The Subsidy Problem

A couple of weeks ago I sat down and made sure I could still get my work done if Claude vanished tomorrow. Not because I think it will. I’d been reading about the money underneath all of this, and almost nobody is paying what this stuff actually costs.

Last year the AI companies pulled in something like $25 billion against more than $250 billion spent on the infrastructure to deliver it , which is about a dime of revenue for every dollar poured into the ground. That gap is wider now than it was heading into the 2001 telecom bust , the one where a lot of smart people lost a fortune laying fiber nobody needed yet. And this isn’t off in some corner of the economy. AI spending was roughly three-quarters of US GDP growth in the first quarter of this year .

I’m not going to sit here and call a crash. I don’t know, and I’ve gotten suspicious of anybody who says they do. Either way, that dime-per-dollar means something for you and me right now, today. The price we all pay for AI is one somebody else is heavily covering. There’s a land grab for users, and a mountain of venture and hyperscaler cash is quietly eating away at the difference between what you’re charged and what it actually costs to serve you.

And it’s fine, mostly. New technology is introduced to the world all the time and subsidized. A lot of businesses have leaned on the cheap price without seeming to notice it’s on loan. They’ve priced products around it. They’ve built margin models on it. I’ve talked to people whose whole plan only pencils out because a model call costs what it costs this month, and I don’t think many of them have stopped to ask what happens when that number moves.

The version everyone imagines is the machines going dark, but that’s the least likely thing to happen. The models already trained still run on the chips already bought. The open-weight ones (Qwen, Llama, and the rest) are just files sitting on a drive, and a company going under doesn’t reach across the internet and delete them. What’s far more likely is boring: spending slows, the weaker labs get bought or fold, and whoever’s left finally charges what inference actually costs to make a dollar.

The risk here isn’t the dramatic one. Not “what if AI goes away.” What happens to your unit economics if the per-call price doubles or triples on a couple of months’ notice? If your product only works at today’s price, then what you’re calling a margin is really somebody else’s fundraising, passing through you until it stops. It’s a judgment call sitting quietly in a lot of 2026 business plans.

So what did I do about it, short of panicking, which I’m genuinely not doing? Nothing fancy. I treated the model like any other supplier whose price I don’t get to set. I put a thin layer in front of it so moving from one provider to another is a config change instead of a weekend of rework. I pulled down the open models I’d want as a backup onto my own machine, so a lab blowing up is an annoyance, not a wall. And I keep the one part that can’t be replaced, my own data and the context I’ve piled up over the years, in formats I actually own, because I can replace a model in an afternoon, but that context took years to build, and I don’t get it back.

None of that takes believing the sky is falling. I don’t, particularly. I’d bet against most of the folks confidently calling the top; they’ve been at it for years now. I did the work anyway because I don’t want my ability to do my job riding on somebody else’s next funding round. That’s true whether the money keeps flowing or it dries up, which is sort of the whole point.

Anyway, I’ll leave you with a boring question. If the price of the thing you built on jumped three, four, five times on ninety days’ notice, would you be okay? If the answer’s yes, good, you’ve already done the thinking. If the honest answer is you’d rather not find out, well, there’s your tell. Cheap AI has quietly become load-bearing, and it’s worth a hard look now, while it’s still cheap enough to plan around.