Jul 2026
Your idle GPU is worth nothing
Trade doesn't happen because two people agree on a price. It happens because they disagree about value — and your sleeping graphics card is the proof.
Right now your GPU is asleep. That idle hour is worth exactly nothing to you — you weren't going to use it. To me, mid-render with a deadline, it's worth forty cents. Neither of us is wrong. That gap, where the same hour is worthless to you and precious to me, isn't a glitch in the market. It is the market.
We're taught that trade is about agreeing on a price. It's the opposite. If we valued your idle hour identically, there'd be no reason on earth to move it. Trade runs on disagreement. The buyer walks away sure he got the better end. So does the seller. And they're both right, because value was never a number stamped on the thing — it lives in the person, in the moment, in whatever else they could've been doing instead. A guy named Menger worked this out in the 1870s and we've been slow to believe him ever since.
Here's the part that breaks people's brains. It doesn't even matter if I'm better than you at everything. Say I've got faster cards and cheaper power and I write tighter kernels. You'd think I should just do it all myself. I shouldn't. My hardware has a best use — the job that pays me the most — and every hour I burn on a cheap job is an hour stolen from the expensive one. So I hand you the cheap work even though I'd technically do it better, because your time isn't competing with a bigger paycheck the way mine is. Ricardo proved two countries both get richer trading like this even when one is better at making literally everything. The same math says your five-year-old laptop has a seat at a table full of H100s.
So if the world is wall-to-wall with these gaps — idle silicon here, starving jobs there — why isn't it already one giant marketplace?
Because the trade costs more to set up than it's worth. That's the whole thing Coase was pointing at: the friction of finding the counterparty, trusting them, and settling up is usually bigger than the gain itself. Nobody rents you their GPU for forty cents when checking that you didn't quietly swap in a smaller model costs four dollars and a lawyer. The gap is real. The plumbing to close it just never paid for itself. So the hour stays idle and the value stays stranded, forever, at the exact moment someone three timezones over would've paid for it.
That's the actual thing we're building. Not a new kind of value — the gaps were always there, humming under everything. We're killing the transaction cost that kept them stranded. The job ships to the machine, not the machine to the job, because moving a model's weights across the internet costs a million times more than running where they already sit — so the compute goes to the data and the gap closes at the cheap end. You pay in a bearer token that clears in a single message: no account, no invoice, no net-30. And you don't have to trust the stranger who ran it, because the work is deterministic and you can re-run a sample and catch a liar for pennies. Finding, trusting, settling — Coase's three taxes — all driven toward zero.
Strip everything else away and a network like this is one sentence: a machine for noticing that your nothing is my something, and moving it before the moment passes.
Your idle GPU is worth nothing. That's not the sad part. That's the opening.