If you invest in AI, build in AI, or are deciding whether to bet your next fund or your next decade on this industry — start with the most expensive question in the room: where does the money come in, whose hands does it pass through, and where does it finally settle?

Consider two facts from 2026 that should not be able to coexist.

Fact one: the fastest revenue growth in business history. According to multiple reports, Anthropic’s annualized revenue went from $9 billion to past $45 billion in half a year — roughly doubling every six weeks — and it filed for an IPO in early June at a reported valuation near $1 trillion. Per company guidance, Microsoft, Amazon, Google, and Meta plan a combined ~$725 billion in capital expenditure for 2026 — close to the GDP of a mid-sized country.

Fact two: the most dazzling companies on this value chain barely make money. The fastest-growing are burning cash, the highest-valued depend on continuous fundraising, and when the word “profitability” appears, it usually arrives with an accounting footnote.

Most AI books explain fact one — how powerful the models are. This is about the crack between the two facts.

One Company, Two Statements

Stay with Anthropic. You’ve seen the revenue line. Now the spending line. The officially confirmed part alone is staggering: in October 2025, Anthropic announced a multi-year deal for up to one million Google TPUs, officially “worth tens of billions of dollars,” bringing over a gigawatt online in 2026; in April 2026 it added a multi-gigawatt agreement with Google and Broadcom. Reported full-scope figures run larger — cloud commitments on the order of $200 billion over five years, plus a data-center deal around $15 billion a year. Add up the reported numbers and the compute it has committed to pay out is almost the same number as the total revenue it takes in.

This is not mismanagement. Per Epoch AI’s breakdown of its cost structure, training compute is 42% of total spend and inference 28% — seventy cents of every dollar flows upstream. The fastest-growing company in history hands its money, almost untouched, to its landlords: cloud providers, chipmakers, and the landlord’s landlord — power companies.

That is the “Deflation Sandwich”: intelligence is deflating; revenue concentrates in the model layer, but profit settles at the two ends.

One Ruler: Profit = Scarcity × Monopoly × Pricing Power

Why does the middle get squeezed? Measure every layer with one ruler — three factors, multiplied; lose any one and profit leaks out.

The model layer is bleeding on all three. Scarcity is bleeding — open-weight models close the capability gap month by month; a model is a reproducible engineering artifact, not a non-renewable resource. Monopoly is bleeding — for bare-API customers, switching models is a one-line config change. Pricing power is bleeding — the price of a unit of intelligence is collapsing; that is what “deflation” literally means.

The two ends light up on all three. Upstream profit comes from scarcity: EUV lithography, advanced packaging, grid interconnection — nobody conjures these up in the short run. Downstream profit comes from lock-in: habits, defaults, workflows. Open source can match the capability; it cannot match these.

So the spine of the book I’m writing is one impolite sentence: the model layer has no profit pool of its own. It is an arena — the winner’s prize is the right to become a company spanning both ends; the loser’s fate is commoditization. Watch what the leading labs are actually doing: one earns most of its revenue from subscriptions (becoming a consumer distribution company); the other is sprinting on workflow products (becoming an enterprise workflow company). The winners aren’t finding profit inside the model layer — they are moving out of it.

The Strongest Counterevidence, Face Up

Honest arguments must face their best rebuttal: per SemiAnalysis, Anthropic’s inference gross margin rose from 38% to the mid-60s in a year. If the middle is doomed, why are margins expanding?

Two cuts, then a showdown. First: that’s gross margin, not profit — training, 42% of total cost, isn’t in the numerator; on an all-in basis the company still burns cash. Second: cost-cutting comes from engineering, but keeping the savings depends on the frontier gap — while the capability umbrella holds, the savings are yours; once it closes, open source forces you to hand every cent of cost reduction to customers. The first is durable; the second is rented.

The showdown: if bare-inference all-in margins (training amortization included) keep expanding for three years under open-source pressure, the model layer is a real third profit center and my thesis is dead. That falsification condition is written into the book in black and white.

Three Bets, Each Built to Be Refuted

The book stakes three bets: ① the model layer earns revenue but cannot keep profit; ② inside this $725-billion-a-year capex cycle hides a bubble amplified by circular financing — Bloomberg tallies over $800 billion of “vendor invests in customer, customer commits to purchases” arrangements, and the bluntest specimen sits in an SEC filing: a chipmaker’s warrant to its biggest customer that vests as purchase milestones are hit — the more the customer buys, the more it earns on its supplier; ③ slow enterprise AI ROI is a permanent constraint of organizational physics, not a transition phase.

Each bet carries its falsification condition. And the book predicts no dates — books that predict dates die on those dates. It hands you a dashboard instead: the open-source-to-frontier gap in months, the slope of bare-API pricing, neocloud credit spreads, the scissors gap between chip generations and depreciation schedules. Calendars get embarrassed; causal chains don’t.

Intelligence is getting cheap, and cheap doesn’t mean everyone profits — in the age of falling electricity prices, the money went to grids and appliances, not power plants. The revenue map and the profit map have never been the same map. Knowing whose hands cheap intelligence flows through is worth more than predicting intelligence itself.

So — is the model layer the next profit center, or the next “power plant”? Which layer is your portfolio or your product betting on? Comments open.

Data sources and evidence levels (June 2026): Anthropic ~$45B annualized revenue, six-week doubling, IPO valuation (multiple media reports; replace with official figures once the S-1 is public); up to 1M Google TPUs / “tens of billions” (Anthropic official announcement, Oct 2025); cost structure 42% training + 28% inference (Epoch AI); inference margin 38%→mid-60s (SemiAnalysis, 1Q26); Big Four 2026 capex ~$725B (company guidance); circular financing arrangements >$800B (Bloomberg, 2026); warrant terms (AMD–OpenAI agreement, SEC exhibit, Oct 2025).

— From the prologue of a book in progress, working title The Deflation Sandwich: Profit Migration and the Capital Cycle of the AI Industry

#AI #AIIndustry #DeflationSandwich