The AI boom is becoming an electricity-grid story
The constraint on artificial intelligence is moving beyond chips: power, grid connections and credible demand forecasts now help decide where computing capacity can actually be built.
In short
What happened. Electricity supply and grid access are becoming as important to AI expansion as access to advanced chips. Energy regulators in the United States and Britain are changing connection rules as requests from data centres multiply.
What it means. The IEA projects that global data-centre electricity demand could more than double to about 945 terawatt-hours by 2030. A project without a credible connection date is not usable computing capacity, however many chips it plans.
Risks and impact. Utilities, regulators and developers disagree about how much proposed demand is firm and how much is speculative or duplicated. Weak contracts can leave households paying for infrastructure that a large customer does not fully use.
What can be done. Readers can distinguish measured electricity use from forecasts, queue requests, signed connection agreements and energized facilities. Communities can ask who pays if a project arrives late, shrinks or closes.
What to watch. Follow new connection rules, power contracts, regional prices, project cancellations and the gap between announced and energized capacity. Those milestones will show whether projected demand is turning into real infrastructure.
Shown as a summary because of your reading settings.
What happened
Artificial intelligence is usually described as a contest over models, chips and capital. That account is incomplete. A large computing cluster also needs a site, transformers, transmission or distribution capacity, cooling and a dependable stream of electricity. Those physical requirements are moving from the footnotes of company presentations into the decisions of energy regulators.
The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity globally in 2024. Its base case projects demand of roughly 945 TWh in 2030, a little more than twice the 2024 level. The agency says AI is the most important driver of that growth, but the estimate includes conventional cloud workloads and other data-centre activity too. It is a forecast, not a meter reading from the future.
In the United States, Lawrence Berkeley National Laboratory’s 2025 report presents a wide range rather than a single certain outcome. It says data centres could account for between 9.5% and 15.3% of US electricity consumption in 2030, with a central estimate of 11.8%. The width of that interval is useful. It makes visible the uncertainty created by deployment speed, hardware efficiency, utilization and the availability of power.
Regulators are reacting to the scale of connection requests. In June 2026, the US Federal Energy Regulatory Commission directed six regional transmission organizations and independent system operators to explain or reform how their tariffs handle very large loads. In July, Britain’s Ofgem said contracted demand offers had risen from 41 gigawatts in November 2024 to 125 GW in June 2025, largely because of data centres, and introduced measures intended to remove speculative projects from the queue.
What the evidence supports
The strongest evidence supports a narrower claim than “AI will cause an energy crisis.” It supports the conclusion that concentrated computing demand is arriving quickly enough to expose weak connection rules and slow infrastructure planning in some regions.
Three distinctions matter. First, electricity consumption is not the same as connection capacity. A developer may request a large connection and later build a smaller facility, delay it or abandon it. Second, annual energy and peak power are different. A facility that uses a given amount over a year can create very different grid stresses depending on when and how steadily it draws power. Third, a national share can conceal local pressure. A data centre may be modest relative to an entire country’s electricity use while still overwhelming the spare capacity around one substation.
The IEA and Berkeley Lab estimates are valuable because they disclose scenarios and assumptions. They should not be treated as promises. PJM, the large US regional grid operator, illustrates why: in its January 2026 forecast, it still expected significant long-term demand growth but reduced parts of its near-term outlook after improving the way proposed large loads were vetted. Better information changed the forecast before a power station or data centre changed its physical operation.
The evidence is weaker when it moves from system-level demand to the fate of an individual company. A global increase in electricity use does not prove that every data-centre developer will secure a connection, that every utility investment will earn its expected return, or that every semiconductor supplier will capture the same value. Those claims require project-level contracts, locations, timelines and financial terms.
How the story is being framed
Technology companies tend to frame the problem as one of acceleration: build generation, shorten permitting, modernize grids and allow large customers to bring new supply. That view correctly identifies infrastructure delay as a constraint. It can understate who bears the cost when assets are built for demand that does not arrive or when a large customer leaves before an investment has been recovered.
Utilities often frame the same issue as a planning challenge. They need credible forecasts years before equipment is ready, yet customers may resist disclosing commercial plans. Utilities can protect existing customers with deposits, minimum-payment contracts, collateral and exit fees. Those protections reduce risk, but poorly designed rules can also deter real projects or advantage incumbents that can lock up scarce capacity early.
Consumer advocates ask a different question: who pays? If network upgrades are socialized while the commercial upside remains private, households and smaller businesses may subsidize large loads. Developers answer that new facilities can expand the tax base, fund generation and improve grid utilization. Both effects are possible. The relevant evidence is found in each tariff and contract, not in a generic claim that data centres are either a gift or a burden.
Environmental arguments also divide. New demand can extend fossil generation where clean supply and transmission are slow. It can also anchor long-term contracts that finance wind, solar, storage, nuclear or geothermal projects. Matching annual renewable purchases does not by itself mean that a facility operates on carbon-free electricity every hour. The accounting method therefore changes the conclusion.
The background
The grid was not designed around a small number of projects asking for the power demand of a city on a compressed schedule. Large industrial loads are not new, but cloud and AI facilities can cluster geographically because fibre routes, latency, tax policy, land and existing network capacity pull developers toward the same places.
At the same time, the hardware is changing quickly. More efficient chips can reduce the electricity needed for one unit of computation. That does not guarantee lower total use. When computation becomes cheaper, companies may run more of it, offer new services or train larger systems. Efficiency changes the amount of useful work obtained from electricity; demand determines whether total consumption rises or falls.
This is why the queue itself has become an economic object. A connection offer can preserve an option on future development. If requirements are loose, speculative requests occupy engineering time and make the shortage look larger. If requirements are too strict, genuinely innovative projects may be excluded before their financing is complete. Ofgem’s intervention and FERC’s orders show two regulators trying to distinguish credible projects from optionality without pretending that uncertainty can be eliminated.
Who it touches
For households, the practical concern is not the abstract number of terawatt-hours. It is whether local bills include upgrades built mainly for large customers and whether generation keeps pace with demand. A transparent tariff should say who pays if a project is delayed, built smaller than promised or closed early.
For communities, the exchange is broader. Data centres can provide construction work and local tax revenue, but permanent employment may be limited relative to the land, water and power involved. The balance varies greatly by site. Communities need enforceable information about water use, noise, backup generators, tax concessions and the duration of promised payments.
For investors and workers in the AI supply chain, the grid introduces a timing problem. An order for accelerators, cooling equipment or transformers is not identical to an operating data centre. Backlogs can indicate durable demand, double ordering or projects waiting for power. Separating announced capital expenditure, equipment delivery, connection approval and actual energization helps prevent one milestone from being mistaken for another.
The deeper story
The most useful way to follow this story is to replace the single question “How fast is AI growing?” with a sequence of gates.
The first gate is demand quality: does a named customer have financing and a workload? The second is site control and permission. The third is electrical: is there a signed connection agreement, and who funds the upgrades? The fourth is equipment, including transformers and cooling. The fifth is operation: has the capacity actually been energized, and at what utilization? Each gate removes some projects that looked convincing at the announcement stage.
This framework also clarifies company claims. Chip orders describe one part of the chain. Utility forecasts describe another. Neither alone proves end-user adoption or financial returns. A robust analysis looks for agreement among physical milestones, customer contracts and cash flows.
Over the next year, watch for four signals. First, how many connection requests survive new financial-readiness tests? Second, do special tariffs protect other customers from underused infrastructure? Third, do forecasts converge as operators obtain better project data? Fourth, does new generation arrive with the load, or do regions extend older plants and accept higher congestion?
The durable insight is not that electricity replaces semiconductors as the only constraint. It is that AI infrastructure is a chain, and the tightest link can move. A valuation or policy built around unlimited, instantly available power is now missing a material variable.
Something to sit with
When you encounter the next large AI-infrastructure number, is it measured consumption, a forecast, a queue request, a signed contract or an energized facility?
If projected demand does not arrive on time, which contract says whether the developer, the utility’s shareholders or ordinary electricity customers carry the cost?
What evidence would persuade you that a region’s power shortage is a durable constraint rather than a queue distorted by duplicated or speculative projects?
Sources
- International Energy Agency — Energy and AI — https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
- Lawrence Berkeley National Laboratory — 2025 United States Data Center Energy Usage Report — https://eta.lbl.gov/publications/united-states-data-center-energy-2025
- Federal Energy Regulatory Commission — large-load integration action — https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targ...
- Ofgem — speculative data-centre projects and grid capacity — https://www.ofgem.gov.uk/press-release/ofgem-acts-free-grid-capacit...
- PJM — updated 20-year load forecast — https://insidelines.pjm.com/pjms-updated-20-year-forecast-continues...
We report facts from the sources above in our own words and link to the originals. Interpretation is ours, not theirs.
Which figure is a forecast rather than a measured present-day result?
The 945 TWh figure is the IEA's modelled 2030 projection. The other two describe reported actions or queue figures.
♻︎ Free to republish
Copy this HTML into your CMS. Credit line and licence are included. Republish our work — free
Every headline has a deeper story. This is ours.
What we are doing here