AI data centre power — why the unit changed from megawatts to gigawatts

Until recently, data centre news came in megawatts. Twenty megawatts was a large site; fifty made headlines.

It now comes in gigawatts.

The unit moved three orders of magnitude. Here is why, and what it means.

1. How large a gigawatt is

Two reference points.

By generation — a large nuclear reactor's nameplate capacity is around 1 GW. So a "1 GW data centre" consumes what one reactor produces.

By settlement — 1 GW roughly matches a mid-sized city's electricity demand.

Both are approximations that shift with region, season and capacity factor. The direction is unambiguous all the same: data centres have become objects of generation planning.

This page covered a 55 MW Korean project on August 6 in Sejong's 55 megawatts — Naver spends 14.6 trillion won on a 'sovereign AI factory'. The Ohio project announced in August 2026 plans 4.25 GW initially with a further 3.75 GW optional. That is roughly 77 to 145 times larger.

2. Why — the load shape is different

The point is not the total. It is the shape of the load.

The middle row is the problem.

Training a model requires tens of thousands of accelerators running simultaneously and without pause. Interrupting the run breaks it, so the load cannot be dialled down. What the grid sees is a flat maximum sustained for months.

For a grid operator that is the least tractable form of demand. It cannot be shifted off peak, and demand-response programmes cannot trim it.

Inference then arrives on top. Once a model is in service, each user request consumes compute, and that load rises with the user base. The demand does not fall when training finishes — a different kind of load begins.

The aggregate reflects it: AI-optimised servers are estimated at 31 percent of all data centre power in 2026.

3. Global scale — the IEA numbers

Those come from the IEA's April 2026 update. A doubling in five years.

For calibration: world electricity demand grew about 3 percent in 2025. AI data centre consumption grew 50 percent over the same period — more than 16 times faster than overall demand.

Year-by-year actuals differ between sources; some put 2026 at 565 TWh. The divergence comes from what counts as a data centre and whether on-site generation is included. This page uses only the IEA's 2025 and 2030 figures.

4. The bottleneck moved — from chips to the grid

In 2023–24 the constraint was accelerator supply. Projects slipped because chips were unobtainable.

The 2026 constraint is different.

You have the chips and nowhere to plug them in.

Goldman Sachs has put the structural US shortfall at:

Roughly fivefold in two years.

The shortfall is not only generation capacity. Grid interconnection is the larger constraint. Substations and transmission lines take years; new generation takes longer. Data centre construction is much faster than either.

That mismatch in lead times is the substance of the industry's current problem.

5. Three workarounds

The Ohio project announced in August 2026 is the third kind. Nvidia is backing up to $105 billion in financing for an OpenAI data centre in Ohio, with SB Energy building the PORTS-Pike campus in Pike County under a 20-year lease to OpenAI. Initial capacity is 4.25 GW, with an option for 3.75 GW more.

The 20-year lease is the striking term. Data centre equipment turns over far faster than that. What a twenty-year contract secures is not hardware but grid access and land. The asset being locked in is electricity.

6. Where this lands in the accounts

As power scale grows, the shape of the balance sheet changes with it.

Cloud businesses were relatively asset-light: buy servers, depreciate them over three to five years. Now generation equipment, transmission assets and twenty-year site commitments sit alongside.

This page covered the two related questions separately:

Power infrastructure enlarges both problems at once. It increases the capital that must be raised, and it lengthens and blurs the useful life of what must be depreciated.

7. Why this is also a semiconductor indicator

The third is the channel through which power news reaches equity markets in supplier economies. A larger data centre power plan means a larger accelerator order book, and those accelerators carry memory made in Korea and elsewhere.

An energy story is also a leading indicator of semiconductor demand.

8. What is not confirmed

• Year-by-year actuals — sources differ for 2026 (565 TWh among them). Only the IEA's 2025 and 2030 figures are used here.
• The 31% share — depends heavily on the compiler's definitions; underlying data was not checked.
• Goldman Sachs figures — 9.3 GW and 45 GW come from citing coverage, not the original report.
• Ohio capacity — 4.25 GW and 3.75 GW are announced plan figures, not operating capacity.
• The analogies — comparing 1 GW to a mid-sized city or one reactor is approximate and varies by region, season and capacity factor.
• Next checkpoint — what share of announced capacity actually reaches an interconnection. The gap between plan and operation sets the real growth rate of this sector.