The Power Wall: Why Electricity, Not Chips, Will Decide the AI Race

When chip stocks had their worst day in months on 22 October, most of the commentary focused on Lam Research’s warning that clean-room and supply constraints would run into 2027, and on Meta’s first negative free cash flow in more than a decade. One detail drew less attention. On the same afternoon that the Philadelphia Semiconductor Index fell almost 7%, US utilities rose more than 2%. Investors selling the AI trade were, in effect, buying the companies that keep the lights on inside it.

That rotation captures a shift that engineers and site developers have been talking about for a couple of years and that boardrooms are only now absorbing. You can order a cluster of accelerators and have it delivered in months. You cannot order 300 megawatts of firm electricity the same way. In a growing number of markets, the wait for power is measured in years, and it has become the single biggest variable in the cost and timing of AI capacity.

The Numbers Behind the Squeeze

The International Energy Agency estimates that data centres consumed around 485 terawatt hours of electricity worldwide in 2025, roughly 1.7% of global demand. Its central projection has that figure close to doubling, to about 950 TWh, by 2030. Some private forecasts are more aggressive still. On a global scale those numbers sound manageable. Electricity demand grows every year, and 2% of it is not an existential threat to anyone.

The trouble is that data centres do not spread themselves evenly. They cluster where fibre, land, tax incentives and existing campuses already sit, which means the strain lands on a handful of local grids. In Ireland, data centres now use more than a fifth of all metered electricity, and the grid operator has effectively rationed new connections around Dublin. Northern Virginia, the largest data centre market on the planet, has pushed its utility into one of the steepest load forecasts in the United States. A share of national demand that looks trivial can be a crisis for a single county.

AI also changes the shape of the load. A conventional enterprise data hall might draw a few kilowatts per rack. Racks built around the latest accelerators are designed for well over 100 kW, with liquid cooling to match. Training runs pull steady, enormous loads for weeks. The result is that a single new AI campus can ask a utility for as much power as a mid-sized city, and ask for it within two or three years.

The Queue Is the Real Bottleneck

Getting that power means joining a queue, and the queues have become extraordinary. ERCOT, which runs most of the Texas grid, has received large-load connection requests from data centres totalling around 160 gigawatts. To put that in context, the entire Texas grid peaks at a little over 85 GW on the hottest summer days. The IEA rightly notes that many of these requests are speculative, with developers filing in several places and expecting to drop most of them. Even so, the sheer volume slows down the genuine projects stuck behind the tyre kickers.

The physical supply chain is no faster. Large power transformers, the unglamorous steel boxes that step voltage down from transmission lines, now carry lead times of two years or more in the US. Switchgear, high-voltage breakers and specialist cable are similarly stretched. Gas turbine makers have reported order books reaching towards the end of the decade, which is why a developer who wants to build its own generation on site often finds the same wall waiting there.

Analysts tracking large US projects suggest only about a third of announced capacity is actually under construction, with the rest waiting on transmission upgrades or interconnection studies. For investors, this is the gap between the capex numbers announced on earnings calls and the capacity that can actually be switched on. For enterprise customers, it is the reason a cloud region can quietly run short of GPU instances even as its parent company promises record spending.

Nuclear’s Unexpected Second Act

Few people predicted in 2020 that big technology companies would become the most enthusiastic buyers of nuclear power in the Western world. Yet that is where the industry has landed. Microsoft signed a 20-year agreement with Constellation Energy to restart Unit 1 at Three Mile Island, now renamed the Crane Clean Energy Center. Amazon has backed X-energy’s small modular reactor programme and bought a data centre campus next to Talen’s Susquehanna plant in Pennsylvania. Google agreed to buy power from a fleet of Kairos Power reactors. Meta has assembled agreements with Vistra, Oklo and TerraPower that together cover more than 6 GW of nuclear capacity.

The appeal is simple. Nuclear plants run around the clock, produce no carbon and occupy little land. For a company that has promised net zero and needs firm power, nothing else ticks every box. The catch is timing. The first commercial small modular reactors in the US and Europe are not expected much before 2030, and first-of-a-kind nuclear projects have a long history of running late and over budget. Most of the nuclear electricity that will actually reach data centres this decade will come from existing plants, through power purchase agreements, uprates and the occasional restart.

That has a knock-on effect the technology firms would prefer not to discuss. When a hyperscaler contracts the output of an existing reactor, that clean power is no longer available to everyone else on the grid. The emissions accounting looks good for the buyer while the wider system may end up burning more gas to fill the gap. Regulators in several US states have started asking exactly that question.

The Bridge Is Built From Gas

Between today’s queues and tomorrow’s reactors sits a decidedly less fashionable answer: natural gas. Meta’s huge Hyperion campus in Louisiana is tied to new gas-fired plants built by the local utility. Elon Musk’s xAI ran mobile gas turbines at its Memphis site while it waited for grid capacity, prompting complaints from local residents about air quality. Across Texas and the southern US, developers are designing campuses around on-site gas generation from day one.

Behind-the-meter generation, where a data centre builds its own power plant and connects to the grid only as a backup, has gone from niche to mainstream in about 18 months. Fuel cells, large battery systems and solar paired with storage fill in parts of the picture. None of them yet offers the combination of speed, scale and reliability that a gas turbine does, which is why the climate pledges of the largest technology firms are under more strain than at any point since they were made.

A New Map of Where AI Gets Built

Power scarcity is redrawing the geography of AI faster than any trade policy. Texas has become the default destination for American mega campuses because it combines cheap land, abundant gas, fast permitting and its own grid. The Gulf states have pitched themselves as the place where power is simply not a constraint. The UAE and the US agreed in 2025 on a 5 GW AI campus in Abu Dhabi, and Saudi Arabia’s Humain is building on the kingdom’s ability to generate electricity cheaply at scale. The Nordics sell cold air, hydropower and political stability.

India is a more complicated story. Google has committed $15 billion to an AI hub in Visakhapatnam, and Reliance is building large AI capacity in Jamnagar alongside its energy complex. India’s grid has added capacity at an impressive pace, but it still leans heavily on coal and on imported LNG for flexibility. As we noted in our analysis of the Strait of Hormuz, that makes Indian power prices sensitive to Gulf shipping risk in a way that a Norwegian hydro plant never will be.

Britain and much of continental Europe sit awkwardly in this race. Electricity is expensive, connection queues are long, and planning is slow. The UK has designated AI Growth Zones and reformed its connection queue to push stalled projects aside, which should help. But a country that cannot offer large blocks of affordable power within a few years will struggle to host the biggest training clusters, however good its universities are. Energy policy has quietly become AI policy, and the governments that understand this first will win the investment.

Who Gains When Power Is Scarce

For investors, the clearest winners so far are the companies selling picks and shovels to the grid. Makers of transformers, switchgear and high-voltage equipment such as Siemens Energy, Hitachi Energy, Schneider Electric and Eaton are sitting on record order books. Cooling specialists like Vertiv have benefited as liquid cooling moves from experiment to standard. Independent power producers with nuclear or gas fleets, including Constellation, Vistra and Talen, have been re-rated from sleepy utilities into AI infrastructure stocks.

Regulated utilities are a more mixed bet. Higher demand means more investment on which they can earn a return, but it also brings political risk. When data centres arrive faster than new generation, everyone else pays more. Capacity prices in PJM, the grid that covers Virginia and much of the mid-Atlantic, rose roughly ninefold in the auction for 2025/26, and household bills followed. Several US states are now writing tariffs that force large new loads to commit to long contracts or fund their own grid upgrades. Expect that model to spread.

What Boards Should Do Now

The first lesson applies well beyond the technology sector. Any company planning significant AI deployment should treat energy as part of the AI business case rather than a line hidden inside a cloud bill. Ask your cloud providers which regions have headroom, how they price capacity in constrained markets and what happens to your service levels if a region runs short. The answers are often less reassuring than the marketing.

Second, companies that operate factories, warehouses or offices in data centre hotspots need to look at their own exposure. Connection queues do not discriminate between a hyperscaler and a manufacturer wanting to electrify a production line. Firms in Northern Virginia, Dublin, Frankfurt or parts of Texas may find their own expansion plans waiting behind AI projects, and their power bills rising to pay for grid upgrades they did not ask for.

Third, efficiency is now a strategic lever, not a sustainability talking point. Every gain in chip and software efficiency stretches the power a company can secure. The growing focus on the cost of running AI models, rather than training them, is partly an energy story. A model that delivers the same answer for half the compute delivers it for half the electricity.

Finally, boards should read capex announcements with a sceptical eye. A promise to spend tens of billions on AI infrastructure means little if the power to run it is three years away. When assessing suppliers, partners or investments, the more useful question is not how many chips they have ordered. It is how many megawatts they have secured, and on what timetable.

Where This Leaves the AI Race

The chip shortage of 2023 and 2024 was painful but temporary. Factories expanded and supply caught up. The power constraint is different in kind. Grids are built on decade-long planning cycles, permitted by local politics and financed by ratepayers who vote. Nobody can wish a transmission line into existence, however much money they raise.

That does not mean the AI build-out stops. It means it slows, relocates and becomes more expensive in ways that favour companies with patient capital, energy expertise and good relationships with utilities and regulators. The winners of the next phase may look less like software start-ups and more like the industrial conglomerates of the last century: owning generation, managing grids and thinking in decades. For an industry used to moving fast, that is an uncomfortable adjustment, and one that every business depending on AI will feel in its costs.