Glossary

Every piece of jargon we use on this site, in one or two plain sentences. If a term on a page confuses you and isn't here, that's our bug — tell us.

Hyperscaler

The giant cloud companies — Google, Microsoft, Amazon, Meta, Oracle — writing the checks for AI infrastructure. Their capital spending is the demand engine of the buildout.

Capex

Capital expenditure: money spent on long-lived assets like data centers, chips, and power equipment, as opposed to day-to-day operating expenses.

PJM

The grid operator for the eastern US. Its capacity auctions set the price of reserving future electricity — a price that rose roughly 12x in three years as data centers competed for power.

Capacity price

The cost of reserving future electricity supply in a grid auction. PJM's ran from $28.92/MW-day (2024/25) to $333.44/MW-day (2027/28) — the clearest signal of data centers bidding against everyone else for power.

Interconnection queue

The waiting line for new power plants to plug into the grid. About 2,300 GW of US generation and storage is waiting, with a typical project waiting roughly 4.5 years to connect.

PPA (power purchase agreement)

A long-term contract to buy electricity — for example, Microsoft's 20-year PPA taking 100% of the restarted Three Mile Island Unit 1's output.

Training vs. inference

Training builds an AI model (an enormous one-time compute job); inference runs the model for users (ongoing compute per query). Both drive data-center demand, on different timelines.

Circular financing

Money flowing in loops between AI players — for instance, a chipmaker investing in a cloud customer that buys its chips — which can make demand look larger than end-user need alone would support.

Ratepayer

An ordinary electricity customer. The central 'who pays' question of the buildout is whether data-center grid and water upgrade costs land on ratepayers or on the data centers themselves.

Confidence: Official

The figure comes from a primary or official source — a company filing, government data, or a grid operator's own numbers.

Confidence: Press-reported

The figure comes from reputable press reporting rather than a primary source. Credible, but one step removed.

Confidence: Proxy / estimate

The figure is modeled, estimated, or a projection. Methodology-dependent — read it as an order of magnitude, not a measurement.

Direct (link)

A connection where A mechanically causes B.

Indirect (link)

A connection where A contributes to B through other factors.

Correlated (link)

A connection where A and B move together, but we can't prove one causes the other.

Circular (link)

A connection where A and B feed back into each other.

Vintage

The as-of date stamped on a dataset. Figures get revised, so the vintage tells you how fresh the numbers you're reading are.

Who wins / who pays

The site's organizing frame: for every AI-economy development, who captures the value and who bears the cost.

Sovereign AI

Government-funded national AI infrastructure efforts — for example, the UK's Sovereign AI Fund — where states, not just companies, buy into the buildout.

AI chip (GPU)

The specialized processors that AI training and inference run on. The scarcest physical input to the buildout — and the reason one company, Nvidia, sits at the center of the market map.