If you take one chart away from this book, take this one. For allocators: how deep each station’s profit pool runs, what its quality is, and where it migrates — in one read.

Two rules before drawing.

Rule one: map profit, not revenue. Revenue tells you where money flows; profit tells you where it stays — and in AI, the two maps diverge shockingly: the brightest spot on the revenue map (the leading labs) is dark on the profit map.

Rule two: grade every pool’s quality. The same dollar of profit comes in three grades: real (backed by end-user payment), circular (backed by loop financing — orders bought with the supplier’s own investment), and borrowed (backed by accounting policy — depreciation schedules slower than actual asset decay, profit borrowed from the future).

The 2026 Map: A Pyramid With Its Peak Upstream

Walking from source to user, station by station:

Chip design: the deepest pool (roughly 90% share, ~70% gross margin, by widely cited industry estimates), but the order book carries substantial echo demand — customers it funded itself. Deep; ~60% real, ~40% circular.

Foundry & HBM: foundries and memory makers don’t play the circular game — they collect cash and allocate capacity. Deep, almost entirely real.

Power & data-center physical layer: this round’s money isn’t only becoming chips — it’s becoming concrete, copper, and transformers. Order books stretch years out, yet the segment carries almost no “AI valuation premium” — the market still prices it like traditional industrials. Deepening, all real, and priced late.

Cloud providers: AI revenue rising, healthy margin structure — but every dollar of profit immediately exits as larger capex, on generous depreciation schedules. Large; real with a borrowed component.

Model layer: the revenue peak and the profit depression — seventy cents of every dollar flows upstream; all-in, cash-burning. Shallow (or negative), with echo in the revenue.

Applications & distribution: subscriptions, seats, outcome-based pricing — all end-user money, not a cent of echo. Shallow, all real, deepening.

Three-Year Migration: Four Arrows

Arrow one: from silicon to watts. Chip design’s monopoly has three erosion mechanisms (customer-built silicon, a maturing second supplier, workloads shifting to cheaper dedicated chips); power scarcity has none — plants and interconnection take five to ten years, and there is no “open-source electricity.”

Arrow two: from the model layer to workflow and distribution. Revenue will keep setting records in the model layer, but the retainable part increasingly comes from “two-ends businesses.” Three years from now, “model company” may no longer be a category — the survivors will have changed names.

Arrow three: the forced conversion from circular to real. The scissors and the loop won’t hang forever — through write-downs, defaults, or orderly wind-down, the circular part gets cleared. A slice of today’s upstream profit will be handed back; downstream’s real profit won’t.

Arrow four: exogenous repricing. The power gap deepens arrow one; geopolitics splits compute into two pricing regimes; and if regulators mandate interoperability, distribution rent itself gets shaved.

The map folded into one sentence:

Today: profit is deep upstream and real downstream. In three years: depth migrates from silicon to watts, and realness compounds from subscriptions into rent; the middle completes its decay from “industry” to “arena”; and the circular part gets collected, all at once, in some future quarter.

Three takeaways. Upstream, owning watts beats owning silicon — the erosion mechanisms are asymmetric. Apply a circularity discount to every upstream valuation — the more financing-dependent the counterparties, the bigger the discount. And the best risk-adjusted seat on the map is the late-priced corner: power equipment and physical-layer engineering.

Which station is your portfolio sitting on? And is that station’s profit real, circular, or borrowed? Comments open.

Data sources (verified, June 2026): model-layer cost structure (Epoch AI); chip share and margins (widely cited industry estimates); circular financing >$800B (Bloomberg); depreciation policies (company filings; to be anchored to 10-Ks before publication).

— From Chapter 4 of a book in progress, working title The Deflation Sandwich

#AI #ProfitPoolMap #AIIndustry