The position. The only company holding all six layers of the AI stack — chips, models, tools, distribution, an always-on agent, and media — sitting on ~$129B of real, audited operating profit.
The one honest crack. AI is almost pure upside here, with a single genuine exception: AI answers could erode the ~$234B ad engine that funds everything. Watch it, don’t dismiss it.
My position. The safest long in AI, and the cheapest place to run high-volume work — because Google owns the factory the others rent.
When Buffett buys a gold rush
Warren Buffett famously avoids technology and avoids hype. So when his Berkshire Hathaway took $10 billion of Alphabet’s $80 billion equity raise in June — the largest raise ever by a profitable tech company, and at a negotiated below-market discount, the same structure Berkshire used to bless Goldman Sachs in 2008 — it wasn’t a trade. It was a verdict (Reuters). Buffett doesn’t chase gold rushes; he buys the company selling the shovels and owning the land. His money said the quiet part out loud: AI infrastructure has graduated from venture speculation to a toll-road value investment — the kind a pension fund, an insurer, or a sovereign wealth fund can finally justify on its mandate. That one placement reaches past Alphabet and reprices how the entire sector gets financed.
AI infrastructure just graduated from venture speculation to a toll-road value investment.
The number you can bank
Across this series, the discipline has been to read every number by the evidence behind it — because most frontier “facts” are run-rates and projections. Google is the exception. This is the one file where the big numbers are audited, pulled straight from Alphabet’s filings. So the grammar barely matters here — but read it anyway, because the green is the point.
The number, and what stands behind it
Notice the honesty the badges force. Google does not report an “AI revenue” line; its big AI-adjacent number is Cloud — audited, banked, checked by accountants — and its profit is real money, not a forecast. That’s the mirror image of the famous labs, whose loudest numbers are unaudited run-rates. When Anthropic’s or OpenAI’s number is big, it’s a claim awaiting an auditor. When Google’s number is big, it’s already in the bank. The whole company sits on the firm side of every badge in this series.
$20.0B in Q1’26 (+63% YoY, ~$80B run-rate). This is Cloud — not an “AI revenue” line.
$129.0B in FY2025 — real, booked, checked by accountants. $39.7B operating income in Q1’26 alone.
3.2 quadrillion — up from 9.7 trillion in 2024 (~330× in two years), most of it on Google’s own TPUs.
$80B equity ($10B to Berkshire) — the largest ever by a profitable tech company. Stock +118% in a year.
Cloud revenue and operating income, Alphabet’s FY2025 10-K and Q1 2026 results (audited); the $80B raise and Berkshire placement, Reuters; token volume, Google I/O 2026.
Defend the toll road
Google’s AI is built by DeepMind and distributed across surfaces no rival can match — Search at ~8.5 billion queries a day, a ~2.5-billion-user Workspace, ~3 billion Android devices, Chrome. The strategic core is simple and brutal: Google profits whether its own model wins or a competitor’s does. It processed 3.2 quadrillion tokens a month as of I/O 2026 — roughly 42× OpenAI’s estimated volume — and ran most of it on its own TPUs, not rented NVIDIA chips.
The growth curve is the whole story. On a linear scale, 2024 is barely a sliver against where the meter sits today.
From a sliver to a quadrillion
This is the toll booth. Two years ago Google counted tokens in the trillions; today it counts them in the quadrillions — a ~330× climb. Drawn to scale, the 2024 bar all but disappears. Volume like this is only economic if you own the road it runs on — which is exactly why the next section matters.
Token volume disclosed at Google I/O 2026. The ~42× comparison vs. OpenAI is a derived estimate, not a Google disclosure.
That’s why Google can price Gemini Flash and Flash-Lite at the bottom of the market ($0.10/$0.40 per million tokens for Flash-Lite) and still run healthy margins — the flywheel of more tokens → lower cost → undercut everyone → more tokens. I won’t hang a precise dollar on the savings (the “$108B avoided” figures circulating are derived models, not disclosures), but the structural logic is unarguable: own the factory and you keep a margin no NVIDIA-renting rival can match. And it has begun renting the factory to the competition: the ~$200B Anthropic TPU deal turns a direct rival into Google Cloud’s largest single customer from 2027, adding a reported $10–15B a year in incremental revenue. Google competes at the product layer and profits at the infrastructure layer at the same time. For Google, the model is insurance on a $300B+ franchise, not a bet-the-company gamble.
The only complete stack
What makes Google singular isn’t any one product; it’s that it’s the only company holding all six layers of the AI stack at once — while OpenAI holds four and Anthropic two. Google owns the whole column, silicon at the bottom, media at the top.
Six layers. One owner.
Silicon → Models → Tools → Distribution → Ambient agent → Media — all owned.
The piece rivals can’t copy is Gemini Spark, the ambient layer unveiled at I/O 2026: a personal agent that runs 24/7 on dedicated Cloud VMs — no prompt required, no laptop kept open — with native access to twenty years of your Gmail, Calendar, Maps, and Search history, plus 30-plus third-party apps wired in through MCP (TechCrunch). Spark runs while your laptop is shut; ChatGPT and Claude wake only when you ask. No competitor can replicate that data depth without Google’s two-decade incumbency in your inbox.
And underneath the products, Google is quietly winning a layer nobody was contesting: the compliance standard. Its SynthID watermark has been adopted by Microsoft, OpenAI, Kakao, and ElevenLabs, and is becoming the default the EU AI Act’s content-labeling rules will effectively require. When even OpenAI and Microsoft authenticate their AI content through Google’s infrastructure, Google is becoming the HTTPS of AI provenance — the standard everyone routes through, whoever’s model made the content.
Three moves, one direction — a competitor turning into a customer, a watermark turning into a default.
Most of those 3.2 quadrillion tokens run on Google’s own TPUs, not rented NVIDIA. That’s why Gemini Flash can sit at the bottom of the market and still earn margin no renting rival can match.
TPU economics · price-to-floorThe ~$200B Anthropic TPU deal turns a direct competitor into Google’s largest cloud customer. Google profits from rivals’ growth at the infrastructure layer even as it competes at the product layer.
~$200B compute dealGoogle’s SynthID watermark — adopted by Microsoft, OpenAI, Kakao, and ElevenLabs — is becoming the default the EU AI Act will effectively require. The HTTPS of AI provenance: the landlord is also writing the rulebook.
Industry watermark standardThe most integrated stack in AI still couldn’t ship on the CEO’s date.
At I/O 2026 Sundar Pichai said Gemini 3.5 Pro was “coming next month.” June passed. July 7 passed. Then Google DeepMind did something unusual: it scrapped the original base model and restarted from a deeper pre-training foundation, pushing the target to July 17 — a date that has also slipped by the time of writing (The Information).
Read this as evidence, not a stumble. The company with four custom-silicon partners, Ironwood TPUs shipping in the millions, Broadcom’s Sunfish and MediaTek’s Zebrafish on TSMC’s 2nm process for late 2027, and end-to-end ownership of the surfaces where the model ships — still cannot make frontier capability arrive on the date its CEO announced. If this stack can’t make model timing predictable, no stack can. That is the thesis in evidence form: the moat of 2026 is not a smarter model.
The same recent stretch produced the cleanest illustration of that thesis. In June, Google agreed to pay SpaceX $920 million a month from October 2026 through mid-2029 for roughly 110,000 NVIDIA GPUs — described as “bridge capacity” for Gemini Enterprise demand (Reuters). Google sells TPUs to Anthropic. Google rents NVIDIA GPUs from SpaceX. The company with the most advanced custom chip program in the world is simultaneously a buyer and a seller in the same compute market — because frontier demand is growing faster than any internal build program can match. Vertical integration is not self-sufficiency; it is a hedge against the year the auction goes against you.
The one crack
On an otherwise green page, there is exactly one amber box — and it’s self-inflicted and structural.
Search-ad cannibalization
Google’s ~$234B advertising engine depends on people clicking links; AI answers hand them the answer without the click. I won’t dress one quarter up as proof — but the early signal is now visible enough to name: Q1 2026 revenue of $109.7B came in below Q4 2025’s $113.5B, a sequential dip that coincides with AI Mode scaling across Search.
The counter-evidence is real too — Alphabet’s stock rose ~7% after those same earnings, AI-mode ad units carry materially higher CPMs than blue links, and $100–200/month Gemini subscriptions monetize power users at ARPUs ordinary Search never reached. So the question is genuinely open: if AI answers erode click-through faster than higher-CPM AI ads and subscriptions can replace it, the engine that funds all of this faces a real headwind. The Q1 dip is a warning, not a verdict — the one thing on this page I’d watch through Q2 and Q3.
The tests the next two quarters will actually run.
The green page turns amber if Q2 2026 earnings (July 22) show Search revenue growth decelerating below 15% YoY as AI Overviews expand to more query types — the first hard signal that cannibalization is beginning; if Cloud growth compresses toward Azure and AWS pace once initial AI contract wins lap; if the Gemini 3.5 Pro delay proves to be a capability problem rather than a schedule problem, letting model quality converge to the pack despite the infrastructure edge; if Ironwood and the 2nm TPUs deliver at scale but fail to produce a durable inference-cost advantage over NVIDIA in production; if the SpaceX bridge becomes a recurring dependency rather than a one-time gap-filler; or if the Antigravity / Spark agent layer stalls at “retention feature” and never becomes a monetization surface. The tell is which of these breaks first.
The Flash fleet and the cost pivot
The most telling thing Google shipped in July was not a bigger model. It was a fleet of small ones: additional Flash-tier models tuned for specific jobs, including a Flash Cyber variant aimed at security work, sold on speed and cost rather than leaderboard position.
For a company that owns its own chips, this is the highest-leverage move available. Every rival buying capacity pays a margin to a landlord; Google pays itself. Pushing volume onto small models it runs on TPUs it owns widens the gap between what a query costs Google and what the same query costs an Anthropic or an OpenAI renting the same class of compute.
It also confirms the pivot this series has been tracking all month. The competitive question moved from “whose model is smartest” to “what does one unit of useful work cost.” Google is answering that question with vertical integration, Microsoft with in-house substitution, DeepSeek with price, Anthropic with a cheaper flagship. Same question, four answers — and Google’s is the only one that does not depend on someone else’s balance sheet.
My read — the structural winner, with one thing to watch
If you made me pick the single highest-conviction long in this set, it’s Google, and it isn’t close. The $80B raise plus Berkshire’s underwrite de-risk the infrastructure thesis permanently and pull a new universe of institutional capital into the sector behind them; the TPU cost edge widens every year; Cloud is outgrowing AWS; the Anthropic deal converts a rival into the biggest customer; Spark builds the deepest personal-AI moat in the field; SynthID is quietly becoming the regulatory default. It is the only company that profits whether or not it wins the model war — which is exactly why its profits are the realest in this whole story. The one discipline I’d hold is to keep the amber box in view and not let the green page lull me: watch Search monetization through the back half of 2026.
It’s the only company that profits whether or not it wins the model war — which is exactly why its profits are the realest in this whole story.
If I’m building: Gemini Flash and Flash-Lite are my default for cost-sensitive, high-volume work — the owned-TPU economics land straight in the price. And if I’m choosing where to run infrastructure for the next decade, the company Buffett just underwrote is the boring, correct answer.
Sources: the $80B raise and the Berkshire placement structure, Reuters and Bloomberg; Cloud revenue and operating income, Alphabet’s FY2025 10-K and Q1 2026 results (audited); token volume, Google I/O 2026; Gemini Spark, TechCrunch; SynthID adoption, Google and CNBC; the Anthropic TPU deal, Anthropic. The honest frame: Google’s are the firmest numbers in this series — but “Cloud” is not “AI revenue,” and the Search-cannibalization risk is real, early, and not yet a trend.
This deep dive is part of “Deep Research on Frontier Companies.” Crafted with intent by Ravi Teja Palanki · June 2026.
Read the rest of the system
Google is one of seven. The synthesis reads all of them together; each deep dive establishes the verified facts for one.