Fifty-six cents on the dollar

The most useful number published about AI economics this year is a ratio, and it moved fifteen cents in a quarter. What it does not tell you is whether it moves again.

Compute cost per revenue dollar went from 71 cents in Q1 2026 to 56 cents in Q2, per figures given to investors.

Why a ratio beats a headline number

Revenue tells you about demand. Capex tells you about conviction. Neither tells you whether serving the product costs less than selling it, which is the only question that determines whether any of this is a business.

A cost-to-revenue ratio does, and a quarter-on-quarter delta does more — it says which direction.

The decomposition nobody published

Fifteen cents can come from several places: better inference efficiency, cheaper contracted capacity, a mix shift toward higher-margin products, or fixed capacity spread across more revenue. Only the first two repeat on their own.

A ratio going 71, 56, then somewhere in the forties is a trend. A bounce back means it was composition.

What it implies for the buildout

If serving cost keeps falling while volume rises, enormous forward compute commitments look like sensible pre-purchasing of an input that gets more productive. That is the bullish case, and this is the first quantitative support it has had.

It also explains where the silicon competition is headed. Intel and Qualcomm are aiming at inference, which is projected at two thirds of all compute in 2026 — and inference is precisely the cost line this ratio measures.

The four biggest clouds spent about $410bn on AI infrastructure in 2025 and guided over $700bn for 2026. Somebody is going to want to know what fraction of that comes back, and this ratio is the first honest look.

Let's Data Science — Anthropic Projects $559 Million Q2 Operating Profit → · Deloitte — Compute power and AI → · Barchart — Intel Sets Sights on Nvidia and AMD With Upcoming AI Data Center Chip Launch →