Digital Realty CTO weighs in on AI's insatiable thirst for power It's no secret the GPUs used to train and run generative AI models are power hungry little beasts. So hungry in fact that by some estimates, datacenters in the US could use as much as nine percent of the nation's electricity generation by the end of the decade. In light of this explosive growth, some have warned that AI …
Datacentres are going to slurp shit-tonnes of power and shit-tonnes of water in order to run completely useless stochastic bullshit generators that have become the latest fad.
The sector is "growing exponentially" without any real purpose, a bit like er, a tumour.
But it's OK because we will be using Renewables and Nuclear!
(Just don't mention that the Nuclear probably won't materialise, and when the renewables go behind a cloud we'll be gobbling up gas and Diesel, belching out emissions, and sending the price of energy through the roof - all so that your job can be replaced by a statistical model of your past actions and ideas, which is cheaper than you in the short term, but can't actually innovate in the long term)
Then there's the awkward question asked by the Reg hack about GPU financing and all of this being a potentially catastrophic bubble ...
I can imagine this guy selling railway bonds ...
Then there's the awkward question asked by the Reg hack about GPU financing and all of this being a potentially catastrophic bubble ...I can imagine this guy selling railway bonds ...
I think there are regulatory solutions. One sorta happened with Amazon and a DC PPA (Power Purchase Agreement) with Talen Energy in PA. Downside, that was a PPA for existing generating capacity. I think what should happen is whenever a new, power hungry DC is proposed, permits/consent should include a requirement to provide new generating capacity to cover the load. So a bit like housing development in the UK. Want to build 20 flats? X% are supposed to be affordable. Which is a nice idea, but doesn't really work because developers can buy out that requirement.
But it's something that could probably be securitised and turned into DC or AI bonds, with the bonds covering the generation costs. Those could be pooled, but would obviously impact the investment case in building a new DC if the PPAs were brought forward. So using Hinkley as an example, that's around £7bn per GW. The big DC builders/operators could in theory pool and invest in that kind of bond.. But probably won't want to given the highly speculative nature of AI. But it's also kind of normal for DC projects, ie they may have to pay for other infrastructure improvements, so why not also make power capacity part of the same planning consent. If you don't provide enough new generating capacity to cover the load, you don't get consent.
This would be a fairer solution given the way electricity generation is subsidised. The public are expected to take the risk of building new capacity, but don't really get any of the benefits if DC loads are soaking up the power and inflating bills. As investors, bond holders also stand to benefit given any surplus power generated under DC bonds would be sold on the open market.
"So using Hinkley as an example, that's around £7bn per GW"
In what universe? Most recent data puts Hinkley outturn at current prices at £46bn, for 3.2 GW. So north of £14bn per GW.
It's about the same £ per GW for offshore wind power once you adjust for the capacity factor. Eg, Hornsea 2 is about £6bn for 1.3 GW, but that's going to see about 46% capacity factor compared to about 92% for nuclear. Onshore wind is cheaper say £1.2bn per GW, but capacity factor is dismal 22%, so you'd need 4.5x as much plate capacity to match nuclear.
In what universe? Most recent data puts Hinkley outturn at current prices at £46bn, for 3.2 GW. So north of £14bn per GW.
Hmm.. I got those numbers from a quick search for any papers on Hinkley's costs published in the last year. Can't find it again but it was interesting in that it broke down most of it's costs into a per-kWh number. Which seemed an odd way of doing it. From another quick search, the answer seems to depend on how costs are calculated, ie construction costs vs construction + operating costs over the estimated lifespan, the latter obviously giving a much higher figure. But I found EDF's £46bn figure, and hopefully the NAO is crawling all over those numbers given they're far, far higher than Olkiluoto or Taishan for the same reactor design. So I'm curious if we're bailing out EDF (again), or they're just that incompetent.
But all part of the fun trying to make sense of this stuff and 'levelised costs' that are supposed to give a like-for-like comparison, but don't.
Also it doesn't really change the idea of user pays, and securitising that so that DC operators are made to invest in new capacity rather than having it subsidised by other energy users.
The last question concerning data centre of the future 10 years down the line. Feeding energy back to to the grid, providing heat to grow tomatoes....That's the Data Centre of Now, at least in Europe. Energy Aware DC's enabled by Eaton started 2-3 years ago and Heat exchange/transfer schemes have been widely used for the last decade, at least in Europe.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
