Chips ship in a quarter. Power for them can take a decade

Chips are the part of the AI boom that can be bought. Power is the part that has to be built, and that is why the race will be decided at the grid connection and not at the chip counter.

AMD gave the boom a fresh number on November 3. The chip maker reported record quarterly revenue of $13.4 billion, up 45% on a year earlier, of which $8.5 billion came from its data-centre business, up 97% ^[2]. The company said demand for its EPYC server processors and the launch of its Instinct MI450 Series GPUs drove the result ^[2]. It forecast about $17 billion for the next quarter, give or take 3%, which at the midpoint would be growth of about 66% on the year-earlier quarter ^[2]. That forecast is the company's own outlook, and AMD's release warns that actual results could differ materially ^[2].

It is an astonishing run of numbers. It is also silent on one thing. The release does not say how much electricity the chips AMD sells will need once they are switched on ^[2]. That is not a criticism of AMD, which sells silicon and reports revenue. It is a clue to where the argument about AI has been looking, and where it has not.

Two clocks

The AI industry runs on two clocks, and they tick at very different speeds. The International Energy Agency says a data centre can be built in one to three years, while planning, permitting and completing the grid infrastructure to feed it can take five to 15 ^[4]. A chip order can be placed this quarter. A transmission line is a decade of planning, permitting and construction.

The queue that results is already enormous. The agency counts more than 2,500 gigawatts of projects stalled in grid connection queues worldwide, a pile that includes power plants and storage as well as big electricity users like data centres ^[4][6]. To meet forecast demand, it estimates, annual grid investment would have to rise by about 50% by 2030, from roughly $400 billion a year today ^[4][6].

Set that beside the demand figures. The IEA projects global data-centre electricity use to nearly double, from 485 terawatt-hours in 2025 to 950 in 2030, with the use of AI-focused data centres tripling ^[3]. Last year the total grew 17%, and the AI-focused part grew 50% ^[3]. The demand is not a forecast waiting to arrive. It is already arriving.

What a queue does to a builder

The place to see the squeeze is Northern Virginia. The US Energy Information Administration reports that commercial electricity sales in Virginia rose by nearly 30 million megawatt-hours between 2019 and 2025, with data centres a major driver ^[5]. It describes the Dominion zone of the PJM grid as the world's largest concentration of data centres ^[5]. PJM is the regional operator that runs the high-voltage grid across much of the mid-Atlantic and neighbouring states, Virginia included. Summer peak load in that zone hit 23,905 megawatts in 2025, 23% above 2019, and PJM forecasts average growth in summer peak load of 5.4% a year for the next decade ^[5].

Now imagine being a developer there with the buildings, the money and an order book of chips. TechCrunch, summarising a New York Times investigation, reports what some of them are doing instead of waiting. They are buying converted engines from retired aircraft, rebuilt to burn natural gas, and generating their own electricity ^[1].

The arithmetic that makes this tempting is plain. The energy analyst Audun Martinsen told the Times that a data centre using jet-engine units could be running in 24 months, while a 100-megawatt grid connection in Northern Virginia takes about seven years ^[1]. Seven years against two. You can see why someone reaches for a jet engine.

The Times, according to TechCrunch, found that xAI had installed eight converted jet-engine turbines in Mississippi to help power its Colossus 2 data centre, based on state permits and satellite imagery ^[1]. TechCrunch's account carries no response from xAI, so the practice should be read as reported, not as proven across the industry. How widespread it is has not been counted.

The cost is the part the speed hides. TechCrunch reports that converted jet engines are less efficient and more polluting than turbines designed for conventional gas plants, and that more of them are needed to produce the same power ^[1]. An industry that is short of electricity is, in this corner, solving its scheduling problem with a dirtier way of making it.

The best case for the optimists

The optimists have a real argument and it deserves its full weight. Energy used per AI task has fallen by at least an order of magnitude a year in recent years, according to the IEA ^[3]. Hardware and software keep getting leaner. If that continued, a forecast of doubled demand could look silly by 2030.

The same agency adds the other half of the sentence. Newer uses such as video generation, reasoning and agentic tasks, where AI systems carry out multi-step jobs on their own, can use hundreds or thousands of times more energy per query than simple text generation ^[3]. Efficiency has so far been spent on doing more, not on doing the same for less. The honest position is that nobody knows the size of the 2030 load. But when you do not know how big a thing will be, you do not decline to build the road. You plan for the range.

And the workaround is not the free lunch it looks like. The IEA says reliable on-site gas generation for data-centre loads that swing up and down may need capacity 30% to 70% above the demand itself, and that a shortage of gas turbines means on-site power is not necessarily a faster route to market at scale ^[3]. The jet engine is a fix for one developer in one year. It is not a policy for an industry.

What I think follows

Here is my view. The chip makers are not going to solve this, and it would be unfair to ask them to. A company that sells processors reports revenue, and its release says nothing of the wattage of the buildings its customers have yet to switch on ^[2]. The people who can move the clock are utilities, grid regulators and governments, and they move slowly by design, because the cost of a bad line is paid for decades.

So the industry should do what it rarely does in public. It should talk about power with the seriousness it gives to chips: publish consistent estimates of the electricity its sites will need, say where those sites are and when they expect to connect, and back the dull work of grid upgrades with the same enthusiasm it brings to a product launch.

There are quicker routes than a jet engine. The IEA says grid-enhancing technologies, meaning equipment and software that let existing power lines carry more electricity safely, together with regulatory changes could free up capacity on existing networks. It estimates that such technology upgrades could connect 450 to 700 gigawatts of advanced-stage projects, with the caveat that these are global estimates and not project-specific ^[4]. That is the unglamorous prize. Whether it can deliver power sooner and at lower environmental cost than gas on site is still an open question, but it is the one worth testing before the rebuilt engines multiply.

The next AMD quarter will be judged on how many chips it ships. The better question is how many of them will be plugged in on time. On the evidence of the queues, the answer will be set less in a design lab than in a permitting office, and the companies that treat electricity as someone else's department are the ones who will learn it last.

Watch for the next round of data-centre announcements. The ones that name a power source, a connection date and a number of megawatts are the ones to believe.