Alphabet AI, Part II: Gemini Stopped Taking Share in March. It Barely Matters. The TPU Carry Trade Is the Whole Story.

Companion to this morning's GOOGL deep dive · Published Monday, August 24, 2026 · Reference prices: Friday Aug 21, 2026 close · @dailyanalysts
The answer in three lines. (1) Yes, Gemini lost momentum — but not the way the headlines say. Its user count is still rising; its share has been flat since roughly March 2026, and every point ChatGPT has bled has gone to Claude, not to Google. (2) No, I do not expect a consumer share re-acceleration before Q1 2027, because the flagship Gemini 3.5 Pro appears to have been shelved in favour of the Flash efficiency line and four of Google's most senior researchers walked out the door on Aug 5. (3) And it barely matters to the stock, because Alphabet's AI P&L is not the chatbot — it is renting TPUs to everyone else's models, backed by a credit guarantee that gives TPU tenants a 220bp borrowing advantage over Nvidia tenants. That guarantee is the real moat, and it is also the real risk.

1. Is Gemini losing momentum? Three independent measurement systems, one answer

The reason this question produces contradictory headlines is that people quote two different metrics as if they were one. Users and share are moving in opposite directions.

Users: still climbing

Share: flat since roughly March 2026

Measurement systemWhat it measuresGemini trendChatGPT trendClaude trend
FirstPageSage (Jan–Jun 2026)Chatbot market share14.0% → 14.1% → 13.8% → 13.2% → 13.5% → 13.3% (−0.7pp)67.1% → 58.6% (−8.5pp)8.5% → 11.5% (+3.0pp)
Similarweb (through May 2026)Web visits~27.5% share, 2.65B visits53% share, 9.5B visits, slipped below 50% for the first time in June9% share, 0.9B visits
Sensor Tower / Similarweb (via tech press, Mar→May 2026)AI-assistant share25.5% → 27.7%50.0% → 46.4%6.0% → 10.3%
BNP Paribas (Aug 14, 2026)Daily active usersHeld at 19.8% — flatRegaining momentum

Methodology note, stated up front: the absolute levels are not comparable across these sources — FirstPageSage puts Gemini at 13% while Similarweb puts it near 28%, because one measures assistant usage and the other measures web visits. Do not average them. What is comparable is the direction inside each series, and all four say the same thing.

The number that actually answers the question. Between January and June 2026, ChatGPT gave up 8.5 points of share. Gemini captured none of it — it went backwards 0.7pp. Claude took 3.0 points. The remaining ~5.5 points scattered to Grok, Perplexity, DeepSeek and open weights. Gemini's billion users are a distribution artifact — Android, Chrome, Search AI Mode, Workspace — not a preference win. Google is putting Gemini in front of a billion people and converting them into flat share. That is the honest read, and it is worse than the "1 billion users!" headline and better than the "Google is collapsing" headline.

2. Why it stalled: the cadence broke before the people left

My read, and it is a negative one: the order of events matters. The flagship slipped before the departures, not after. Shipping efficiency models instead of a frontier model, while compute-constrained enough that your own DeepMind researchers are reportedly queuing behind TPU capacity you sold to Anthropic and Meta, is the signature of an organisation that chose to monetise its silicon rather than defend its model crown. That was a deliberate trade, and I think it was the correct one — but it has a cost, and the cost is showing up exactly where you would expect: consumer share.

Will it come back? My specific, falsifiable call

No meaningful consumer share re-acceleration before Q1 2027. Base case: Gemini plateaus at 25–30% of Similarweb visit share and ~20% of DAU through year-end.

What would change my mind (the falsification test): a Gemini 4-class frontier release that moves Similarweb visit share more than 5 percentage points within two months of launch. Nothing smaller counts — Gemini 3.1 and the Flash line have already proven that competent releases without a frontier step do not move share.

What would confirm the bear case: a second wave of senior DeepMind departures before the Oct 26 print, or Gemini visit share printing below 24% in any month through year-end.

3. Why it barely matters: the chatbot is a narrative asset, the TPU is a P&L asset

Here is the quantification nobody puts next to the Gemini headlines. Alphabet does not break out consumer Gemini subscription revenue; it sits inside subscriptions/platforms/devices. Be generous and call it $10B a year. Against a ~$479B annualised revenue base, that is ~2%.

Now look at what happened in the same quarter that Gemini's consumer share went flat: Google Cloud grew 82% to $24.87B at a 35.6% operating margin, and backlog hit $514B. Grok 4.6 launched on Google Cloud Vertex AI. Anthropic trains and serves on TPUs. Meta is testing them. Alphabet monetises its competitors' models. Losing the chatbot war and winning the compute war is not a contradiction — it is the strategy.

The tradeable inefficiency: on Aug 5 the market took 4–5% — roughly $200 billion — off Alphabet because researchers left the model lab. Almost none of the earnings power that has actually re-rated this stock since Q4'25 depends on those researchers. The market persistently prices Alphabet's AI as a chatbot story and Alphabet persistently earns it as an infrastructure story. Every leadership headline is therefore a buyable event until the backlog breaks.

4. The TPU strategy — and the part nobody underwrites

4.1 What is actually shipping

GenerationDetailConsequence
TPU v7 "Ironwood"4.6 petaFLOPS per chip; 42.5 exaFLOPS in a 9,216-chip superpod; shipping in volumeThe inference-era workhorse; the chip Anthropic's 1GW+ 2026 buildout runs on
TPU 8t (Cloud Next, Apr 2026)Broadcom-designed, training-optimised, TSMC 2nm, late 2027Broadcom keeps the training socket
TPU 8i (Cloud Next, Apr 2026)MediaTek-designed, inference-optimised, TSMC 2nm, late 2027; claimed ~80% better inference performance-per-dollar than Ironwood; 121 exaflops headlineBroadcom loses the inference socket — and inference volume dwarfs training volume over time
Attach silicon$120B Marvell arrangement for CXL and custom NICs around TPU inference; Intel also in the supply chainA deliberate four-vendor supply chain
Anchor demandAnthropic: up to 1M TPU chips, >1GW online in 2026, 3.5GW committed from 2027Single-counterparty concentration — see 4.3
External customers namedAnthropic, Meta, Salesforce, Midjourney, Replit (plus SSI and xAI in supply-chain reporting)Meta is the validation that matters — a direct Nvidia substitution by a peer hyperscaler

My read: the two-way split of TPU v8 is the single most important strategic disclosure Google made in 2026, and it was almost entirely covered as a chip-spec story. It is a procurement story. Google first used custom silicon to break Nvidia's pricing power; it is now using MediaTek and Marvell to break Broadcom's. A company that multi-sources both its training and its inference silicon at 2nm has structurally capped what any single vendor can charge it — forever.

4.2 The real moat is a financing spread, not a FLOP count

Per reporting on Aug 4, 2026: Alphabet has built a credit-guarantee structure covering up to $43.8 billion of lease payments if TPU tenants default. The result is that data-centre operators deploying TPUs borrow at 7.1%, while operators deploying Nvidia hardware pay 9.3% for equivalent debt. Broadcom carries a parallel backstop, absorbing roughly $30B of the initial $35B of Compute-SPV financing if Anthropic stops paying.

Do the arithmetic, because it is decisive. A 1-gigawatt AI deployment costs on the order of $35–50B. A 220 basis point financing advantage on $40B is roughly $880 million per year of interest saved. No plausible perf-per-watt difference between Ironwood and Blackwell/Rubin generates that much economic value per year on the same capital base. Google has turned its balance sheet into a chip subsidy, and the subsidy is larger than the silicon advantage.

And it ties directly to this morning's finding. Alphabet issued $25B of senior notes on Aug 10 including 6.500% due 2066 — 123bp over the 5.273% 30-year Treasury. So the structure is: Alphabet borrows 40-year money at 6.5%, and lends its credit rating to tenants so they can borrow at 7.1% to buy Alphabet's chips and rent Alphabet's cloud. That is a carry trade with a chip business bolted on, and it is brilliant right up until the counterparty stops paying.

4.3 The exposure nobody is underwriting: Alphabet is four ways long Anthropic

This is my central original finding in this piece. Alphabet's relationship with a single private, cash-burning counterparty now runs through four separate channels simultaneously:
  1. Equity holder. The Anthropic stake is a core part of the $99B of unrealized equity gains booked in Q2 — $98.0B of net other income, ~80% of reported quarterly profit.
  2. Compute supplier. Anthropic is among Google Cloud's largest customers, with up to 1M TPU chips and 3.5GW committed from 2027 — a material, undisclosed slice of the $514B backlog.
  3. Accounting beneficiary. Alphabet books GAAP income when Anthropic's private valuation rises, and that valuation is itself partly a function of the compute Alphabet finances.
  4. Credit guarantor. Up to $43.8B of contingent lease-payment protection, which does not appear inside the $98.2B of on-balance-sheet debt reported at June 30.
These four exposures are perfectly correlated and all fail on the same day. If AI private marks compress, Alphabet simultaneously (a) books a large negative other-income line, (b) sees backlog quality questioned, (c) faces guarantee calls, and (d) is doing it while free cash flow is already negative $5.9B a quarter. The $43.8B guarantee is 18% of Alphabet's $242.5B cash and ~1.0% of market cap — trivial in isolation, and not trivial when it triggers in the same week as everything else.

Consequence, and a correction to my own morning note. In Part I I framed an Anthropic IPO priced below the last round as a clean "fake miss, buy the dip." That was too clean and I am tightening it. A soft Anthropic print now has three transmission channels into Alphabet, not one. I am keeping the trade but raising the bar: I will only buy that dislocation if backlog holds and no guarantee is drawn. If a guarantee is drawn, it is not an accounting event.

What to demand on the Oct 26 call. Alphabet discloses a $514B backlog and does not disclose customer concentration within it. Every analyst on that call should ask for the top-customer percentage. Until they do, the strongest bull pillar in this story is an unaudited black box. I am treating it as a real number with an unquantified concentration discount — not as a bond.

5. Where Alphabet's AI sits against the peer group (Fri Aug 21 close)

PriceMarket capP/E (TTM)Rev growth YoY% off 52wk high
GOOGL$344.82$4.227T17.3 headline / ~26 adjusted+20.1%−15.6%
NVDA$214.72$5.423T34.0+70.7%−9.2%
MSFT$483.24$3.700T27.7+17.8%−12.7%
AVGO$368.45$2.004T68.4+32.3%−25.6%
META$549.90$1.516T22.3+27.7%−30.5%

Trailing P/E and revenue growth from the fundamentals feed. Forward P/Es from that feed are unreliable in this dataset — it reports AVGO's forward multiple above its trailing multiple, which would imply falling earnings at a company growing revenue 32%. I am using trailing only and saying so.

The honest ranking on adjusted earnings: META (22x) < GOOGL (~26x) ≈ MSFT (27.7x) < NVDA (34x) << AVGO (68x). Alphabet is not the cheapest mega-cap AI name — that was Part I's correction and it holds here. But it is cheaper than Nvidia while owning the only credible full-stack alternative to Nvidia, and it is trading at a fraction of Broadcom's multiple while being the customer that decides how much of Broadcom's TPU franchise survives.

6. The trades

[GOOGL-1] REAFFIRMED — HIGH CONVICTION LONG (updated risk)

Unchanged from Part I: buy GOOG at $341.75, not GOOGL at $344.82 (0.90% cheaper, identical economics). Entry $322–$350 GOOGL-equivalent. Target $410 (1–3 months), $460 (6–12 months). Invalidation: weekly close below $300, or on Oct 26 a backlog below $500B or cloud operating margin below 30%.

What this AI review changes: nothing in the levels, one thing in the reasoning. The bull case does not require Gemini to win the chatbot war, and I now say that explicitly. It requires (a) cloud backlog conversion and (b) the TPU financing moat holding. New named risk added: the four-way Anthropic exposure, including up to $43.8B of contingent lease guarantees that sit outside reported debt. Size accordingly — this is a full position, not an oversized one.

[GOOGL-AI-1] NEW — SPECULATIVE: LONG GOOGL / UNDERWEIGHT AVGO on TPU multi-sourcing

[GOOGL-2] UPDATED — the paper-gain reversal buy, with a harder gate

Original trigger stands: a quarter with a negative other-income line from AI equity marks (likeliest catalyst: Anthropic IPO pricing below the last private round) plus an >8% selloff → buy $280–$310, target $400.

New mandatory conditions before buying, added today: (1) cloud backlog must still be flat-to-up in that same quarter, and (2) no credit guarantee may have been drawn. If either fails, stand aside — it is a solvency-chain event, not an accounting event, and the floor is much lower than $280.

7. Bull / base / bear on the AI franchise specifically

ScenarioProbMeasurable triggersGOOGL
Bull30%Gemini 4-class frontier ships and moves visit share >5pp within two months; TPU 8i tapes out on schedule at 2nm; a second named hyperscaler-scale external TPU customer beyond Anthropic and Meta; Oct 26 backlog >$600B$410–460
Base50%Gemini share plateaus at 25–30% of visits / ~20% DAU; Flash line carries the roadmap; TPU roadmap on track; backlog $540–600B; FCF still negative$340–390
Bear20%Second wave of DeepMind departures; Gemini visit share prints below 24%; or a guarantee is drawn / Anthropic funding stumbles, hitting equity marks, backlog quality and the contingent liability at once$270–300

8. Key sources

Data limitations stated openly: chatbot share figures come from panel-based estimators with incompatible methodologies and should be read for direction, not level. Third-party API token-share rankings (e.g. OpenRouter) understate Google badly because most Gemini API volume runs through AI Studio and Vertex rather than aggregators — I have deliberately not leaned on them. Not investment advice.