TL;DR — Generative AI was supposed to break Google Search. Through January 2026 the numbers said otherwise: Search revenue kept growing double digits while Gemini moved into AI Mode. The real contest is cost, and Alphabet’s distribution, custom TPUs and a milder-than-feared antitrust remedy explain why its 2026 capex is misread.
Entering 2026, the argument that generative AI would dismantle Google Search has lost most of its force. The bear case was a Kodak moment: users would trade ten blue links for chatbots, and the ad auction would shrink with them. The reported numbers have not cooperated. In the third quarter of 2025, Google Search & other revenue rose 15% to $56.6 billion, according to Alphabet’s earnings release.
Alphabet’s answer to the threat was to put its strongest model inside the product it was supposed to replace. Gemini 3 launched in November 2025 and went straight into AI Mode in Search instead of living only in a separate app. That choice re-engineered the search experience without breaking the ad model, and it is why we think the 2026 capex budget, which we were modeling above $110 billion before the February print, is the most misunderstood investment in the S&P 500.
Distribution Keeps Alphabet’s Cost of Acquiring AI Users Near Zero
Rivals are spending heavily on marketing and subsidized subscriptions to win AI users one download at a time. Alphabet does not have to. Its models ship inside surfaces people already open every day, so each new feature reaches an installed base at close to zero customer acquisition cost.
The scale is documented. On its second-quarter 2025 call, Google said AI Overviews had 2 billion monthly users, and the third-quarter release put the Gemini app above 650 million monthly active users. Those numbers came from distribution, not ad spend. The main channels are:
- Search, where AI Overviews and AI Mode sit on top of the existing query stream;
- Android and Chrome, where Gemini can be offered as the default assistant on hardware Alphabet does not have to sell;
- Workspace, where Gemini is bundled into paid business seats;
- The Gemini app, the only channel that looks like a classic standalone product.
A chatbot startup has to buy each of those users. Alphabet only has to avoid losing them, which is a far cheaper problem. We covered the revenue side of this advantage in our analysis of how Alphabet is defending its search moat.
Custom TPUs Turn AI Search Into a Cost Contest Alphabet Can Win
Answering a query with a large model costs more than serving a list of links. That makes AI search a unit-cost problem first and a product problem second. Alphabet is the only large search operator that designs its own accelerator at scale, the Tensor Processing Unit, now in its sixth (Trillium) and seventh (Ironwood) generations.
Google announced Ironwood in April 2025 and moved it to general availability in November 2025. The same quarter, Anthropic said it plans to use up to one million TPUs to train and serve Claude, which tells you outside buyers see the chip as competitive. Microsoft and Meta are building their own silicon too, but both still rely far more on Nvidia, and they pay Nvidia’s margin on every GPU.
In the AI era, the winner is not the company with the loudest chatbot. It is the one with the lowest cost of inference and the most entrenched distribution.
Owning the chip does not make inference free. It does mean every efficiency gain in model serving accrues to Alphabet’s gross margin rather than to a supplier’s. That is the lens through which the capex number should be read: as spending on a cost curve competitors cannot easily copy.
The Antitrust Remedy Protected Distribution Instead of Breaking It
Department of Justice scrutiny is real, but its impact has been mispriced in both directions. On September 2, 2025, Judge Amit Mehta declined to order a divestiture of Chrome or Android, calling a forced Chrome sale messy and highly risky. Structural remedies are off the table for now.
The behavioral remedies are narrower than many feared. Google can no longer sign exclusive default deals and must share some search index and interaction data with qualified rivals. It can still pay partners such as Apple for placement, which means traffic acquisition costs stay on the books ($14.9 billion in Q3 2025) but so does the distribution they buy. Our antitrust paradox analysis walks through why that outcome reads more like a legal floor than a ceiling.
Update: The February Print Confirmed Search Growth and a Far Bigger Capex Bill
Update, October 2026. Alphabet’s fourth-quarter release on February 4, 2026 reported Search & other revenue of $63.1 billion, up 17%, and Sundar Pichai described AI as driving an expansionary moment for Search. He also said Google cut Gemini serving unit costs by 78% over 2025, which is the cost-curve argument above expressed as a single figure.
Our capex estimate proved far too low. Management guided 2026 capital expenditures to $175 billion to $185 billion, well above Street estimates according to Bloomberg, then raised the range to $195 billion to $205 billion in July. Search kept pace: second-quarter 2026 Search & other revenue rose 17% to $63.3 billion, and the Gemini app reached 950 million monthly active users. The thesis held, but the price of holding it went up.
What to watch next: the trend in Search & other growth as AI Mode expands, any disclosure on AI Mode monetization relative to classic search, the gap between capex and operating cash flow, and the progress of appeals in the search case. If Search growth slows while capex keeps rising, the cost-curve argument has to be re-examined.
Sources: Alphabet quarterly earnings releases (Q3 2025, Q4 2025) and annual and quarterly filings on SEC EDGAR; earnings materials and call remarks at Alphabet Investor Relations; capex reaction via Bloomberg. Disclosure: the author holds a long position in Alphabet (GOOGL/GOOG). This article is analysis, not investment advice. Full disclosure is on our About page.
