Wouldn't it be really nice if these values were required to be published, with strong legal oversight, and meaningful financial penalties should there be mistakes?
Why should I need to trust a corporate white paper/preprint from an AI company? Why should I need to trust a CEO from an AI company? What fine points are omitted?
But, let's do it this way. How does something with the energy demand of only a tenth of a kettle per day per person put such a strain on the UK's grid - a country which boils many kettles every day - that there's a demand to increase power generation by 50%?
"NESO received a total of 229 responses to its CFI. The CFI identified around 140 (50 GW) data centres, the majority of which are likely to receive a Gate 2 offer. 71 (around 20 GW) data centres reported they have achieved financial commitment with FID."
and
"Since November 2024, total contracted offers in the demand queue rose sharply from 41 GW (17 GW Transmission, 24 GW Distribution) to 125 GW (97 GW Transmission, 29 GW Distribution) in June 2025. For comparison, peak electricity
demand in GB on 11 February 2026 was 45 GW."
It's almost as if scaling up the median Gemini text prompt use from last year's preprint has very little bearing on the total AI power use.
After all, we know there are far more AI uses than the two presented options of "thirty Gemini queries per day" and "every adult makes one AI video a day".
A growth in power generation by 50% isn't even enough to meet the anticipated increase in demand for electric vehicles and appliances. Data center usage is barely a drop in the bucket, in comparison.
Also, there's much, much more going on in data centers than AI, so data center growth isn't necessarily AI growth.
First, the numbers I quoted were only for data centers: "The CFI identified around 140 (50 GW) data centres, the majority of which are likely to receive a Gate 2 offer. 71 (around 20 GW) data centres reported they have achieved financial commitment with FID."
That's 50 GW potential demand, of which I quoted only the 20 GW with financial commitment, to be conservative. That in a country where current peak demand was 45 GW, which is how I got the 50% number.
Second, the report says that the data center growth is primarily for AI: "The significant growth in data centres, driven by investment in artificial intelligence and the expectations for its economic and social potential, is a global challenge."
(What large amount of non-AI data center growth do you know about?)
Third, it highlights that while there will be other demands, these lag behind the more immediate demand for data center power:
"Data centres must be central to any solutions we implement. Data centre growth has brought into sharp focus the challenges within the demand connections process. However wider electrification of the economy and the associated rise in electricity demand, would, over time, have exposed these challenges. This would have made demand connections reform unavoidable."
There is no way to get from those numbers to conclude the AI power use per capita is "a tenth of a kettle of tea."
Thanks for this, it was part of what pushed me to dig further, and you were both partly right. Globally, data centres are under a tenth of electricity demand growth to 2030, with industry, cooling and transport far bigger, and AI-focused data centres used about 155 of the 485 TWh all data centres used in 2025. But AI was most of last year's growth, and in the US data centres are around half of all demand growth. Written up here: https://allyburns.com/musings/where-the-power-goes/
From the IEA report: "Simple text queries now typically consume less electricity than running a television over the same period of time."
Given that most people have smart televisions which gather and send information for, among other things, selling to advertisers, I can't help but wonder if their calculation is just for the TV or if it includes the data center overhead - which may well be running an AI model.
You wrote: 'My hunch going in was that a lot of this is waste: dozens of companies racing to build the same thing in secret, each burning energy so that one of them can win.'
Part of that waste is all the data scrapers draw power (from botnets running on residential proxies with less efficiency than a data center) -- and the scraped data isn't shared. The scrapers cause more demand on the servers, which need more power. The server operators deploy anti-bot protections like proof-of-work, which requires more power, and the anti-anti-bot methods which use real hardware or automated full browsers which again require more power.
Any idea what that power use might be?
BTW, you did a bit of a cheat there, with "The oldest big AI workload isn’t chat at all." If recommendation systems (or Bayesian spam filtering) count as AI then we've long known that AI for everyone is easily possible, with well less than a tenth of a kettle of tea per person per day.
If you're serious about including recommendation systems, you'll need to include the advertisement infrastructure as part of the AI overhead, as all use recommendation methods to figure out how much to bid for an advertisement on your screen. Or probably LLMs these days.
How much power is going into powering online advertisement?
9 comments
[ 0.33 ms ] story [ 5.5 ms ] threadWhy should I need to trust a corporate white paper/preprint from an AI company? Why should I need to trust a CEO from an AI company? What fine points are omitted?
But, let's do it this way. How does something with the energy demand of only a tenth of a kettle per day per person put such a strain on the UK's grid - a country which boils many kettles every day - that there's a demand to increase power generation by 50%?
Ofgem reported earlier this year, at https://www.ofgem.gov.uk/sites/default/files/2026-02/2026-02... :
"NESO received a total of 229 responses to its CFI. The CFI identified around 140 (50 GW) data centres, the majority of which are likely to receive a Gate 2 offer. 71 (around 20 GW) data centres reported they have achieved financial commitment with FID."
and
"Since November 2024, total contracted offers in the demand queue rose sharply from 41 GW (17 GW Transmission, 24 GW Distribution) to 125 GW (97 GW Transmission, 29 GW Distribution) in June 2025. For comparison, peak electricity demand in GB on 11 February 2026 was 45 GW."
It's almost as if scaling up the median Gemini text prompt use from last year's preprint has very little bearing on the total AI power use.
After all, we know there are far more AI uses than the two presented options of "thirty Gemini queries per day" and "every adult makes one AI video a day".
Also, there's much, much more going on in data centers than AI, so data center growth isn't necessarily AI growth.
First, the numbers I quoted were only for data centers: "The CFI identified around 140 (50 GW) data centres, the majority of which are likely to receive a Gate 2 offer. 71 (around 20 GW) data centres reported they have achieved financial commitment with FID."
That's 50 GW potential demand, of which I quoted only the 20 GW with financial commitment, to be conservative. That in a country where current peak demand was 45 GW, which is how I got the 50% number.
Second, the report says that the data center growth is primarily for AI: "The significant growth in data centres, driven by investment in artificial intelligence and the expectations for its economic and social potential, is a global challenge."
(What large amount of non-AI data center growth do you know about?)
Third, it highlights that while there will be other demands, these lag behind the more immediate demand for data center power:
"Data centres must be central to any solutions we implement. Data centre growth has brought into sharp focus the challenges within the demand connections process. However wider electrification of the economy and the associated rise in electricity demand, would, over time, have exposed these challenges. This would have made demand connections reform unavoidable."
There is no way to get from those numbers to conclude the AI power use per capita is "a tenth of a kettle of tea."
Given that most people have smart televisions which gather and send information for, among other things, selling to advertisers, I can't help but wonder if their calculation is just for the TV or if it includes the data center overhead - which may well be running an AI model.
You wrote: 'My hunch going in was that a lot of this is waste: dozens of companies racing to build the same thing in secret, each burning energy so that one of them can win.'
Part of that waste is all the data scrapers draw power (from botnets running on residential proxies with less efficiency than a data center) -- and the scraped data isn't shared. The scrapers cause more demand on the servers, which need more power. The server operators deploy anti-bot protections like proof-of-work, which requires more power, and the anti-anti-bot methods which use real hardware or automated full browsers which again require more power.
Any idea what that power use might be?
BTW, you did a bit of a cheat there, with "The oldest big AI workload isn’t chat at all." If recommendation systems (or Bayesian spam filtering) count as AI then we've long known that AI for everyone is easily possible, with well less than a tenth of a kettle of tea per person per day.
If you're serious about including recommendation systems, you'll need to include the advertisement infrastructure as part of the AI overhead, as all use recommendation methods to figure out how much to bid for an advertisement on your screen. Or probably LLMs these days.
How much power is going into powering online advertisement?