The AI capex bubble, and what I am building while it lasts

As of this writing, the AI build-out is priced in trillions. Four hyperscalers alone plan roughly 725 billion dollars of capex (capital expenditure) for 2026; that equates to almost two billion dollars of concrete, silicon, and cooling per day, about 1.4 million dollars a minute, around the clock. Supplying the accelerators for that build-out has made Nvidia the most valuable company in history, first past five trillion dollars and peaking at 5.5 trillion this May, with other infrastructure firms trailing it at hundreds of billions each. The promise financing all of it is a step change in what a working human can produce.

Now set the measured reality next to the price tag, because the research I keep returning to points the other way. METR ran a randomized controlled trial with 16 experienced open-source developers across 246 real tasks: they believed AI made them 20% faster and were measured 19% slower, respectively the perception and the stopwatch. In Denmark, Humlum and Vestergaard linked chatbot adoption across 25,000 workers and 7,000 workplaces to the national earnings registry and found precisely estimated zeros: no detectable effect on earnings or hours, with self-reported time savings of 2.8% of work hours. The MIT Media Lab’s EEG study of essay writers adds the uncomfortable part: the LLM (large language model) group showed the weakest brain connectivity of the three conditions, and the authors coined “cognitive debt” for what accumulates when the tool does the thinking. The blunt reading is that average use is not making the average user sharper. Trillions sit on one side of the scale; 2.8% and a slower stopwatch sit on the other, and that distance between price and measured output is, in my reading, what a bubble is.

What a bubble leaves behind

A bubble does more than destroy capital on a delay; it is also how expensive infrastructure gets financed and then left behind for others to use. The telecom crash left the fibre that carried the broadband decade, and the companies that walked out of the dot-com wreckage were the ones people kept using while the tickers fell. In other words, the capital evaporates, whereas the infrastructure and the habits remain.

For anyone building right now, therefore, the overbuild is a subsidy: frontier capability rented below replacement cost, open-weight models one product cycle behind, tooling improving monthly. My conclusion has been to spend this period harnessing that capability into tools I need: the agent pipelines, the review gates, this notebook’s own publishing machinery. The marginal cost of trying an idea has never been lower, and the bubble is paying for it.

Squaring the decision to build on AI with the studies above means holding both results at once. The averages measure average use: chat in a side window, no verification loop, no harness. The METR developers lost their minutes to prompting, reviewing, and waiting inside workflows that were not built for the tool. My own ledger, in contrast, is positive exactly where the work is harnessed, gated, and verified, and I treat that as the variable under my control rather than a refutation of the data; it is a hypothesis this notebook exists to test in public.

Three rules for the downturn

Everyone has the same models, the same subsidized compute, the same agents; capability no longer provides a durable advantage this time, however impressive the demo. I am operating by three rules.

(1) Traction and retention are the hard part, and they were always the hard part. Treat them as the product problem, and treat the AI as supporting infrastructure. A tool nobody returns to has no value once the subsidy ends. (2) Build tools you will use yourself no matter what any customer says. Using them daily guarantees a retention floor of one user, and one retained user who feels the pain daily beats a hundred sign-ups who churned. In addition, it is the only feedback channel that survives a downturn. (3) Keep the cost structure honest. Anything whose unit economics only work at subsidized token prices will break when those prices normalize.

What is not proven yet

It is not proven, and the studies above are exactly why I say so out loud. “Build useful tools and retention follows” is a thesis, not a law, and the sample size on my side of the bet is one researcher’s tooling: the agents already carry a real share of my daily work, this site included, and the time they return each week is not hypothetical to me. That would not convince a skeptic. It convinces me enough to keep going, which is what a lab notebook is for; the market will grade the rest.

What I will watch next, dated 2026-07-09: whether my own tools survive their first quarter of real use without me subsidizing them with enthusiasm. If retention is the hard part, I should feel it first-hand before the quarter is out.