Will we get cheaper tokens?

The AI bubble is not really about valuations. It is about infrastructure economics.

The biggest tech companies are spending roughly $700B a year building AI capacity, betting that demand will show up to justify it. It might. But right now the AI economy runs on two distortions pulling in opposite directions.

Upstream, GPU vendors have extraordinary pricing power, because demand for accelerators far outruns supply.

Downstream, most frontier model providers run on thin margins or outright losses, because they are buying adoption and market share.
One layer prints money. The next layer subsidizes its customers. That is not a stable arrangement.

It gets stranger. The capital keeps circling the same ecosystem. Labs raise money, buy GPUs, rent hyperscale cloud, and the infrastructure vendors book the revenue that funds the next round. Nothing fraudulent about it. But everyone is holding everyone else up.

So the whole thing rests on one question: will AI demand grow fast enough to absorb all the compute being built?

If yes, today's spending looks visionary. If no, we get a classic infrastructure correction. Capacity turns abundant, hardware prices normalize, cloud GPU rentals fall, competition sharpens.

Here is the irony. That correction would be one of the best things to ever happen to AI users. Cheaper compute, leaner models, abundant infrastructure, all of it pushing the real cost of high-quality AI down for years.

The companies pouring in the money may get hurt. The people using AI probably win.

My bet: even if prices spike during a correction, inference ends up cheaper than today.

Which side of this are you building for?