I keep thinking about a number Dario Amodei has started saying out loud. Five trillion dollars of compute over five years. A trillion per year by 2027. He says it not as a boast but as a line he refuses to cross. "If I'm just off by a year in the growth rate," he told Dwarkesh Patel, "you go bankrupt." That is the CEO of Anthropic describing the shape of the industry he helped build, and calling it ruinous.
The moved bar
The interesting part is not the number. The interesting part is the word Amodei keeps using. Responsible. He is the responsible one, because he is only planning to spend a couple hundred billion, not a trillion a year. The competitors, in his framing, are yoloing the global economy. This is probably accurate. It is also a language problem.
We are in an industry where "responsible" now means "my capex is only the size of a G7 economy's annual defense budget." The bar for prudence has drifted up by two orders of magnitude in about thirty months. Nobody threw a party when it happened. It just is.
I am not mocking Amodei. In relative terms, he is saying something careful. That is the point. Careful is now shaped like this.
Who is actually in the boat
The phrase "burn the boats" gets used a lot in these conversations. It sounds decisive. It leaves out a detail, which is: whose boats, and who is in them.
The capex is investor capital. The externalities are not.
U.S. data centers are projected to go from about 80 gigawatts of draw in 2025 to 150 gigawatts by 2028. In Virginia, 26 percent of the state's electricity already goes to data centers. In Iowa, 11 percent. Roughly two-thirds of new hyperscale sites built since 2022 sit in areas already under water stress. Hyperscalers are projected to consume 16 to 33 billion gallons of water per year by 2028.
These costs do not appear on a pitch deck. They appear on a municipal utility bill, in a county that did not get asked whether trading its aquifer for a training run was a fair deal.
The distribution problem, not the total
It is tempting to treat this as a referendum on AI itself. It is not. Every large infrastructure wave has asked the same question and received a messy answer. Railroads, the electrical grid, the telephone system, the internet backbone. Each absorbed capital and materials at scales their critics found obscene. Each, in retrospect, looks different depending on where you lived and what the local deal was.
The issue isn't whether the compute is worth it in aggregate. The issue is that the aggregate is a useless abstraction. Somewhere, the ledger is real. Someone pays the power bill. Someone notices the well ran low. Someone waits for a model that rate-limited them because the training run had priority. The aggregate calls this efficient. The person at the bottom of the ledger calls it something else.
What would actually be responsible
A minimum version of honest would be calling things by their names before deciding. Not "responsible scaling," but "I am spending $200 billion of other people's money and drawing 50 gigawatts from grids I do not own, and I think it will be worth it." That is a defensible sentence. Maybe even a true one. It is a different sentence than the one being printed.
We might, in the end, decide the trade is worth making. That would be a choice, and choices are fine. What we are doing right now is not choosing. We are listening to the people inside the boat tell us, politely, that they are the responsible ones, and that the boat is on fire for our benefit.