Preview

Prologue · Three Counterintuitive Bets

This book answers two questions

The AI industry in 2026 contains two facts that should not comfortably coexist.

The first is the fastest revenue growth in corporate history. Leading model companies are expanding annualized revenue at a pace that, in reported cases, looks like a doubling every six weeks. The four largest technology platforms are on track to spend roughly $725 billion in capital expenditure in a single year, up about 77% from 2025 and near the annual GDP of a mid-sized country.

The second fact is more uncomfortable: the most visible firms in the AI value chain mostly do not earn much profit. The fastest-growing companies burn cash. The most highly valued firms need continuous financing. Even when the word “profit” appears, it often arrives with a footnote about accounting treatment.

Most books about AI explain the first fact: intelligence is getting stronger, faster, and cheaper. This book is about the gap between the two facts. Where does the money enter? Whose hands does it pass through? Where does it finally settle? For investors and founders, the real questions are simple:

Which layer should I bet on? And is now the right time?

The spine of the argument

The entire book rests on one claim:

Intelligence is deflating. The model layer is not an independent profit pool. It is an arena. Winners earn the right to become companies at the two ends of the stack; losers are commoditized.

Upstream profit comes from scarcity: lithography tools, advanced packaging, grid-connected power, and capacity that cannot be conjured quickly. Downstream profit comes from lock-in: user habits, defaults, workflow ownership, memory, permissions, switching costs, and trust. Open source can chase model capability. It has a much harder time chasing a deeply embedded workflow or a regulated distribution channel.

The model layer sits between those two ends. It can capture extraordinary revenue, but it leaks profit toward both sides.

The measuring stick used throughout this book is:

Profit = Scarcity x Monopolizability x Pricing Power.

If any factor is missing, durable profit becomes hard to keep. Each chapter applies that single test to one layer of the industry and asks whether the layer is worth owning.

Three counterintuitive bets

An honest thesis should be breakable. This book makes three bets and states what would prove each wrong.

Bet 1: The model layer will capture revenue but fail to retain profit.
Profit migrates upstream to compute and power, and downstream to distribution and workflow lock-in. Revenue concentration in leading labs does not contradict this. It is part of the mechanism. Look at the mix: the money increasingly comes from subscriptions, products, and workflows, not from naked model APIs. The survivors stop being pure model companies. OpenAI is moving toward consumer distribution. Anthropic is moving toward enterprise workflow. The prize for winning the arena is the right to become something else.

Referee variable: the all-in gross margin of naked inference APIs, including training amortization. If by 2029 this margin is positive for four consecutive quarters and above its 2026 level, the sandwich thesis is wrong. This test does not add an after-the-fact condition about whether open-source pressure was “strong enough.” The number itself is the judge.

This is the most important gauge in the book because it has a deadline. The claim cannot hide forever behind “the trend is still forming.” If the naked API can carry training amortization, serving cost, and frontier refresh cost while expanding all-in margin, then the model layer is not merely an arena. It is a real third profit pool. If the number remains negative, or improves only at the inference-gross-margin line while all-in economics stay weak, the sandwich thesis stands. The point is not to win the argument by definition. The point is to make the argument falsifiable.

Bet 2: This capital spending cycle contains a bubble amplified by accounting and financing structure.
Demand is real. But part of demand is an echo created by capital itself. More than $800 billion in circular financing arrangements make suppliers, customers, and investors overlap. The real profit pool is smaller than consensus believes, and it is more upstream than consensus wants it to be.

Falsification: if leading buyers generate positive returns on AI capex within the depreciation cycle, and circular arrangements are digested without disorder, the bubble argument fails.

Bet 3: Slow enterprise ROI is mainly an organizational constraint, not a temporary technology gap.
Inference demand will not come from universal enterprise adoption. It will come from the minority of workflows where ROI stands up quickly: coding first, then support, legal, finance, and other narrow domains. This explains why profit moves toward workflow lock-in. Demand itself lives inside workflows, not inside generic APIs.

Falsification: if agents produce repeatable returns across large enterprise processes without organizational redesign, organizational transformation cost is no longer the main constraint.

A dashboard, not a calendar

Books that predict dates usually die by dates. This book does not try to say when open source catches the frontier, when the bubble breaks, or when agents cross a magic threshold. Calendars embarrass analysts. Causal chains are more useful.

The alternative is a dashboard. Each chapter identifies observable leading indicators. When they turn red, the window for that layer is closing. Six gauges matter most:

  • the number of months between open-source and frontier capability
  • the slope of naked API pricing
  • agent portability across tools, context, memory, permissions, and workflows
  • credit spreads of circular-financing counterparties
  • the gap between chip replacement cycles and accounting depreciation
  • the length of grid interconnection queues for data centers

Readers do not need to accept the book’s conclusions. They need to learn how to read the gauges. A useful dashboard must be able to prove the author wrong. If the open-source/frontier gap remains wide enough to sustain frontier pricing, if all-in naked API margins turn positive and expand through 2029, or if enterprise conversion rates steepen sustainably, the core thesis must be revised. A dashboard that can only confirm a thesis is not analysis. It is fortune-telling.

One gauge watches the book itself. If agent portability turns fully mature and downstream switching costs collapse toward zero, distribution rents also collapse. At that point even “becoming an end company” may not save the winners. The spine of this book would break.

The same discipline applies in the other direction. A single weak quarter, a single spectacular open-source release, or a single failed financing deal is not enough to declare victory. Each gauge must be read against the mechanism it is supposed to test. Open-source capability tests the price umbrella. API margin tests the model layer’s right to be called a profit pool. Neocloud credit tests whether circular demand is being refinanced or marked down. Grid queues test the duration of upstream scarcity. Portability tests downstream lock-in. Mixing these gauges together after the fact would be another way of refusing to be falsified.

Only the invariants matter

This book does not try to rank today’s leading model or today’s most visible company. Those facts expire in months. It focuses on four invariants: physics (energy, latency, speed of light), capital structure (who pays, who depreciates, who carries the liability), incentives (where profit flows, behavior follows), and power (who can block whom).

Companies and models appear as examples. The argument rests on invariants.

The book has four parts. Part I follows the money: capital cycles, circular financing, and the profit map. Part II walks the stack layer by layer. Part III studies the three forces that set prices: physics, politics, and law. Part IV turns the map into two playbooks, one for investors and one for founders.

Intelligence is becoming cheap at a measurable speed. Intelligence equivalent to GPT-3.5 fell from roughly $20 per million tokens in late 2022 to about seven cents by late 2024, a decline of around 280x, according to the Stanford AI Index. This book’s wager is that understanding whose hands cheap intelligence passes through is more valuable than trying to forecast intelligence itself.

The reader should also notice what this book refuses to do. It does not rank models. A model ranking expires too quickly. It does not crown a single winner. The industry is too layered for that. It does not argue that AI is a bubble in the loose sense that “nothing is real.” Much of the demand is real, and the technology is real. The book argues something narrower and more useful: real technology can still produce misallocated capital, and real demand can still fail to arrive fast enough for the first capital stack.

The method is to separate five things that public discourse often merges.

First, capability is not the same as scarcity. A model can be astonishing and still hard to price if substitutes arrive quickly. Second, revenue is not the same as profit. Money can pass through the most visible layer and settle elsewhere. Third, adoption is not the same as ROI. Employees can use AI widely while the enterprise’s financial statements barely move. Fourth, lock-in is not the same as habit. A habit can be broken by a default change; deep workflow state is harder to move. Fifth, infrastructure value is not the same as investor return. A society can benefit from overbuilt infrastructure after the first owners are written down.

These distinctions are the book’s operating system. If the reader keeps them separate, the rest of the argument becomes easier to audit. If they are collapsed, almost every AI debate becomes circular. Bulls point to capability and revenue. Bears point to cash burn and overbuilding. Both can be right at the same time because they are describing different layers and different time horizons.

The book also treats falsification as a practical discipline, not a philosophical ornament. Every serious claim should have a gauge. The gauge does not need to be perfect. It needs to be observable enough to force revision. A thesis without gauges becomes a mood. A dashboard with only confirming signals becomes marketing. This matters because AI will produce contradictory evidence for years: better models and worse margins, higher usage and slow ROI, more capex and real bottlenecks, more open source and more concentration. Without explicit gauges, every side can explain away every inconvenient fact.

Finally, the book is written for decision-makers, not spectators. Investors need to know which layer deserves duration and which layer deserves only tactical exposure. Founders need to know where to build before the window closes. Operators need to know why pilots do not become profit automatically. The argument is therefore less interested in whether AI is “good” or “bad” than in where value can stay after the excitement is translated into cost, liability, power, workflow, and distribution.