Gigawatts Are the New Benchmark

There is a new vanity metric in AI, and unlike parameter counts, this one is load-bearing: the gigawatt.

Meta's Hyperion campus in Louisiana is pitched at 5 GW. Stargate's Abilene site talks in gigawatt phases. Nearly every hyperscaler earnings call now includes a capacity number denominated in power, not servers. The International Energy Agency projects global data-center electricity consumption roughly doubling from the mid-2020s to somewhere near 945 TWh by 2030 — more than Japan consumes today.

How to read a gigawatt claim

Not all announced gigawatts are equal, and the gap between a press release and an energized substation is where most of the risk in this cycle lives. A useful mental model is a funnel:

Stage	What it actually means
Announced	A land option and a rendering
In queue	An interconnection request filed with the grid operator
Contracted	Signed PPA or energy services agreement
Energized	Electrons flowing to racks

The attrition between stages is enormous. Grid operators report interconnection requests that sum to multiples of their entire existing load — ERCOT alone has fielded large-load requests in the tens of gigawatts, far more than will ever be built. Some of the queue is real. A lot of it is the same project shopping three utilities, or speculative land plays wearing a data-center costume.

The benchmark that matters

The honest metric — the one that separates operators from announcers — is energized megawatts per quarter. It compounds like revenue, it cannot be faked with a rendering, and it is auditable from utility filings.

Watch that number, and the league table looks different from the press-release league table. The players who quietly locked in turbines, transformers, and switchgear in 2024 are hitting energization dates. The ones who announced campuses first and asked about transformers second are discovering that the lead time on a large power transformer is measured in years, and that GE Vernova's gas-turbine backlog stretches toward the end of the decade.

Chips made AI possible. Power decides who actually gets to run them.