The prediction first, with the confidence attached, so it can be held against me later. A correction is coming. High confidence, because the pattern has run five times in two centuries and the financing gap is real. Moderate confidence that 2027 is when it starts to feel like one. If it arrives later, nothing else here changes.
That is the least useful part of the argument.
Nearly every conversation I have about the bubble is about whether and when, which is entertaining and beside the point. The question that pays is what a correction does to the price of what you are doing now.
Carlota Perez traced the same sequence across steam, railways, electricity and information technology: financial capital builds infrastructure far ahead of demonstrated demand, a frenzy inflates the claims, a correction forces the technology to prove itself, and only then does deployment put what was built to real use. Railway mania destroyed enormous capital and left Britain a network no sober parliament would have funded, which then carried the second half of the industrial revolution. The dot-com collapse wiped out trillions and left the fibre that carried the internet economy. The people who fund the frenzy are rarely the people who profit from the deployment.
Which brings me to the part that has changed how I plan my own work.
There is a line doing the rounds that I have come to find misleading: this is the worst AI will ever be. It is probably true. It is also a claim about the frontier, not a claim about your desk.
What you have is not the frontier. It is the frontier multiplied by how much of it you can afford to run, and those are two curves rather than one. Cost per token keeps falling, which is the honest counterargument and a strong one. But the tokens a single piece of work consumes have been rising faster, because agentic workflows burn through them at rates that would have looked absurd two years ago. And the price is being held down by capital that expects a return. Whether inference at today's prices covers its own cost is an open question; the providers do not publish, so treat that as informed inference rather than fact.
My own experience, and one desk is not a dataset. Over the past year, tightening usage limits have affected my daily work more than any model improvement has. The models are better. What I can get done with them in an afternoon is not straightforwardly better, because I keep meeting the ceiling before I meet the limits of the tool. And the rationing arrives as limits, tier changes and deprecated models rather than as a price rise, which is exactly why it never reaches a procurement review. The invoice does not change. What it buys does.
So here is the question I cannot answer. Is the improvement curve going to outrun the access curve? If capability wins, that is a bonus. If access wins, an organisation whose working assumptions are built on today's provisioning has a continuity problem it has not named.
The asymmetry is why I would plan for the second. And it sharpens the timing. The cheapest period an organisation will ever get for learning with this technology is right now, while someone else's capital pays most of the bill. That means running the experiments, making the mistakes, and building judgment no vendor can sell you. Wait for the market to settle and you will be learning at post-correction prices what your competitors learned at frenzy prices.
Something to try this week. Take one piece of work that now runs on AI assistance and that you would not want to hand over badly. Do it once on the cheapest thing available to you: a smaller model, a lower tier, whatever you would fall back to if the price doubled tomorrow. Same task, same standard, one sitting.
There are three possible findings and all of them are worth the afternoon. It holds up better than you expected, in which case you have just discovered that your fallback is viable and it cost you nothing to learn. Or the quality drops and you can say precisely where, which is more useful still, because now you know what you are actually paying for.
Or you cannot tell whether it got worse. That is the most important result of the three, and it is the subject of a later piece in this series.