The five largest tech companies will spend roughly $690 billion on AI infrastructure this year. That number is so large it stops meaning anything. So let’s try to make it mean something.
The Large Hadron Collider cost about $9 billion to build. That’s the most ambitious pure-science project in human history — 27 kilometers of superconducting tunnel under Switzerland and France, built over 20 years, funded by contributions from over 100 countries, designed to answer one question: what is matter made of? It found the Higgs boson. It changed physics. It took a generation.
We now spend the equivalent of one LHC every five days on AI compute.
The Three Gorges Dam — still the largest power plant ever built — cost roughly $37 billion. The International Space Station, assembled over 13 years by 15 nations, came in at around $150 billion. The Great Pyramid of Giza, rebuilt with modern materials and machinery, would run maybe $5–10 billion. These are the kinds of projects that define an era, end up in textbooks, acquire mythology.
AI training infrastructure is that scale of project. Happening simultaneously. In a dozen locations. Every year. With no coordinating body, no shared purpose document, and no agreed name for what we’re building.
The LHC is the most instructive comparison, because it’s the only other entry on the list that was built purely to learn something — not to generate power, not to move goods, not to house people. It was a bet that the question was worth it. A hundred countries signed a piece of paper saying: we don’t know what we’ll find, and we’re doing it anyway.
AI training infrastructure is the same kind of bet, at a hundred times the scale, made competitively. There’s no shared question. Microsoft, Google, Amazon, and Meta are each building their own version, racing, not collaborating. The LHC was the last time humanity pooled enormous resources around pure curiosity. What we’re doing now looks more like an arms race that happens to produce something useful.
I don’t think that makes it wrong. But it does mean we should be honest about what it is. The wonder of the ancient world we’re most likely to be remembered for is being built right now — in server farms in Virginia and Iowa and the Arizona desert — and we haven’t even agreed on what to call it, let alone what question it’s supposed to answer.
Maybe that’s fine. Maybe the question emerges from the thing. It usually does.
Or maybe we should write it down first.
Leave a Reply