There is a number making the rounds right now that most of us can’t really get our heads around. In 2026, the four largest hyperscalers — Alphabet, Amazon, Meta, and Microsoft — are on track to spend a combined ~$725 billion in capital expenditures, up roughly 77% from a 2025 figure that was already a record. McKinsey puts the cumulative data-center bill at $6.7 trillion by 2030. The IEA expects AI to pull nearly a full Japan’s worth of electricity out of the global grid by the end of the decade. Goldman sees U.S. data-center power demand more than doubling from 31 gigawatts in 2025 to 66 gigawatts by 2027.
These are not technology numbers. They are infrastructure numbers, the kind the modern economy has produced only a handful of times — the railroads, the telephone and internet networks, the interstate highways, and the one that rhymes loudest with this moment, the electrification of America.
I’ve spent a lot of time lately thinking about what that last one actually teaches us, and I’ve written it up in a new white paper called The Dynamo and the Token. It’s the first in a series I’m doing on infrastructure cycles and where value migrates.
The short version: electrons, bits, and tokens are a lineage, not a set of analogies. Electrification gave the economy the electron. Computing gave it the bit. Generative AI is giving it the token. Each one followed the same arc — heroic capex, hype peak, price collapse, and then a long, profitable second act lived almost entirely on the layer above. The fortunes from electrification did not go to the utilities. The fortunes from the fiber buildout did not go to Global Crossing or Worldcom. And if the pattern holds, the fortunes from AI will not go to the people building the grid either. They will go to whoever reorganizes their business around the assumption that capable intelligence is cheap, abundant, and everywhere.
That is an uncomfortable argument to make on a network that runs on advertising from many of the vendors most exposed to it. I am making it anyway, because I think it is right, and because the only thing an analyst on an ad-supported platform actually owes the audience is the willingness to say the inconvenient thing.
Inside the paper: why intelligence commoditizes even in the bull scenario, why the neoclouds are this era’s regulated utility, why the integrated giants win as appliance makers who happen to own a power plant, and where the durable moats actually live — in workflow, memory, proprietary data, and the reorganization of work around cheap cognition.
If you are building anything in AI or on top of it right now, this is the essay I’d want you to read first.
No registration required
