AI can do anything. It’s going to transform the global economy. It’s going to give us the productivity boom we have awaited for so long – and in doing all our boring jobs, give us the leisure time economists predicted we would have decades ago. It’s going to save lives through early diagnoses by “smelling” the chemicals associated with diabetes and cancer. It will drive our autonomous cars, predict and prevent flight delays, design new types of climate-resistant rice and possibly even do the one thing that no amount of money or actual people has yet been able to do – save the UK’s National Health Service. The list is endless. But you get the idea. By next year, says Elon Musk on his X, “AI will probably be smarter than any single human. By 2029 AI is probably smarter than all humans combined.”
But here’s the catch: AI isn’t magic. It can’t perform its miracles — or do anything — without electricity; lots and lots of electricity. No one is entirely sure how much, but a much quoted recent estimate comes in at around the same as Sweden uses each year or 0.5% of total world electricity use. Another way to get a sense of the scale is to look at the size of data centers. According to the latest data center report from the commercial-property firm JLL, it is “not uncommon” for developers today to announce new-build data centers with 10 times the load of only a few years ago.
So while we can’t say exactly how much electricity AI will take, we do know it’s going to require a lot more generation and transmission capacity. That’s particularly the case given the other two shifts underway – the attempt to green the grid with renewable energy and the effort to electrify more and more energy use in order to green that too. The European Union expects demand for electricity to rise 60% by 2030.
That capacity doesn’t yet exist. US electricity generation has been more or less flat since the early 2000s for example. This, says JP Morgan Asset Management’s Michael Cembalest in a recent report, hasn’t been much of a big deal: “The energy needs of a growing population have been offset by improving energy efficiency.” However, the arrival of AI, with its ever-expanding list of potentially life-changing applications, its unexpectedly rapid adoption — ChatGPT became the fastest growing app in internet history — and its rapacious demand for energy changes that. Now it is a big deal.
To meet demand in the US, and in the UK where the grid is just as inadequate, involves a new world of power generation and many billions to be spent on the grid needed to carry that power. Yet, look to the growth in high-voltage direct lines (the ones that optimize the transmission of renewables from their usually remote locations); you won’t find much. In the UK, National Grid Plc admits to a “bottleneck” in connecting generation projects. Look to the US, and you can see that in the early 2000s there were rarely more than 100 gigawatts of capacity in the interconnection queue at regional transmission firm PJM Interconnection LLC, for example, says Cembalest. Now there are more like 250. Not all these projects will get approval and complete. But you get the idea. The grid: might it be where our AI dreams go to die?
There are solutions.
First, there’s the Elon option: AI might itself produce a solution we haven’t yet thought of; perhaps fission, fusion or even the beaming of solar energy from space using giant mirrors are closer than we think. AI might also bring new efficiency into the grid - stabilizing it with better forecasts of the load from and intermittency of renewable sources, better demand forecasting and predictive maintenance that cuts downtime. It is also entirely reasonable to imagine that the productivity gains elsewhere thanks to AI will offset some of the increased demand it creates. This report from Google and Boston Consulting Group claims AI can mitigate 5-10% of global emissions by 2030.
Second, the regulatory option: stop AI using lots of energy. A report produced for the US government last month notes that the rise of AI and AGI (artificial general intelligence) has “profound implications for democratic governance and global security” – and not in a good way. The report suggests a fairly broad range of mitigations, one being making it illegal to train AI models using more than a low-ish level of computing power. The EU’s new AI Act (among many other things) sets out requirements that compel various AI systems to offer full transparency on energy consumption.
Finally there is, the nuclear option. It should be no surprise to regulators or power companies that in the wake of an insistence that we switch to using electricity for everything, demand for electricity is rising fast, along with massive investment plans. But the large-scale adoption of nuclear (AI-assisted, of course) would solve all the supply problems – and if built locally, much of the transmission problem too: There is no need to place small modular reactors miles out in the North Sea, The analysts at Doomberg reckon that this will be the final push to a nuclear renaissance – for the simple reason that it is the obvious and, “for all practical purposes,” the only sustainable way to create the electricity to work with our tendency to grow our computing power exponentially. In a possible sign of things to come, a few weeks ago Amazon Web Services announced that its new data center in Pennsylvania is to be powered directly by an existing adjacent nuclear power station. The AI revolution might be coming hand in hand with a power revolution — something for investors in uranium to watch.
A message from Advisor Perspectives and VettaFi: To learn more about this and other topics, check out our podcasts.
Bloomberg News provided this article. For more articles like this please visit
bloomberg.com.
Read more articles by Merryn Somerset Webb
