If you’re valuing an AI company — or holding an AI research report and a buy button — try one exercise first: follow a single dollar of AI spending from the person who pays it to the people who keep it. The trip will change how you see the entire industry.

Walk it through: a business or consumer pays a model company (subscriptions, API, agents) → the model company hands 60–70% of its costs to cloud providers (training + inference compute) → cloud providers pour capex into chipmakers and data centers → chipmakers and data centers pay foundries, HBM, and electricity.

How much each hop keeps depends on what that hop holds. Per Epoch AI’s breakdown of Anthropic’s cost structure: training compute is 42% of total spend, inference 28% — seventy cents of every dollar flows upstream. The fastest-growing company in history hands its money, almost untouched, to cloud providers, chipmakers, and the landlord’s landlord — power companies.

Why can’t the model hop keep the money? One ruler: Profit = Scarcity × Monopoly × Pricing Power. Three factors, multiplied; lose any one and profit leaks out.

The model layer is bleeding on all three. Scarcity — open-weight models close the capability gap month by month; a model is a reproducible engineering artifact. Monopoly — for bare-API users, switching is a one-line config change. Pricing power — the unit price of intelligence is collapsing. Multiply the three and the conclusion is structural: bare model capability cannot collect rent in the long run.

At the two ends, the formula flips. Upstream scarcity is physical (EUV lithography, advanced packaging, grid interconnection), and the monopoly is real (one company holds roughly 90% of the training-chip market, by widely cited industry estimates). Downstream lock-in is behavioral (workflow embedding, user habits). The two ends each own the formula; the middle owns nothing.

The Hardest Counterevidence, Face Up

Someone will throw this at you: per SemiAnalysis, Anthropic’s inference gross margin rose from 38% to the mid-60s in a year. Margins are expanding — how does “can’t keep the money” survive?

Two cuts, then a showdown.

Cut one: that’s gross margin, not profit. The 38%→mid-60s number measures inference revenue minus inference compute. Training — 42% of total cost — isn’t in the numerator. And the shorter a model’s shelf life, the more training becomes a recurring cost, not a one-off. On an all-in basis, the company still burns cash. The margin improvement is real. So is the absence of profit.

Cut two: cost-cutting comes from engineering; keeping it depends on the frontier gap. Batching, distillation, custom silicon — real engineering wins. But margin also depends on the price line, and the price line leans entirely on the capability umbrella. While the umbrella holds, the savings are yours; once it closes, open source forces you to hand every cent 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 profit center and my thesis dies — that falsification condition is written into the book. You don’t have to believe me. Just watch that one number.

The dollar’s journey ends with one sentence: the revenue map and the profit map are two different maps. The most dazzling revenue numbers in AI sit in the labs; the most solid profits sit with whoever sells chips, sells electricity, and collects subscriptions. The first lesson of valuing an AI company isn’t predicting how smart models get — it’s knowing which map you’re looking at.

Open your portfolio: are your AI positions plotted on the revenue map, or the profit map? Comments open.

Data sources (verified, June 2026): Anthropic cost structure 42% training + 28% inference (Epoch AI); inference margin 38%→mid-60s (SemiAnalysis, 1Q26); ~90% training-chip share (widely cited industry estimates, pending primary anchoring).

— From Chapter 1 of a book in progress, working title The Deflation Sandwich

#AI #AIIndustry #DeflationSandwich