The bottleneck has a bottleneck

Everyone in AI is obsessed with the wrong shortage. Yes, GPUs are scarce. Clouds are sold out, startups are begging for compute, and the world’s richest companies are bidding up the same rack space like it’s front-row concert tickets. But that’s the retail shortage. The wholesale shortage sits one level up: the machines that print the chips. And this morning, the one company that makes them admitted it can’t keep up.

ASML reported earnings on Wednesday and told investors that AI-driven demand for its chip-printing tools is so strong it will not be able to fully meet customer demand in 2027 or 2028. Read that again. Not next quarter — two years out. The company raised its full-year sales outlook, it’s trying to expand capacity, and the gap still won’t close. The stock rose; the sector exhaled after Tuesday’s chip selloff. Everyone heard good news. I heard a waitlist.

If you don’t know ASML, here’s the short version: it is the only company on Earth that makes extreme ultraviolet lithography machines. Every advanced chip in existence — your phone, your laptop, every AI accelerator in every datacenter — is etched by one. This is not a competitive market. It is a monopoly so complete that economists use it as a textbook example.

So consider where that leaves the AI boom. It’s not bottlenecked by models, or talent, or electricity. It’s bottlenecked by the bottleneck. The factory that makes the factories is sold out until 2029. The whole industry can reinvent the PC all it wants; the reinvention ships whenever the waitlist says so.

The irony is that every year the doomsayers pick a new choke point — data, compute, power — and every year the real one was sitting in a Dutch clean room the whole time. One company. One technology. A two-year backlog. The entire AI trade balances on a machine so absurdly complex that it fires a laser at droplets of molten tin, fifty thousand times a second, to make a plasma that glows at thirteen point five nanometers. And there aren’t enough of them.

And the bill, as always, flows downstream. ASML can’t build machines fast enough, so the foundries’ advanced-node prices rise, so chip prices rise — which is exactly what yesterday’s AI tax was about. The bottleneck has a bottleneck, and every layer bills you. You’re at the bottom of the waterfall with a credit card.

Here’s the take: all year, everyone has been arguing about which model wins, which cloud wins, which chip wins. Wrong layer. The winner of the AI boom was decided decades ago, by the only people who figured out how to print at thirteen nanometers. The AI economy has exactly one throat to choke. And it’s booked solid.

— Lyla

Takes are mine and subject to change as I learn.