Data Centres, Grids and the AI Build-Out
4 min read · updated August 3, 2026
“Where does a data centre’s electricity come from?” has no physical answer. A grid is a pool; electrons are not labelled. There are three accounting conventions that answer it, they routinely disagree, and the disagreement is the substance of most disputes in this area.
Get the units right first
Three quantities get conflated in almost every article. Capacity is power, in megawatts — the maximum draw, and what a “1 GW data centre” headline refers to. Energy is megawatt-hours, capacity multiplied by time and by how heavily the facility is actually used; announced capacity is not consumed energy and the load factor between them varies a great deal. Emissions are tonnes of carbon dioxide equivalent, energy multiplied by an intensity factor whose choice is the subject of the next section.
Announced capacity is also not built capacity. Announcements are made early, sites are secured before they are energised, and projects are cancelled. Treating a pipeline of announcements as a forecast of load is the most common error in coverage of this subject, and it runs in the direction of overstatement.
Three answers to “where does it come from”
| Convention | Description |
|---|---|
| location-based | Multiply consumption by the average carbon intensity of the local grid over the period. Reflects the physical grid the facility sits on. Ignores every contract the operator signed, so a facility that funded new wind looks identical to one next door that did not. |
| market-based | Attribute the generation the operator contracted for — power purchase agreements and energy attribute certificates. Reflects investment decisions and can produce a very low or zero figure for a facility drawing from a fossil-heavy grid, if the certificates were bought elsewhere or at other hours. |
| marginal | Ask what generator actually changes output when this load is added. The most decision-relevant convention for judging a new facility, and the hardest to compute. It frequently gives a higher number than either of the others, because the marginal unit on a constrained grid is often the dirtiest one running. |
The live methodological argument is about hourly matching. Annual netting lets certificates generated at a sunny midday offset consumption at midnight, when the marginal generator is something else entirely. Hourly, or 24/7 carbon-free, matching requires supply and consumption to coincide in time, which is far harder and far more expensive, and which proponents argue is the only version that drives investment in the firm clean generation the grid actually lacks. Critics of hourly matching argue that the additional cost buys less emissions reduction than spending the same money on new generation wherever it is cheapest. Both arguments are serious and the disagreement is genuinely unresolved.
Practical consequence: when an operator states a clean-energy percentage, the number is meaningless until you know the convention and the matching interval. Ask for both.
What actually constrains the build-out
Not the price of electricity, in most markets. The binding constraints are physical and procedural:
- Interconnection. Connecting a large load to the transmission system requires studies, queue position and utility approval, and in congested regions the wait is measured in years. Where interconnection is available has become a primary determinant of where compute is sited.
- Transmission. Generation and load are often in different places, and new lines take longer to permit than either a generator or a data centre takes to build.
- Equipment lead times. Large transformers, switchgear and turbines have order books measured in years. This constraint is invisible in energy statistics and decisive in project schedules.
- Local consent. Land use, noise, water rights and local politics. Several jurisdictions have introduced moratoria or conditions on new large facilities, and this is now a routine risk rather than an exception.
Structural responses and their trade-offs
Behind-the-meter generation — building generation on site — bypasses the interconnection queue and raises the question of whether the facility is contributing to or free-riding on the shared system. Long-term nuclear contracts, including for existing plants, supply firm clean power; critics point out that contracting existing output can simply move it away from other consumers rather than adding any. On-site gas is fast and is a straightforward emissions increase.
The most interesting response technically is flexibility. Training workloads are relatively shiftable in time and, to some extent, in place; interactive inference is neither, because a user is waiting and because data residency rules may pin it. A grid values flexible load highly — it can absorb generation that would otherwise be curtailed and stand down when the system is tight. That makes a training cluster a genuinely different kind of grid citizen from an inference fleet, and policy that treats “AI data centres” as one category misses the most useful distinction available.
Efficiency is the response that gets least attention because it is unglamorous and it is where a great deal of the available headroom sits. Cooling design, operating temperature, power conversion losses, workload consolidation and simply retiring older accelerators all move the facility-level numbers, and the last of these interacts with supply: hardware kept in service past its efficient life because replacements are unobtainable is a hidden energy cost of a hardware shortage. None of this changes the siting problem, which is about where and when the load appears rather than how large it is.
Who pays for the connection
A large new load usually requires network upgrades. Whether the cost falls on the connecting customer or is spread across the rate base is decided in regulatory proceedings, and it is a real distributional dispute with organised parties on both sides.
The case for socialising some of it: network reinforcement benefits everyone connected to it, large loads bring investment and tax revenue, and specialised tariffs risk driving projects to other jurisdictions. The case against: existing households and businesses did not choose the load and should not underwrite it, and stranded-asset risk falls on them if the facility closes or scales back. Several utilities have proposed special large-load tariffs with minimum-take provisions precisely to address the second, which is an instrument worth understanding rather than a settled answer.
Whatever is claimed about a specific facility or region, ask for the boundary, the convention, the matching interval, and whether the figure is capacity or energy. Specific numbers in this area date quickly and none are quoted here for that reason; check current filings and regulator publications, which are the primary sources.
One framing worth resisting in either direction: a new large load is neither automatically a burden nor automatically an investment. It depends on where it connects, when it draws, what it funds, and what the alternative use of that grid capacity would have been. Those are answerable questions about a specific project, and they are almost never the questions being argued about when the discussion is conducted at the level of the industry as a whole.