Part I: Foundations
Chapter 1
Why AI Is Not Just Software
On Wednesday 30 November 2022 OpenAI put a chat interface on one of its language models and released it as a "research preview", free to anyone who signed up. The model was not new. It was a version of GPT-3.5 fine-tuned for dialogue with reinforcement learning from human feedback, the technique OpenAI had described in its InstructGPT paper in January 2022, and developers could already buy similar capability through the company's paid API. The team, led by John Schulman, had built the interface in a few weeks. OpenAI had about 375 staff in early 2023 (reported) and its revenue in 2022 was about $28 million (reported, verify), mostly from the API.
On 5 December Greg Brockman, OpenAI's president, said ChatGPT had passed one million users. Through December its servers were often at capacity, and a waitlist and a "ChatGPT is at capacity" page became familiar to the people trying to use it. Sam Altman, the chief executive, described the cost of running it as "eye-watering" and put the compute cost at "probably single-digit cents per chat". By January 2023 the product was estimated to have about 100 million monthly users (UBS, citing Similarweb), the fastest growth of any consumer application on record by the estimates of the time. OpenAI launched ChatGPT Plus at $20 a month on 1 February 2023, and Microsoft committed a further ~$10 billion to the company in January (reported), much of it in the form of the Azure computing that every answer consumed.
The pattern set in those weeks has not changed. By early 2026 OpenAI claimed about 900 million weekly users and more than 50 million paying subscribers. Its revenue rose from ~$3.7 billion in 2024 to ~$13.1 billion in 2025 (reported), but it reported an operating loss of about $5 billion on 2024's revenue (reported, verify), and by November 2025 Altman said it had committed to about $1.4 trillion of compute over about eight years (company claim), a target reported in February 2026 to have been reset to about $600 billion by 2030. In March 2026 investors valued it at $852 billion. A free product made from an existing model had become the most valuable private technology company in the world, and also one of the most expensive to run.
The preview ends here. The full chapter is in the book.
In this chapter
- 1.1Four problems that ordinary software does not have
- 1.2A stack of businesses
- 1.3How AI differs from ordinary software
- 1.4The free-user bill
- 1.5Where the money has gone
- 1.6The intelligence equation