Artificial intelligence is not virtual, clean, or weightless. It has a rapidly escalating physical cost in electricity, water, and emissions—and ordinary people will pay the price.
Research shows that AI data centres could soon consume electricity on the scale of entire nations. At the same time, AI cooling systems are diverting vast quantities of water in a world already facing severe shortages.
This video asks the questions politicians are refusing to confront: who pays for AI’s energy and water use, who profits, and whether unlimited AI growth is compatible with planetary limits, democratic accountability, and basic human needs.
AI may promise growth—but at what cost, and to whom?
Richard J Murphy gains the benefit: he uses AI to produce his slop after all. At what cost? The diminution of the collective IQ of the species, obviously.
Generally speaking, I think the people running the data centres will pay for the electricity and water.
Hope this helps.
“Research shows that AI data centres could soon consume electricity on the scale of entire nations”
So do Christmas lights… let’s ban Christmas.
Don’t give him ideas. No, he really will!
Yes, I do sometimes think he sees himself as a second Oliver Cromwell (who was from somewhere not far from Ely).
Even allowing for it being Murphy, this is galling. An intellectual and emotional swindler, who uses ChatGPT to write much of his drivel, condemns AI in the style of every Leveller and Luddite in history. And his supremely credulous victims suck it all up yet again.
Ah, but once AI is controlled by a government Commissioner, its use will be restricted to approved persons.
Looking at the spelling & syntax, I’d guess he’s used ChatGPT to write this. Raw Spud isn’t as polished. There was a piece Tim quoted yesterday. Had his trademark misspellings, typos, catch-as-catch-can punctuation & jumbled thoughts was undoubtedly the real thing. Spud actually doesn’t write very well. A lot of it’s stream of consciousness rather than rational composing. And he obviously doesn’t often read what he’s written before posting.
Even if you accept the premise, anybody who can invent clever ways to reduce the energy usage will make a lot of money. I don’t really accept the premise. We generate energy to do stuff people want. If they want AI we use it to provide AI.
.
The thing with AI is that most of the people talking about it have no grasp of what levels of energy we’re talking about.
“AI uses lots of energy”. Compared to what?
Yes, compared to things like getting a bank balance. But compared to say, driving a car, heating a house? No. Running an automated speed camera (which is a trained AI model) uses a lot less energy than driving 2 police officers out to a spot. And that’s before we get into the cost of human labour.
What’s the energy requirements for a night in a bar? Getting to the bar, heating the bar, the energy needed to distill vodka. Is that more or less energy than Tinder? Which is why (along with all the labour) it’s about the same price to get Tinder for a month than a night in a bar. “Pubs are dying”. Yeah, partly because people found a more efficient way to find a shag.
No, the energy demand for AI is humungus:
AI data center power consumption is skyrocketing, driven by energy-intensive GPUs for model training, with global usage projected to more than double to ~945 TWh by 2030 (from ~415 TWh in 2024),
Not even 1.21 jiggawatts, it’s hundreds of terawatt hours per annum.
it’s about the same price to get Tinder for a month than a night in a bar.
Grim. Remember what they took from you.
Isn’t this is the technocratic “trends are fixed” fallacy though? If you’re right that it’s a bubble then it won’t happen. If more efficient technology comes along it won’t happen etc. Even if it does happen then it’s 3% of global power generation. I think the sky will stay up.
OTOH we could have blackouts tomorrow thanks to Mad Ed and co.
My arithmetic has 945 TWh world wide as being about 1/8th of a MWh per person.
About 3.5 days of energy (1% of a year) if you are a typical UK energy consumer.
It’s not nothing, but we’ll survive that.
Hoping someone good with big numbers can corroborate or ferguson that
“the energy demand for AI is humungus: … with global usage projected to more than double to ~945 TWh by 2030 (from ~415 TWh in 2024)”
Just like Tim’s ‘rubbish in the Rhine’ article today, the response to big numbers is “yes, but what’s that out of?”.
Total global energy use is apparently 633,000 TWh p.a., so we’re looking at an increase from ‘a bit under 0.1%’ to ‘a bit over 0.1%’.
Idk, but the extra power requirements won’t be split across the globe equally. Massively concentrated in the US, energy-fucked Europe, and select Asian nations.
AI is projected to use hundreds of TW hours in the next 4 years (!) When Hinkley Point C comes online, it’ll provide 3.2 GWh.
We’ve been trying to build it since the 80’s….
Like everything else related to AI, the arithmetic ain’t encouraging.
Britain’s inability to build shit is the problem, not AI. In 2009, the UAE awarded a contract to Korea to build a 5.6GW power plant. It has been fully operational since March 2024.
When a nation is as useless as the UK, literally everything that happens is a problem.
Of the top 50 listed technology companies in the world (by market cap) the UK HQs precisely none.
The UK has and has had some of the brightest minds in the field but the morons in charge in Government seem to hate success or stifle growth in red tape or tax it out of existence.
Wankers.
Naughty Steve, confounding power and energy. Over 4 years, Hinkley C would provide over 100TWh, assuming we can ever get it started.
“with global usage projected to more than double to ~945 TWh by 2030 (from ~415 TWh in 2024)”
Is there an Excel spreadsheet available explaining this, or just some wanker at Accenture pulling numbers out of their arse?
I’m very sceptical for various reasons. The bubble thing going on, and that bigger models and more training have diminishing returns. I’ve tried different models and can’t tell much difference.
As Bloke in South Dorset says, 0.1% of global power consumption. And that’s before we get into how AI training is something that could be done in remote places on redundant hydro, or on unreliable power like solar. Just checkpoint the process when the sun stops shining and restart it when it comes out again.
Estimated total global power provided by wind in 2024 is about 2500 TWh.
Which is something like 1% of total electricity, let alone total power.
It looks like it’s a big number, but it really isn’t.
If AI has a vulnerability then it’s one or both of (i) it’s massively over-hyped, and (ii) the financing of these data centres looks like an inverted pyramid of mutual debts, of the sort that could collapse in the blink of an eye.
Naturally Murphy ignores these possibilities to promote leftie claptrap instead.
How so?
https://www.theatlantic.com/economy/2025/12/nvidia-ai-financing-deals/685197/
Ace, so if it all collapses, only a handful of overvalued firms go under. As much of the debt comes from PE funds, there won’t even be a banking crisis.
Thanks, Steve. In your link a guru declares “But AI is an exponential technology.” I wonder whether it will prove existential.
It looks a lot like the ‘dark fiber’ buildout leading up to the Dotcom Crash.
I wish these Greta Groupies would get their stories straight. We’re all going to drown due to Gerbil Worming and rising sea levels! There’s a shortage of water! Electricity is going to be so cheap it’ll be almost free! We’re going to be using too much electrcity! Ordinary people are really rich now but when AI is up and running they’ll be really poor! The Labour Government have conquered poverty!
Confused? Perhaps I should read more Spudonomics…
Yes, if such clever ways are invented (and allowed).
The other bit, I dunno.
We generate energy to do stuff people want. If they want AI we use it to provide AI.
To me this is a bit like Tim’s argument that we can just import everything because comparative advantage means if the Chinese can make steel cheaper than we can why would we make our own?
See also ‘Jobs are a cost’, and many similar arguments, many of which were developed by (or based upon the ideas of) men living in countries at a time in which they made a lot more sense than they do now.
Yes, imports are great, until there’s a war, or a blockade, or a pandemic, or some other bollocks.
Back in Ricardo’s day, pandemics were a fact of life largely unaffected by bizarre worldwide government action (as they still should be), and wars wouldn’t have affected Britain’s ability to arm or feed itself to the extent a big one might today (and nobody was going to run a successful blockade against the Royal Navy), so it all sounded and was logical.
And of course jobs are a cost – but so are welfare, and crime, neither of which existed to anything like today’s extent in the 18th and 19th centuries, when a lot of these theories were developed.
All a long way of saying that if New Orleans’ power goes out*, and murder and mayhem ensues, I don’t think people will link that to their use of AI because people don’t think like that.
I hate government just about as much as anyone, and I completely understand that planners fuck everything up, and I prefer free markets etc etc.
But that’s not an argument for no planning, IMO, just the least amount possible.
Similarly:
Generally speaking, I think the customers of the people running the data centres will pay for the electricity and water, one way or another… But so will everyone else.
If there was a way of ensuring that
a) the data centre people passed that cost on only to their direct customers, and those customers passed on their extra costs only to their own direct customers who paid directly for their services, so that it didn’t treble the cost of power for little old ladies who have no fucking idea what AI is (OK maybe there’d have to be some increase in everyone’s bills to reflect that we all perhaps benefit to some extent, though I’m not sure to what extent these things actually are all benefits)
and
b) probably more importantly, the power and water supplies were not existentially threatened, I’d be reasonably OK with it.
*Currently, Meta is building a vast data centre in Louisiana which will consumer twice and perhaps three times as much power as New Orleans, in a deal that bypasses regulations and conceals how much Zuck will contribute to the cost of building the power supply.
I doubt the grid will cope, and I think the whole thing is fraught with danger.
I think you might be reading more into my words than what was intended. Technology isn’t fixed and I’ll take my chances on markets and clever people rather than the fat controller in Ely.
And I’d take the serial catastrophes promised by the media more seriously if they were at all interested in the likely ones we’ll get from the Nut Zero policies Spud fully supports.
To be honest, while I replied to you I was being more general.
Agree re the media – I fucking hate them with a passion.
the customers of the people running the data centres will pay for the electricity and water, one way or another… But so will everyone else
AI (specifically Sam Altman, who belongs in Bernie Madoff prison if not Gitmo) is already imposing enormous costs on consumers and businesses, even if they don’t want or use AI.
The RAM crisis is inflating the cost of every device. 40% of global memory supply being squatted on by one shit company that will never be profitable, just so they could fuck their competitors and keep blowing air into the most ridiculous bubble any of us have ever seen in our lives. Do you want an AI enabled fridge, washing machine or toothbrush? Whaddayamean, “no”?!?
AI (which stands for Actually Indians) is also making people stupider. Nobody’s learning anything at Yooni anymore, they’re all lying and cheating using AI, all the time (like Indians). Can we calculate the cost of an externality that lowers the average IQ by 30 points? Maybe on Gemini.
I’ve been a computer man since my dad plugged a ZX81 into our black and white telly, and I say: Kill the thinking machines, Butlerian Jihad now!
Hence the phrase ‘if there was a way’.
The RAM thing is interesting though – that is a gap which can and I assume soon will be filled.
No, it won’t be I’m afraid – takes years to plan, build and staff a new semiconductor fab. And atm, nobody’s interested in investing in more capacity for non-AI products. Those markets are being abandoned in the gold rush, e.g. Nvidia used to be a graphics card company, now 90% of their earnings are from AI.
Speaking of graphics cards, you used to be able to buy a top of the line GPU for about £400. Now, you’d be looking to add at least another £1000 to the bill. The consumer market never recovered from the crypto mining bubble (an even more pointless waste of electricity than AI – we’re burning leccy, real wealth, to get computers to do nonproductive work which is only necessary because fake internet money requires artificial scarcity. Madness.)
It’s wise that no-one is bothering.
OpenAI and Anthropic are hype. Classic build-to-flip. VCs spend a ton of money on building it up, talking it up. and when they think they can’t do it any more, they go IPO and a load of rubes snap it up, and they cash out. This is 6 months away.
At which point, also, no-one cares much about spending this stupid money on them. The owners want a return, not shovelling more money at Sam Altman’s dreams.
So 12-18 months from now, it could be that all those RAM orders are cancelled, or resold. And you’ll get 64GB of DDR5 with 8 stamps from Costa as the world is awash with it.
And crypto is just going to die. It won’t be long. Everyone jumps all over bubbles, but once they start to fall, it’s Armageddon. They return to intrinsic value, in the case of crypto, about zero.
The consumer market never recovered from the crypto mining bubble (an even more pointless waste of electricity than AI – we’re burning leccy, real wealth, to get computers to do nonproductive work which is only necessary because fake internet money requires artificial scarcity. Madness.)
That’s bollox Steve. If crypto was being used as money (in commerce) the amount of it in circulation would be irrelevant. Money is valued by the quantity of good & services it’s exchangeable for & that’s it. Total of tokens of value = total of production/consumption.
It’s because crypto is viewed as a speculative “investment” that the volume of crypto units becomes important. Because crypto is not money.
If crypto was being used as money (in commerce) the amount of it in circulation would be irrelevant
Why do you think “mining” exists?
Think, McFly. Think.
No. You think. The “mining” exists because the tokens of value needed actual value. The proof of work. That some value was consumed in their production. Energy plus the equipment & effort of the miner. They weren’t just valueless bits of paper like governments churn out. So they had a utility value because the only way to produce them was to consume value. If they’d been used in commerce as intended they would facilitated commerce & would be money.
However, idiots who don’t know what money is have treated them as a physical asset & exchanged government money for them rather than goods or services in commerce. Dead end.
You need to understand money is not a thing it’s a process. So if it’s not part of the process of commerce it’s not money. It’s commerce that values money so the quantity of individual tokens of value in circulation is irrelevant. Which gets proved regularly by inflation. It doesn’t matter how much government money governments print the value going around in commerce is unchanged.
The thing with AI is that it’s currently a lie. As in, ChatGPT and Anthropic aren’t charging you. This is to raise awareness for IPO so rubes buy it up.
Once the company has to break even or make a profit, they’re going to charge you for generating the theme from Neon Genesis Evangelion as a madrigal. And people won’t bother.
There’s uses, because there’s AI stuff that’s profitable, mostly in the corporate space. Credit card fraud detection, ANPR cameras. I don’t need Github Copilot enough to get a subscription but they could charge me £10 a month for what I use and I’d pay it.
Very accurate assessment. AI has to produce value that people will pay for. And there’s not much of that going on yet.
“And of course jobs are a cost – but so are welfare, and crime, neither of which existed to anything like today’s extent in the 18th and 19th centuries, when a lot of these theories were developed.”
Welfare is mostly a product of the welfare system being introduced. Before the welfare system we had “lock up your daughters”, vetting of suitors, chaperones. And if those systems didn’t prevent women getting knocked up, they went on the game. Which acted as a deterrent to the advances of cads and bounders. Get knocked up and you might be sucking strange cock all day.
Once you tell women that the state will look after them, they get a lot less careful.
Crime has risen because the state monopolises dealing with criminals. 100 years ago, you’d just shoot someone breaking into your house, and unless it seemed suspicious, the police would shrug and take the body away. Shouldn’t have been there, tough shit. Once you disarm the population, criminals can have a field day. Then it takes a dozen or more offences for jail, when they really should be executed or sent off to some middle east shithole after 3 (I’m a soft liberal and I believe in giving people a chance).
Why doesn’t Japan have any graffiti? Because it doesn’t take doing it many times to go to jail.
Agreed. I wasn’t suggesting they arose ab initio.
At what cost? The diminution of the collective IQ of the species, obviously.
A bit obvious, of course, but still a great quote from Billy Madison.
Water is not consumed
Quite so. I’ll heat my house with it, thanks, before it goes back for reprocessing to potable.
Not at the hydrogen and oxygen atoms scale, but the increasing demand is not for ‘water’ but for clean water, and that does get consumed.
If you drink it and piss it away it goes into the sewers (or in my case into a domestic treatment plant leading to a soakaway) and needs to be recovered and cleaned.
Yes, I get that point but data centres don’t drink and piss water, they simply run it close to heat so that some of that heat is transferred to the water.
Quite. You could cool a data center with raw sewage if you had a supply.
So build your data center next to Richard Murphy?
There’s a couple of guys online that tried water-cooling PCs with vodka. The lower freezing point means you can overclock a CPU harder.
Smeghead.
“AI cooling systems are diverting vast quantities of water in a world already facing severe shortages”
There is not a global water shortage, there are specific localised water shortages in particular areas.
Presumably they will build AI data centres in places where water is cheap because there’s more of it than is needed, rather than in places where it’s expensive because there isn’t enough.
All of Iceland should be a data centre. Endless power, water and cooling. The M4 corridor, not so much.
The largest problem with using Iceland is that you still need people to maintain things. A “lights out” facility still needs the occasional visit.
Getting there in winter is a real problem, and incentivizing techs to move there even more so.
But there are already people who live there and like it, called “Icelanders”. I know this because my wife’s been, and met some. They’re smart enough to have built enough geothermal-powered greenhouses to provide the entire island’s fruit & veg requirements. Yes, I was astonished to learn this, too, thinking that they must live exclusively on fermented shark and expensive imports.
Pay people enough and they’ll do almost anything. Look at North Sea oil rig workers, or the guys in the Western Australian mines. Given the potentially low cost of running data centres in Iceland you could pay the techs shitloads and still undercut the M4 corridor.
As I see it the problems are probably to do with sovereignty, and vulnerability and latency in the undersea cables. There won’t be many clients for the data farms in Iceland itself; all that I/O will have to travel.
Yes, I did do an album with Bjork. /boast
You’re quite right, Norman. The reason there are lots of data centres in Slough is for those reasons. It’s a cheapish place, but low latency to London.
When you host on Amazon AWS you pick zones. Like London, Madrid, Frankfurt, Canada, Mexico. And some companies will just deploy to all of them and traffic is directed to the the most appropriate.
What Iceland would be useful for is applications where latency and high uptime don’t matter so much. Like if you’re doing CG rendering, you send data to a server and it churns away for a days and then when it’s done you get the results. If the line is down for 5 minutes after it’s done it doesn’t matter. Or you could even just put the hard drives on a boat.
I don’t get this obsession with the use of water. Very few industrial processes actually consume water. In the sense that it’s incorporated in a product or becomes contaminated so it’s not usable for other purposes. What are data centers supposed to be doing? Running a mains supply through for cooling & piping it down the drain? Seems unlikely. The cooling system would end up looking like the inside of a water heater. Mass of calcite. They’ll have a cooling loop with a heat exchanger. Might not even be water. If you’re cooling the exchanger with water why tip expensive water down the drain when you can let it cool & re-use? Even power generating stations don’t use water. They borrow it. Abstract it, run it through the plant, put it back where it came from a few degrees warmer.
If you think about it, you don’t even use water domestically. It’s abstracted, treated for consumption, comes out your tap & goes down the sewer, where it’s treated & released. Not much different from when it was abstracted.
Problem is we’re not great at putting it back where it came from, so the cheap sources tend to run out.
There’s only two places it can go, Norm. It either evaporates or ends up in a river.
Sure, but if you put it in a river it can take a long time to find its way back to the aquifer you abstracted it from, and if you put it back there directly you first have to re-purify it to the equivalent of 10,000 years of filtering through rocks. This costs, so it goes in the river and to the sea, and your aquifer depletes.
You take it out of the water cycle and you put it back. While that may seem to be using a lot, consider where most of the water in the cycle is. In the ocean or in the atmosphere. Not much in the form of fresh water. All the fresh water in the world amounts to about 3%, but that includes the ice caps which contain most of it by far. Looked at that way, you are using a fraction of what’s in the cycle for a fraction of the time. If we ever got short of water we could melt the ice. That we don’t means it is more of a problem of where it is rather than how much we ‘use up’.
Exactly my point, but the aquifer example remains. In the SE of England there’s an over-reliance on extraction from rivers and aquifers because they’re simple, cheap sources. It takes a long time for the Water Cycle to replenish aquifers. We could use desalination plants but we built one that doesn’t get used. We could also build more reservoirs to speed up that part of the Cycle but we don’t do that because Reasons, including solar farms.
Desalination sounds like a useful task for over-producing windmills. But collecting the water that falls from the sky everywhere in the UK seems preferable if they would let us do it.
I don’t think aquifers get used much for industrial processes. Not enough flow, and it’s generally hard water, which is a pain to use for cooling.
Unless it’s farming you’re actually talking about. Water for crops and animal drinking water actually gets used.
Anyway we’ve been there before. The total water usage for an actual data center they were quoting was the equivalent of a bathroom basin tap. They were trying to assign the water used in electricity generation as well. And that doesn’t use water. It simply borrows it from a river & returns it.
Anyone know what ‘physical cost’ means?
Gamecock asked AI, it said it costs about $200 for an annual physical.
I’m guessing it’s using extra adjectives to make it sound profound.
Isn’t this a guy who uses ChatGPT to form absolutely bonkers ‘theories’ of ‘quantum economics’?