Thank you for explaining that, James. I think there's another denominator missing from the discourse, quite possibly because the public only started hearing about datacentres† in the context of AI, is the number of datacentres needed for non-AI uses.
Searching the web requires a datacentre even if no AI is used in the search. Watching a video online or using a streaming service on a TV requires a datacentre to serve that video. Everything anybody does on the internet (or a device does for them) requires a datacentre at the other end of the connection.
LLMs clearly require more datacentres than we've had before, but it doesn't follow that no AI would mean no datacentres. It'd be useful to have a vague sense of ‘new for AI’ datacentre capacity as a proportion or multiplier of existing datacentres for other things.
† I used to work in one, and I'm pretty sure we always spelt it without the space. And each time a new datacentre was built or an existing one refitted, it would be done more efficiently than before — the main driver being the business wanting to reduce its electricity bill, particularly on cooling. For instance, we inherited a datacentre in which all the computers were facing the same way; that was redone with rows of them in opposite directions, so instead of hot air out the back of one row blowing into the intake at the front of the next, there were alternating hot and cool aisles. Then doors were put on the ends of the cool aisles, and the aircon could be focused on those rather than the room as a whole.
Similar to your comment on electricity supplies becoming less carbon-based over future decades, it feels weird to presume innovation in datacentre efficiency will suddenly end. I worked for a medium-sized organization; the big tech players with lots of massive datacentres have a financial incentive to find ways of making them as efficient as possible, even if they don't care about the environment per se.
I looked into this and think your explanation actually undersells how much of an exaggeration the electricity use is. The amount of electricity used is being exaggerated in addition to the decarbonisation of the grid you mention. Here's where ChatGPT and me ended up:
"The EIA assumes roughly full utilisation of the site and, more importantly, applies a fixed electricity carbon intensity of about 140gCO₂e/kWh all the way to 2103, explicitly ignoring future grid decarbonisation. The site’s own newer modelling says 50% server utilisation is the “likely” case, with electricity demand of about 2.65TWh/year, versus roughly 6.9TWh/year in the worst case. Using current DESNZ/NESO projections, grid carbon intensity is around 2g/kWh by the early 2040s and about 1.3g/kWh by 2050. On that basis, lifetime electricity emissions in the likely case are roughly 0.2MtCO₂e, versus about 22Mt if you freeze the grid at 140g/kWh — a difference of roughly 100×."
I think calling this "hysterical" is too kind - the Guardian is dishonestly misleading readers.
The developers certainly did have a choice but seem to have appointed a rubbish consultant. The RICS WLCA methodology includes for grid decarbonisation.
Wait, they applied a 15% uncertainty adjustment, but they adjustment only goes *up*? That's fine if you're losing an upper bound, but if doing that, you need to explicitly say "this is an upper bound, NOT the expected amount. We expect it to be less than this."
James! The Grauniad "needlessly hysterical? "Surely, you can't be surprised. It's more or less the house style, and data centers are the perfect bandwagon for their needless hysteria.
Thank you for explaining that, James. I think there's another denominator missing from the discourse, quite possibly because the public only started hearing about datacentres† in the context of AI, is the number of datacentres needed for non-AI uses.
Searching the web requires a datacentre even if no AI is used in the search. Watching a video online or using a streaming service on a TV requires a datacentre to serve that video. Everything anybody does on the internet (or a device does for them) requires a datacentre at the other end of the connection.
LLMs clearly require more datacentres than we've had before, but it doesn't follow that no AI would mean no datacentres. It'd be useful to have a vague sense of ‘new for AI’ datacentre capacity as a proportion or multiplier of existing datacentres for other things.
† I used to work in one, and I'm pretty sure we always spelt it without the space. And each time a new datacentre was built or an existing one refitted, it would be done more efficiently than before — the main driver being the business wanting to reduce its electricity bill, particularly on cooling. For instance, we inherited a datacentre in which all the computers were facing the same way; that was redone with rows of them in opposite directions, so instead of hot air out the back of one row blowing into the intake at the front of the next, there were alternating hot and cool aisles. Then doors were put on the ends of the cool aisles, and the aircon could be focused on those rather than the room as a whole.
Similar to your comment on electricity supplies becoming less carbon-based over future decades, it feels weird to presume innovation in datacentre efficiency will suddenly end. I worked for a medium-sized organization; the big tech players with lots of massive datacentres have a financial incentive to find ways of making them as efficient as possible, even if they don't care about the environment per se.
I looked into this and think your explanation actually undersells how much of an exaggeration the electricity use is. The amount of electricity used is being exaggerated in addition to the decarbonisation of the grid you mention. Here's where ChatGPT and me ended up:
"The EIA assumes roughly full utilisation of the site and, more importantly, applies a fixed electricity carbon intensity of about 140gCO₂e/kWh all the way to 2103, explicitly ignoring future grid decarbonisation. The site’s own newer modelling says 50% server utilisation is the “likely” case, with electricity demand of about 2.65TWh/year, versus roughly 6.9TWh/year in the worst case. Using current DESNZ/NESO projections, grid carbon intensity is around 2g/kWh by the early 2040s and about 1.3g/kWh by 2050. On that basis, lifetime electricity emissions in the likely case are roughly 0.2MtCO₂e, versus about 22Mt if you freeze the grid at 140g/kWh — a difference of roughly 100×."
I think calling this "hysterical" is too kind - the Guardian is dishonestly misleading readers.
That's astonishing! Amazing work!
SAP is for residential assessments. For general carbon accounting you want to use the government greenhouse gas reporting figures, which in 2026 were 0.129 kGCO2/kWh... (https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2026)
The Treasury Green Book includes consensus predictions on grid decarb as well: https://assets.publishing.service.gov.uk/media/6567994fcc1ec5000d8eef17/data-tables-1-19.xlsx
The developers certainly did have a choice but seem to have appointed a rubbish consultant. The RICS WLCA methodology includes for grid decarbonisation.
Wait, they applied a 15% uncertainty adjustment, but they adjustment only goes *up*? That's fine if you're losing an upper bound, but if doing that, you need to explicitly say "this is an upper bound, NOT the expected amount. We expect it to be less than this."
James! The Grauniad "needlessly hysterical? "Surely, you can't be surprised. It's more or less the house style, and data centers are the perfect bandwagon for their needless hysteria.