"AI bubble" is the argument that the money being invested in artificial-intelligence infrastructure is far larger than the revenue that infrastructure currently produces, and that some of the revenue being reported is the same money going round a small circle of companies.
This page explains the second half — the circular part — because it is the part with actual arithmetic in it, and because it is what our game is built on.
Imagine three companies: one makes chips, one trains models, one rents out datacentres.
The billion is back where it started. Nothing was built that would not otherwise have been built. But three deals worth $3B in total were announced, and each company can report its share as revenue, investment or bookings.
The echo ratio is the number we use for this: total announced deal value divided by the real outside money that ever entered the system.
In the example above, one billion of real money produced three billion of announcements. That is an echo of ×3. Run the same billion round twice more and it is ×9, with the same single billion dollars underneath it.
An echo ratio is not by itself proof of fraud, and this is worth stating plainly: real companies genuinely do buy things from each other, and vertical relationships in a young industry are normal. What the ratio measures is how much of a reported number could be the same money counted more than once — which is a question worth asking before treating that number as demand.
Announced deals raise a valuation. A higher valuation raises how much can be borrowed against it. The borrowing pays for the infrastructure the deals implied. As long as the valuation keeps rising faster than the bills arrive, the arrangement funds itself.
The bills, however, are not circular. Electricity is real. Cooling is real. Interest is real. Those are paid to people outside the circle, in money that has to come from somewhere outside it.
Historically, manias end when the outside bills arrive faster than new outside money does — not when somebody proves the thing was a bubble. Tulip contracts in 1637, the South Sea Company in 1720 and the dotcom crash in 2000 all ran on the same shape, and in each case the people who did best were the ones who left early and looked foolish for a while.
This page explains a mechanism. It does not predict anything, it is not a claim about any real company, and it is not investment advice.
Our free browser game makes you build one. You get sixteen quarters, three actions each, and a board that prints your echo ratio the entire time. Play AI Bubble Tycoon — no account, nothing installed, about ten minutes.