the door, a decade of TPUs treated as plumbing — is one long documentation of the world undervaluing what Google holds. The 2026 risk is the mirror image: that for the first time, after the great re-rating, the world has finally priced it correctly — or a turn past.

One closing symmetry for the series' pattern library. In 2000, Yahoo hired Google to answer its users' questions, and the supplier became the successor. In 2026, two billion people ask Google's AI their questions, Apple rents Google's model for Siri, Anthropic trains on Google's chips, and Meta may deploy them — the whole industry, rivals included, increasingly runs some layer of its intelligence through Mountain View's stack. Alphabet has spent twenty-five years making sure that whichever company owns the next interface, it would own a toll booth underneath. That is the deepest version of the bull case, and it has one precondition the company cannot engineer: that there remains, at the bottom of the stack, a question box someone profitably answers. Watch the five signals; the first one arrives this month.

Sources

Report generated by the Aya Invest research pipeline. Every claim above traces to a public source; where figures are third-party estimates (zero-click share, TPU economics, Waymo unit costs) or reported-but-unconfirmed (Meta-TPU negotiations), the text says so. The pipeline runs on long-context frontier models, a category in which Alphabet is a vendor; we note the fact for transparency. For information and research purposes only. Not investment advice.

FAQ

Is AI killing Google search?

Both sides of the ledger are real. About 68% of Google searches now end without a click, and AI answers erode the click-through economics that funded the web for two decades. Yet search revenue keeps growing, AI Overviews monetize at comparable rates, and the deeper moat — distribution — was just re-certified by two events: a federal court preserved Google's ~$20B/year default payment to Apple (non-exclusive, annual renewal), and Apple chose Gemini, at ~$1B a year, as the brain of the rebuilt Siri. The company most threatened by AI answers is also the one selling the answers.

What is Alphabet's AI backlog and who is Anthropic in it?

Google Cloud's backlog reached $462B, with roughly 40% attributable to Anthropic's multi-year commitments — the same four-sided pattern the series documents at Amazon (whose $364B backlog predates a further $100B Anthropic deal) and Microsoft (whose $627B RPO is ~45% OpenAI). Each hyperscaler's receipts lean heavily on one frontier lab, which is why the series treats backlog concentration, not backlog size, as the number to watch in October's capex season.

Why do TPU merchant sales matter?

For a decade TPUs were Google's private cost advantage; selling them to external customers converts a capex line into a product line and puts the first credible merchant-silicon alternative next to NVIDIA at hyperscale. It also reframes Alphabet's $185B of 2026 capex: part of that spending now manufactures a sellable good rather than only internal capacity — the cleanest kind of receipt this series tracks.

© Aya Research · About · invest.aya-ai.org · For information and research purposes only. Not investment advice.