Earnings — is the machinery paying for itself?

The coverage ratio: the quasi-rents the AI capital actually earns, against the annual cost of holding it. A gap is normal for a young technology — the trajectory is the signal, not the level.

Two bars are often confused. This ratio sets earnings against depreciation plus a required return on the whole installed stock — a stricter test than “revenue vs depreciation” alone, which is why it sits well below 1.0 even as the picture improves. On the looser depreciation-only test, independent bottom-up work (Exponential View, The State of the AI Economy, 25 Jun 2026) finds quarterly AI revenue first cleared quarterly infrastructure depreciation around Q4 2025 / Q1 2026 — consistent with this dial’s upward trajectory. Source.

What the AI capital earns

Realized AI revenue stated on earnings calls (K-holders only; guidance and chipmaker revenue excluded). Each figure is kept with its verbatim sentence.

What it costs to hold — the user cost

Data-centre construction

The durable, building-shaped part of the build-out — monthly, since 2014.

GPU rental prices

The open-market price of an hour of AI compute — a check on what the machinery is worth. Falling prices would be the first hard evidence of overcapacity.