--- title: "Why AI's power appetite is becoming an electricity problem" description: "AI companies now describe their build-outs in gigawatts, units of electrical power, not in numbers of chips. That shift in language is the story: the constraint on artificial intelligence is increasingly the electricity to run it. US power demand hit a record 4.2 trillion kilowatt-hours in 2025, and computing is a growing reason why." category: "Economy" category_url: https://boursel.com/category/economy author: "Hannah Blackwood" published: 2026-07-23T16:16:00.000Z updated: 2026-07-23T16:16:00.000Z canonical: https://boursel.com/article/why-ai-s-power-appetite-is-becoming-an-electricity-problem tags: ["ai", "electricity", "data-centers", "energy"] --- # Why AI's power appetite is becoming an electricity problem AI companies now describe their build-outs in gigawatts, units of electrical power, not in numbers of chips. That shift in language is the story: the constraint on artificial intelligence is increasingly the electricity to run it. US power demand hit a record 4.2 trillion kilowatt-hours in 2025, and computing is a growing reason why. There is a tell in how the artificial-intelligence industry now talks about itself. When AMD and Anthropic announced a partnership this week, the headline number was not a count of chips but a measure of electrical power: a deployment of [up to 2 gigawatts of computing capacity](https://ir.amd.com/news-events/press-releases). When the biggest AI deals are sized in gigawatts, the same unit used for power stations, it tells you where the real bottleneck now sits. Building the models is increasingly a question of finding the electricity to run them. ## What a gigawatt of computing means A gigawatt is a unit of power, the rate at which energy is used. One gigawatt is roughly the output of a large power station, enough to supply a substantial city. So when an AI project is described as two gigawatts, it is saying its data centers will draw power on the scale of a couple of sizeable generating plants, continuously. That framing matters because it reveals the binding constraint. For years the scarce input in AI was the advanced chips themselves. Now, with buyers committing to accelerators years in advance, the harder problem is often securing the electricity and the grid connections to power them, and the cooling to keep them running. The chip is no longer the whole story; the power to feed it is. ## The demand is landing on a grid already at record highs This is arriving on top of an electricity system that is already stretched. According to the [US Energy Information Administration](https://www.eia.gov/energyexplained/electricity/use-of-electricity.php), total US electricity consumption in 2025 was "about 4.20 trillion kWh, the highest recorded," and 14 times the level of 1950. Historically that demand splits fairly evenly between the residential (37.3%) and commercial (36.8%) sectors, with industry taking the rest (25.7%). The relevant trend for AI is inside the commercial figure: the EIA notes that "computers and office equipment" account for the largest share of commercial electricity use, and that this share "has increased over time" even as lighting has become more efficient. Large data centers push that computing demand to an industrial scale, concentrated in specific locations. Crucially, the EIA projects overall US electricity demand to grow "about 2% from 2025 through 2050", a pace that already assumes a meaningful uplift from new sources of demand. When a single AI campus can require the output of a power plant, that kind of load does not fit quietly into a 2%-a-year world; it has to be planned for, generated and delivered. ## Why it matters for money, not just megawatts For a financial audience, the AI-power story cuts several ways at once. It is a **utilities and infrastructure story.** Sudden, large, concentrated demand for round-the-clock power is a boon for whoever can supply it, generators, grid operators, and the firms building transmission and cooling. It also raises hard questions about who pays to upgrade the grid, and whether ordinary consumers end up subsidizing connections for hyperscale data centers. It is an **energy-price story.** Layering gigawatts of new, steady demand onto a system already at record consumption puts upward pressure on wholesale power prices in affected regions, and revives interest in every form of generation that can run reliably, from natural gas to nuclear. And it is a **risk to the AI thesis itself.** If compute is now gated by power, then the growth plans of the AI leaders depend on things outside a data center's walls: permitting, grid capacity, and the pace at which new generation can be built. A chip order can be filled in months; a new power connection can take years. That mismatch is now a real constraint on how fast the industry can actually scale. ## The takeaway The move from counting chips to counting gigawatts is not just a change of vocabulary; it is the industry admitting where its limits lie. AI's expansion has become inseparable from the electricity system, and that system is already running at record output with only modest growth planned. None of this is a forecast or investment advice, but it reframes what to watch. The AI story is now also an energy story: the winners will include not only the companies that build the best models, but the ones that can secure the power to run them, and the utilities and grids that ultimately have to deliver it.