AI Is Priced in Dollars and Constrained by Megawatts. Four Percent of Deals Publish the Megawatts.

Across the 606 AI deals in our deal tracker, 23 state how much electricity they will draw. Those 23 commit 56.9 gigawatts.

Both halves of that sentence matter. The capacity being signed for is enormous. The rate at which anyone discloses it is under four percent.

## A dollar figure is the press release. A megawatt figure is the constraint.

Price tells you what a deal cost. Capacity tells you whether it can physically happen. Chips can be air-freighted. Substations, transmission lines and cooling water cannot, and the queue to connect a large load to a grid is measured in years.

So the megawatt number is the one that determines whether the dollar number is deliverable. It is also the one that usually goes unstated.

Look at what disclosure looks like when it happens. Anthropic is committing $45 billion over six years to lease 460 MW of Nvidia Vera Rubin capacity from Nscale's Monarch campus in West Virginia. That is not a chip order. It is a long-dated lease on power and racks, closer to how an airline finances aircraft than to how a software company buys servers.

## The disclosure rate is the actual finding

Almost every AI deal announces money. Very few announce load. That asymmetry is not innocent: capacity figures invite questions about which grid, whose water, and whose electricity bill, while a dollar figure invites applause.

The consequence is that public estimates of how many gigawatts AI is committing to are extrapolated from a thin set of announcements. Ours is thin too. The difference is that we can show you the whole set rather than model around it.

## Where the capacity concentrated

In 2025, three deals in our corpus disclosed a figure, totalling 16.2 gigawatts. In 2026 so far, twenty deals disclosed, totalling 40.7 gigawatts. More announcements are carrying the number, and the numbers are bigger.

About seventy percent of that capacity sits in compute deals, meaning leases and campus builds, rather than in hardware purchases. That distinction tracks the shift we documented in The AI Deal Mix Shift, where compute contracts overtook acquisitions. Buying GPUs is procurement. Securing somewhere to run them is infrastructure, and infrastructure has a lead time nobody can buy their way out of.

The largest disclosed commitments are national in scale: SoftBank's 5 GW in France, Meta's power purchase agreements for up to 6.6 GW signed alongside funding for TerraPower and Oklo, Nvidia and SK's 2 GW AI factory in Korea.

## The small numbers are more interesting than the large ones

A 5 GW campus is an aspiration with a decade attached. The deals worth watching are the ones sized to what a grid can actually absorb now.

MTN is building roughly 150 MW across South Africa and Nigeria. Reliance is building 168 MW in Jamnagar and leasing it to Meta. Valar Atomics and Nvidia are exploring 30 MW powered by a single small reactor in Utah.

Thirty megawatts is a rounding error against six gigawatts. It is also a project that could be energised before most of the gigawatt announcements have broken ground.

## What to do with this

**When you read an AI infrastructure announcement, find the load figure before you react to the price.** If it is absent, you are reading a financing story, not a capacity story, and you cannot yet tell whether anything gets built. If it is present, check the hedge word: "up to 5 GW" and "5 GW" are different claims, and the first one is a ceiling somebody chose to publish.

**Watch the sub-500 MW deals for the real timeline.** The gigawatt headlines set the narrative, but the projects sized to existing interconnection capacity are the ones that will actually be serving tokens in 2027. If you are forecasting AI supply, weight them accordingly.