In April, the chief technology officer of one of the world’s largest consumer-technology companies admitted — almost in passing — that the firm had exhausted its entire annual artificial-intelligence budget in four months. Not because the projects had failed. Because they had worked. The agents were let loose, they produced, and they consumed. A line meant to last twelve months lasted one third of that.
We keep returning to that admission, because it is the most honest description we have of where the AI economy actually arrived this quarter.
The constraint moved.
For two years, the binding constraint on artificial intelligence lived in the world of ideas — better architectures, better training recipes, better ways to wire models into work. The cost of running these systems was, to a first approximation, free, because the laboratories were paying it for us. Inference was sold below cost to win users, in the time-honoured way of every platform that wanted to own a market before it charged for it. We have spent a decade learning to recognise that pattern; it is the pattern that built the consumer internet.
That era ended in the second quarter. Quietly, without a single defining headline, the constraint moved out of the world of ideas and into the world of physics — into compute, into power, and finally into silicon itself. The cost of intelligence stopped being something a balance sheet could absorb and started being something an economy has to ration. We think this shift — from the subsidy era to the scarcity era — is the most important development of the first half of 2026, and the one most likely to be misread by the people whose job is to price it. So we will spend most of this letter on it.