Follow the Twenty Quid: The AI Bill That Doesn't Add Up
No prompts, no agent architecture, no tooling demo. This is a talk about money. Twenty quid a month. That's what an AI subscription costs, near enough, whether it's ChatGPT, Claude or Cursor. Tens of thousands of pounds for a single chip. Warehouses of them, drawing power like a small city. Somebody is covering the difference, and it isn't me. I went looking for who, and got it wrong twice. Maybe nobody's covering anything, and this simply isn't expensive to run. Except power is only about 1% of what an hour of GPU time costs. The money goes on the chip itself. So the gap is real. Then maybe they're taking the loss on me. Except OpenAI's leaked, FT-verified accounts show serving turns a profit, and the twenty-billion-dollar hole is research. The answer is other users. The 95% who pay nothing, and more surprisingly the quiet payers who barely touch it. The expensive ones are the people in this room running coding agents overnight, and three companies have now said so publicly. That should have been the end of it. Except the labs spent twenty billion a year themselves, so somebody funds them too. Keep pulling and it stops being about software: thirty-eight billion dollars per gigawatt of datacentre, companies buying repurposed jet engines because grid connections take years, and SoftBank selling its entire Nvidia stake to fund a company that hands the money straight back to Nvidia. We have seen most of this film. Netflix burned cash for a decade at £5.99, turned profitable, and only then brought in ads, a password crackdown, and a £12.99 price tag. All three have now appeared in AI, the last of them in April. You'll leave knowing which of the five very different layers of this industry actually makes money, whether you personally are one of the expensive users, and one question worth asking about whatever you're building: what happen if the price of a token doubled tomorrow? Nothing to sell here. Every figure is sourced.
Yuan Zhang
To be update