UK Telegraph Excerpt:
Speaking via a video link to global elites in Davos last week, Donald Trump said the US would need to double the amount of energy it produces to satisfy the ambitions of the AI hyper-scalers.
This was a far bigger estimate than even the highest such prediction I’d seen up until that point – that of the management consultants McKinsey, which forecast that global data-centre demand might grow by 240GW between 2023 and 2030, and that in the US they would absorb some 12pc of all the power generated.
That’s a long way below what Trump seems to imagine, but it would still require a massive upscaling of both US power generation and distribution. And already plans are well advanced to meet that demand.
Writing for BloombergNEF, the UK based clean energy entrepreneur Michael Liebreich says that most of the data centres used to train AI models currently in construction are in the 100MW and 250MW range, much bigger than anything that’s gone before – but what the tech giants really want are training centres of between 1GW and 2GW.
Recently announced in a blaze of publicity, Trump’s $500bn (£400bn) Stargate supercomputer complex would require 5GW. Amazon alone proposes to spend $150bn on data centres over the next 15 years.
All these plans now begin to look massively overblown. DeepSeek’s “Sputnik moment” also points to an already monumental, hype-driven misallocation of capital that threatens to drive significant losses and write-offs across the tech sector for some years to come.
Hundreds of billions of investment dollars are being sunk into AI, but so far for little or no return. For comparison, look back to the turn of the century dot-com and mobile telephone bubbles, which gave us a wonderful new technological infrastructure but also an eventual bust in which many people and organisations lost their shirts.
Some of the same excess can be seen in the rush to secure seemingly scarce energy supply for AI. In anticipation of explosive growth, the US state of Georgia alone is planning for 36.5 GW of new demand in the next decade, or around a half of current peak demand in the UK, including 19.9 GW in the next four years.
Elon Musk, Trump’s new confidante, noted in March last year that “we have silicon shortage today, a transformer shortage in about a year and electricity shortages in about two years”.
Mark Zuckerberg, the chief executive of Meta, has claimed his company would build more data centres if only it could get more power.
Sam Altman, the founder of OpenAI, likewise: “We still don’t appreciate the energy needs of this technology … We need fusion or we need radically cheaper solar plus storage, or something.”
Or maybe not. One of the abiding truths about technological innovation is that what starts off as clunky and barely worth the time and effort required to use it rapidly generates big improvements in efficiency, driving down costs and ease of use, and eventually creating a virtuous circle of rising supply and demand.
As Liebreich points out, some of the projections for AI, to the effect that it threatens to gobble up an ever larger share of the world’s supply of energy, were already looking far-fetched even before the DeepSeek game-changer came along.
In fact, “the average Power Usage Effectiveness (PUE) – the ratio of total power used in a data centre to the power used by its servers – dropped to 1.5 in 2021 from 2.7 in 2007, with the best data centres now delivering PUEs as low as 1.1”, Liebreich notes.
An early version of such reductio ad absurdum logic was on show in what became known as “the great horse manure crisis of 1894”, when such was the dependence on horse-driven carriages for transport that The Times predicted that every street in London would in 50 years’ time be buried under nine feet of manure.
Then along came the automobile, rendering the forecast ridiculous. Technology eventually finds a way. The comparison that Greg Jackson, founder of the UK’s Octopus Energy, likes to use is with the iPhone, whose battery power has not improved that much since it was first launched in 2007, but now supports computing power and connectivity thousands of times greater than the original version.
The same may well be true of AI. The price of a unit of electricity may not change very much, but it will buy you an awful lot more. Nor should we underestimate the power of AI to help mitigate energy demand across the board, by for instance optimising grid utilisation or enabling the invention of new, energy saving technologies, says Jackson.