The Intelligence Equation

Artificial intelligence became an industry in public in four years. On 30 November 2022 OpenAI released ChatGPT as a free research preview; by early 2026 it claimed about 900 million weekly users. Nvidia, which sells most of the chips the models run on, became the first company worth $4 trillion in July 2025 and $5 trillion that October. The four largest US cloud and platform companies spent about $400 billion of capital in 2025, mostly on data centres, chips and power, and guided to more than $700 billion for 2026. AI start-ups took about half of the world's venture funding in 2025 (reported). Some of this spending has created more value for users, faster, than any technology before it. Much of it will be lost, and in AI destruction can arrive as suddenly as success: Nvidia lost about $589 billion of market value in a single day in January 2025 when a small Chinese lab showed what a cheaper model could do, a London start-up valued at ~$1.5 billion entered insolvency less than three months after its founder stepped down in 2025, and a device that had raised ~$230 million was switched off less than a year after its first reviews.

The difference is rarely the quality of the research. This book argues that an AI company creates durable value only when four things are true at once, and that the factors multiply rather than add: if any is zero, the result is zero. It must have capability that survives real work, a system that does a job reliably enough, on evidence a buyer can check, that someone changes how they work and pays for it. It must have unit economics that hold at real usage, so that each token, query, seat or resolved task earns more than the compute it consumes when real customers, not pilots, use it. It must hold a position the next layer cannot route around, something scarce that keeps its supplier from taking the margin and its customer from switching. And it must have the capital and permission to last the cycle, the money to fund compute, chips and power, and the legal and political licence to operate, long enough to get paid. Around those four factors sit two forces that no company controls, the capital cycle and the rules, from copyright rulings and AI laws to export controls, which behave like weather: they do not decide whether a product works or earns its keep, but they decide how long a company can survive if it does not. Most of the failures in this book come from treating one factor as proven because another was, and most of the successes from a company that tested all four before its competitors did.

The twenty-five chapters develop that view. Part I explains why AI is not just software, how it became a business through seventy years of booms and winters, how the money from one query flows down the stack to the chip and the power plant, and why the evidence for a model's capability is so often weaker than it looks. Part II follows the markets one by one: chips, clouds and data centres, model labs, open models, tools and data, applications and consumer AI. Part III examines what it takes to build products that work, from inference economics and pricing to distribution and platform risk, data and copyright, agents and safety. Part IV covers capital and the AI cycle, acqui-hires and Big Tech, regulation, the geopolitics of chips and AI around the world. Part V turns to people: talent and research culture, and how to build an AI company. Part VI reads five cases for the decisions behind them. Each chapter opens with a case, explains the mechanics it illustrates, and closes by naming the factor of the equation on which its case turned.