We're all going to be paying AI's Godzilla-sized power bills When I was a wet-behind-the-ears developer running my programs on an IBM 360, a mainframe that was slower than a Raspberry Pi Zero W, my machine used about 50 kilowatts (kW). I thought that was a lot of power. Little did I know what was coming. Today, a large, AI-dedicated datacenter typically requires 100 megawatts (MW). That …
Whereas the actual intelligence in the UK is powered by a slice of cake and a nice cup of tea.
Well, our artificial intelligence has jetted off to Egypt and thus free-gear Kier will probably be getting some nice cake and tea. Meanwhile, the Pairish notice board I passed on my way to the shops has announced the opening of their warming centres. Tea & cake extra. But rather crazy that we have to have those in modern, civilised UK.
Meanwhile, let them eat cake.
No cake for you. It contains sugar, and sugar is taxable. I mean bad. Meanwhile, is the Bbc quietly revolting?
https://www.bbc.co.uk/news/articles/c9qnqx7znqno
Keir Starmer will pay "particular tribute" to US President Donald Trump at the gathering in Sharm el-Sheikh, Downing Street has said.
I don't think Starmer or Downing Street has explained yet just how much this particular tribute will cost us. But I guess paying tribute is expected of vassals.
Almost. The infrastructure isn't built, yet, and infrastructure projects are notorious for taking longer than they should so over-production of power in the medium to long term is hardly likely. But the bubble will burst though so I doubt tax payers will end up funding GPU power draw – not commonly or for very long, anyway.
The *real* question will be whether the tax payer ends up bailing out the speculators who find themselves holding the bag when the scam reaches its conclusion.
This could go either way. In any kind of just reality, this would be absurd. But it's almost certain that this will become yet another "too big to fail" scenario and we all know how those end: the rich get richer.
Is this scam perhaps the first that's "too big to bail"? You know: I fucking hope so!
I'd love to be glass-is-half-full on this but I predict half-built power stations, contracts with impossible get-out clauses and grid connections gerrymandered to hell as fallout from the likely bursting of the AI bubble.
But if I'm wrong, I'll cheerfully charge your car whilst we have a brew.
Even with avoided cost of construction the land acquisition, planning and permitting cost has to be paid for, and for projects where they've actually started running HV lines or building new substations they will finish those projects because they'll figure a lot of it is sunk cost and they'll need it someday.
If they've built new production infrastructure they'll shut down the old inefficient and more expensive stuff like coal plants. Electricity doesn't work like other markets such as oil where if there's an oversupply prices crash, producers lose billions, and consumers benefit. It is highly regulated and they are allowed to match supply to demand to keep prices stable and profit margins intact.
So best can we hope for is that prices quit rising, but in places where they've already gone up they aren't coming back down. I'm fortunate that I'm on the border where two large utility territories meet, and the "other" territory is where the datacenters are getting built (a reasonable number, I'm not in the DC/VA corridor or other major datacenter hubs) They've seen prices go up by over 50% and mine are up by less than 20%. Mine also benefits from a lot of wind power and modern natgas turbines, that other one has a lot of aging plants like coal and first gen natgas that are expensive to operate and an aging nuke plant they planned to shut down that has had its shutdown postponed because of those hungry datacenters!
"they are allowed to match supply to demand to keep prices stable"
Basic physics I'm afraid. Demand and supply have to match.
Some of that demand may come from energy storage, and some of the supply will later come from energy storage, but demand and supply *must* match, else you lose control of frequency and/or voltage.
If they've built new production infrastructure they'll shut down the old inefficient and more expensive stuff like coal plants.
That's pretty much a lie though. Coal is more efficient than 'renewables' because coal can provide reliable power. Feed coal in one end, electrons shuffle out the other and we're not at the mercy of the weather. Like yesterday in the UK-
minimum: 0.647 GW maximum: 2.343 GW average: 1.332 GW
Old is more true, ie the UK's coal fleet was old, along came the Climate Change Act and calls to reduce carbon emissions by 30%. Then this-
https://en.wikipedia.org/wiki/Kingsnorth_power_station
They had proposed constructing two new 800 MW supercritical coal-fired power units on the site, to be operational "as early as 2012"
So if we'd built that, a single coal power station would have met CCA targets, and provide more power than the entire UK wind fleet. But the neo-luddites forced that to be abandoned and the power station closed completely. Much the same with 'expensive' and coal, gas and nuclear are made artificially expensive via 'renewables' subsidies or loading on carbon taxes.
Hate to be the one to tell you but electricity generated from coal is now more expensive than renewables + battery, at least here in the US. Coal is dead, and has no role in the future for power generation. It may see a brief blip as AI causes severe electric market distortions, but once the bubble pops that'll be the last hurrah for coal. Its use will trend strongly downward and disappear entirely by the end of next decade if not sooner.
Hate to be the one to tell you but electricity generated from coal is now more expensive than renewables + battery, at least here in the US.
Hate to be the one to tell you this, but unicorns on treadmills are cheaper than both. But as always, citation needed, especially when the US is a big country with a diverse range of climates. So solar might be just fine in Arizona or Texas, but won't be cheap in Alaska.
Its use will trend strongly downward and disappear entirely by the end of next decade if not sooner.
Again YMMV. If you're a country that wants cheap, reliable energy and has coal, you might build more coal generation.. Like.. err. Germany.
> But as always, citation needed, especially when the US is a big country with a diverse range of climates
Oh look, a citation:
https://www.theecoexperts.co.uk/news/is-renewable-energy-cheaper-than-fossil-fuels
The IEA reported that in 2023, an estimated 96% of newly installed, utility-scale solar PV and onshore wind capacity had lower generation costs than new coal and natural gas.
Three-quarters of these new wind and solar PV plants offered cheaper power than existing fossil fuel facilities.
Admittedly, I don't know how much that equation has changed since the current US administration decided to relax regulations on fossil-fuel plants, but still.
And even then: between 1999 and 2020, the emissions from coal powered plants in the USA alone are estimated to have killed half a million people.
https://cns.utexas.edu/news/research/coal-power-killed-half-million-people-us-over-two-decades
And that's a lowball number, since "the research does not account for any additional deaths among individuals under age 65 or among uninsured people".
On a brighter note, regulatory improvements (and coal plants shutting down due to lack of competitiveness) did reduce this by about 95%, or from about 25,000 deaths per year to "just" 1250 per year.
But with regulations being relaxed, and coal plants being brought back online as part of the LLM gold rush[*], those death rates are going to start climbing again.
> So solar might be just fine in Arizona or Texas, but won't be cheap in Alaska
Y'know, we have these things called wires, which you can send electricity down. For instance, Path 27 is around 500 miles long, and runs between Utah and California:
https://en.wikipedia.org/wiki/Path_27
There's also the fact that no-one relies purely on solar. There's also plants powered by wind, water, biomass, etc.
> If you're a country that wants cheap, reliable energy and has coal, you might build more coal generation.. Like.. err. Germany
Germany is a unique case: the Green lobby is very strong there, and pushed hard to get nuclear power banned following the Fukushima disaster. Which left them with the unfortunate fact that the only feasible replacement was coal, with all the environmental issues that brings. Especially once Russia's invasion of Ukraine drove up the price (and political consequences) of using natural gas.
As a result, it looks like the pendulum is finally swinging the other way, and Germany is now considering nuclear power again.
https://www.ft.com/content/e99efa2b-338a-4065-89c6-0683d5759ed7
[*] Primarily because it's quicker and easier to fire up a mothballed power plant than to build a new one. And when investors are throwing billions around, who cares if the power plant is actually efficient?
"As a result, it looks like the pendulum is finally swinging the other way, and Germany is now considering nuclear power again."
Oh good - I mean it's not like Germany has a long history with earthquakes and tsunamis... certainly don't recall it being prone to being dropped a full metre and shunted two and half metres east...
"And even then: between 1999 and 2020, the emissions from coal powered plants in the USA alone are estimated to have killed half a million people"
And Muppets burning wood (including painted and chemically treated pallets), coal and who knows what else in their "eco" wood stoves have caused more than 61,000 deaths a year (and that's just woodstoves and not including open fires etc)
https://theconversation.com/wood-burning-stoves-are-a-serious-problem-for-your-health-and-the-environment-245737
Even more ridiculous - in Scotland you now can't have a gas boiler or an oil boiler in a new house or one that's under going to the alteration BUT thanks to Scotland's answer to the Taliban (Forbes) you can put in an open fire or wood stove and burn wood, coal or worse PEAT (one of THE dirtiest fuels) - utter and total idiocracy at work...
Any day now, the AI bubble will burst and electricity will become cheap as chips thanks to the over-built infrastructure.
That depends entirely on:
a) the regulatory scheme and wholesale market structure
b) AI resulting in extra capacity being built...
c) the extra capacity having been paid for.
In the UK, the pretend electricity wholesale market has been entirely knobbled by our government's Canutian fight against climate, so that the old rules of system cost recovery and wholesale marginal pricing has increasingly little relevance to the prices customers pay, hence the UK enjoying amongst the highest energy prices in the world. But as we're a net importer of electricity and rely on France to keep our lights on, and new build of dispatchable capacity won't even address non-AI workloads, its only a theoretical consideration.
In the US, the old system rules probably still largely apply, but because big tech haven't planned ahead there's probably very, very little new capacity (for AI) that's anywhere near completion. And unless the AI firms have splunked the cash up front (which I'm pretty much sure they won't have) then any new capacity still has to be paid for post-bubble by remaining customers.
The AI DC’s should be multi-level
Ground - Server racks
First Floor - Old Folks Home*
Second Floor - Greenhouses
And pipe the heat upwards, based on needs.
*Absorbing the heat must pay more than council contracted care fees solving a secondary problem.
For ‘mur’cans First, Second, Third floor as you can’t count building levels or day a week starts on.
I've often thought that we should follow the Dutch example and build a dementia village, where there is a higher staffing level, that people can roam around the site freely and in safety, that they can take stuff from stores and not face any hassle.
The perfect site for a trial version would have been the (mothballed and now sold for housing ) former open prison HMP Noranside - quiet country location, close to a main trunk though for family visitors, similar in design to a holiday camp from decades past, take down the cell doors and remove the bars from the windows, even put the staff in coloured jackets and the patients would quite likely believe they were at a holiday camp, put on period compliant entertainment, give them a gardening space and flat level nature walks with benches and many would be calmer and likely need less or no medication, thus making them far more lucid and give them better health outcomes.
create a fake gate well before the actual boundary and have the staff tell them that the buses/trains etc have broken down/on strike etc and we apologise for the inconvenience but this lady / gentleman will take you to get a meal/tea/coffee and cake on us while you wait.
Instead lets turn sites like that into more overpriced and under built "luxury developments" and medicate people into zombies
I see a race coming between the bursting of the AI bubble, the cracking of our already overburdened electrical grid, and all of us shivering in the winter and baking in the summertime, as AI-driven costs and brownouts make us miserable in our homes.
It might not take that long, and it might have some benefits with the right regulatory nudges. So you highlighted the problem with solar, which is a part-time solution. Europe also has this at the moment-
https://notalotofpeopleknowthat.wordpress.com/2025/10/11/anti-cyclone-descends-on-europe/
The anti-cyclone will also establish itself over most of NW Europe as well, so it will be interesting to see how much wind power is generated.
Answer for the UK so far is 'not much'-
https://gridwatch.co.uk/Wind
minimum: 0.383 GW maximum: 2.084 GW average: 0.938 GW
For yesterday's wind generation. Being an anti-cyclone, at least there's some clear skies, but we're heading into winter and insolation is dropping. And like you say, providing reliable power takes time. SMRs might be a savior or full-size NPPs better, but those take a long time to build. Well, in the West anyway. The benefits might come from following the 'polluter pays' principle and making it a planning condition for AI bit-barns for 'consumer pays'. So want 100MW? Build 100MW of baseload generation rather than just getting 100MW power contracts which increases demand and thus price. Then if/when the AI bubble does burst, there'll be surplus capacity and then lower energy costs. Especially if the cats get bounced and generating capacity bought out of Ch.11 or administration.
Governments could also probably incentivise the big consumers, ie Apple, MS, AlphaGoo, FaceMelta to become energy producers rather than just consumers, especially as most also happen to be sitting on large piles of cash.
Mine absorbs 50kW (max 100 if almost flat) but will accept down to 1.4kW and that's driven by the charger which can b3 driven by the Grid.
So if you had quite a lot of them plugged in at a moment, you'd have a large buffer against swings in production and other demand. Useful.
Mostly they charge overnight, absorbing 7kW (some on 11kW 3 phase) during 5 hours or so. So 1GW would charge (no napkin so I'll use my head)
Fast DC: 20 cars per MW :: 20 000 per GW simultaneously, all done in 1 hour, so 5 lots in one dead of night. (100 000 if we are all up and zipping around; and ... rather a lot during the whole day)
Slow AC: 140 cars per MW :: 140 000 per GW simultaneously, during most of all or little more than dead of one night.
Perhaps my head, or that napkin, is soggy?
You mean photo-voltaics?
Almost certainly won't since grid-tie inverters are required to turn off when the mains goes out to stop your micro-generators zapping the guys out in the street repairing the cables.
Its possible to keep the house going but the kit is a lot more complicated/expensive and I'm gonna take a wild guess that your new-build builder hasn't gone for that.
"Its possible to keep the house going but the kit is a lot more complicated/expensive and I'm gonna take a wild guess that your new-build builder hasn't gone for that."
Indeed. I've considered fitting solar panels to my house in the past but the costs really jump if/once you factor in power generation whenever the grid-connection is off (i.e. automatic grid isolation, battery storage etc).
Going to get dinner here as soon as I can with a battery
Costa about 12.5 to 13k*. Seen a couple of the electric companies have offers on which may decrease this cost
We are in the SE a d out house issouth facing so we are ideal for it. Payback for us is not too bad, but even quicker if this article does be one true
*From a previous quote that we had to pull which had 2 X 5kw batteries and the latest revised quote is for a Tesla powerwall
Ah, probably the most expensive solution available. There are plenty of other choices of battery that are a hell of a lot cheaper than the powerwall.
I've had 33kWh of battery working at my home for 4 years now. I recently added some more Solar Panels and once SSE gets their finger out of their ass, and issues me a G-99 cert, I can get paid 16p/kWh I send back to the grid.
My EV's V2L feature will provide the 240V needed if the grid goes down (and I have removed the fuse so isolating my house). I can run the whole house inc heating for 2-3 days in that situation.
If we have a good summer in 2026, my leccy bill will be less than zero.
Yes, it takes a lot of capital investment in the house but I'll get it all back when I sell up. Sadly for most people what I've done goes over their heads and will continue with their combi boilers and petrol/diesel cars.
The anti Renewables crowd on this site can sharpen their downvotes coz, I don't give a damm what you say or think. I know that in the long term, I'm onto a winner.
I get accused of being anti-renewable because I don't believe the CO2 the world is ending shite. But I'm all for economic renewables. I have them but only because it economically makes sense for individuals with the leccy prices. But ... if the government actually built low pollution power stations it wouldn't be economical for individuals. I even wanted a small, unobtrusive turbine but can't get planning. Net Zero is that important ...
"The anti Renewables crowd on this site can sharpen their downvotes coz, I don't give a damm what you say or think. I know that in the long term, I'm onto a winner."
So, to be clear, you expect the grid to be available for your use as and when renewables don't meet your demand (lets face it, at 5% capacity factor in winter solar isn't going to heat your house or power your car, it might just keep the lights on), and you expect the grid to be available for you to dump excess power to, but you hope to be a net non-contributor to public system costs? That's generous of you, and it's on ly right that the poor who can't afford eco-cars, battery storage and houses with sufficient space for a big solar array can pay all the system costs.
No wonder you posted AC.
"but you hope to be a net non-contributor to public system costs?"
You seem to be missing the point.
When we export energy we get paid for it, and we get paid less than you're paying for the energy we're supplying.
We still pay the standing charge, we still pay for any energy imported, but we also get paid when we export. That happens in two main periods:
- The middle of the day, when it's bright sunshine and the battery is full (or there isn't one) and we're not self consuming*
- The peak period, when our batteries are exporting because it's economically favourable.
The key variable here is self consumption. It's beneficial for the grid for us to arbitrage and export, so it's made economically beneficial. Else we'd focus on self consumption, using the battery to cover overnight rather than charging at off peak rates, and putting any excess generation during the day into hot water, EV batteries, etc...
Whilst the overall bill for some households might be negative, that's not because they're not paying their share of the public infra costs, it's because they're exporting more than they import, and doing so at times when it's good for the grid, resulting in sales which exceed what they are paying in standing charges.
With my small array it's pretty rare than I get negative price days, but my bill is ~£600 a year, and that covers my car fuel, which would have been about £2k, as well as my household usage, which would have been about £1800 (pretty high usage household with multiple people WFH).
There is a #vimesboots element to anything we do that costs money, but here there is actually a net benefit to all from the personal investment, not just to the individual (though obviously the economic benefit is most focussed on the individual).
is not positive, is it. Although they do use some electricity, and expect the Grid to be there to dump their surplus into.
Had you really thought out your comment?
HP is larger, and has a different contract for supply and use, but size is not the primary determinant.
The anti Renewables crowd on this site can sharpen their downvotes coz, I don't give a damm what you say or think. I know that in the long term, I'm onto a winner.
Ah, the selfless nature of the virtuous. I'm alright Jack, and don't care who's paying to subsidise you. Also perhaps an optimist in assuming you'll get the money back when you sell your property. Jury seems to be out on that one, possibly while insurers & mortgage lenders work out the risks. Especially when some of the talk doesn't seem to differentiate between Lithium and more traditional battery designs.
There are plenty in "the renewables crowd" that find your post not only smug, but also poorly argued because we understand the physics, and can do the maths required to assess the investments, especially if subsidies are removed. 33 kWh is very, very expensive and will only give you at best a week or so more than something closer to your daily consumption. If I was looking for greater autonomy, I'd have a diesel generator. But you also completely neglect to mention your heating requirements which normally dwarf the everything else in winter: ours are around 10:1 over the year.
Anyway with the spot market in most of Europe now usually well below feed-in tariffs for much of the year, and even negative in the summer, feed-in tariffs should discontinued or be phased out as they are effectively transfers from those without solar panels, and many people can't install them, to those with them. And, thanks partly to Chinese (probably coal-powered) overproduction, it's already economical for many who can to reduce their total demand, so additional subsidies just add insult to injury. How do your calculations look with subsidies removed? Should still make economic sense.
Dunkelflauten such as that we're currently experiencing in north western Europe, are a real and increasingly expensive feature of the transition, especially if temperatures fall, because they require we hold additional capacity in stand-by and battery storage at the scale required will never be economical even assuming it can be produced.
Here, we're just below production break-even, partly due to the weather, but probably won't be able to charge the 5 kWh battery fully now before spring even though we don't use a lot of power. Feed-in tariffs are so low that we requested to opt out only to be told by the local utility that that wasn't option. We might get a plugin-hybrid at some point, now that the BYD Seal DM-I is available in Europe, though I can't find it on the UK website: ≥ 100 MPG would encourage more people to switch and would lead to a noticeable drop in petrol demand and shouldn't need subsidies.
But there is so much still to do.
"below feed-in tariffs"
There are relatively few people still on the early FiT, which were high. Even a decade ago they were down at an effective 5p/kWh - and that's not worth it, you're better off ditching the export component and actually selling the energy on the market (not something the local network operator has any control over).
It's going to really blow people's minds to think that the variable rate is regularly over 20p, and last week nearly hit 30p on three days.
Yes - that's a variable rate, and it dips as well as climbs, that's why I mentioned the flat rate instead.
If you look on some of the Chinese wholesale sites, there are indeed bargains to be had. But then add transport costs, taxes and, of course, certification.
Best waiting for a couple of years until sodium ion batteries are produced at scale: these will be excellent for static storage.
As in 10 modules to a 60kW battery which have declined to 70% of rated capacity after a few years use and been salvaged, or swapped out because they declined early (at a discount)
So there's your 4kW battery, with some of its electronics built in, and a 5 year life.
They might be better in a big rack near panels with a caretaker to supervise them, or singly outside houses. There's a business or three there.
A non-grid tie inverter and a switch that detects power flow and switches off the mains connection in the event of a power failure is not that complicated, but those are expensive and its installation may not even be allowed (as we can't have you not dependent on the grid).
..and its installation may not even be allowed (as we can't have you not dependent on the grid).
I'm pretty sure they're not allowed, at least not in the UK, Spain or the US where I've been looking to build. Reasons aren't to encourage dependency but safety, ie not zapping some poor sparky* working on a section of the network that is supposed to be de-energised. I think this includes anything that can feed power back into the grid, so batteries and generators as well. But I'm designing DC-style, so some batteries to but time until the generator fires up. Plus the switchgear also has a programmable delay before going back to mains to eliminate transients and surges. But curious now if export meters include isolation switches. Not really looked at that given I'm working on the assumption that I want to be as independent from the grid (and supplier) as possible.
(* Also why some telcos are reluctant to supply dark fibre given the launch power for LH/ULH (Ultra/Long Haul) optics can be dangerous. Like a field engineer gave me a bit of wood he'd engraved 'Ba*stards!' on because a customer had ignored the instructions in the planned works notice to shut down their lasers. If he'd engraved something a little politer, I'd probably have sent that to the customer's CTO to show them why they're asked to do this.)
It'll help but it wont be enough. We have a big array for a domestic residence and on dull winter days it can drop to ~200W. Get on the right tariffs and you can economically, charge battery cheap overnight and if sunny sell excess solar back to grid. The problem is the control systems aren't that programmable and you can't give them a nice case statement to automate. Then you forget to change something and get annoyed with yourself.
What system - you should be able to rig up control pretty easily.
Probably either based around Siemens Alpha or Schneider equivalent. Both offer pluggable ATS, isolation, distribution in nice, neat cabinets.
Home Assistant is your friend (other automation solutions exist)
Yes and no. No being for a new build, using 'industry standard' solutions makes things like approvals & insurance easier, plus can be less messy than nailing stuff to a bit of ply and hiding it in a cupboard. Maybe costs more, but has safety features like proper lockout switches so I don't start smoking again. Plus Siemens has gadgets like this-
7KM PAC2200 measuring devices
Overview - The energy meter for the DIN rail
The 7KM PAC2200 measuring devices function as versatile energy meters for the DIN rail and indicate the most important measured values directly on the display. Integrated communications interfaces via Modbus TCP, Modbus RTU or M-Bus provide a flexible connection to power monitoring systems. The integrated web server (Modbus TCP version) is used for simple, device-independent visualization of measured values. The devices are incorporated in the familiar and simple system environment of Powerconfig and Powermanager.
That can sit in the cabinet and I can peer at, be hooked into their Sentron management system and sit safely on an isolated network segment. And theoretically, supplier certainty and no cloud dependencies. Have spare critical bits and easy* to maintain. In theory. But a big part of the decision process is not ending up with me standing infront of the smoking ruins of my home, talking to the insurers and them using the excuse of DIY home automation to deny the claim.
*Well, easy assuming no pretty blue flashes. Or easy to call a suitably qualified electrician who can swap the dead bits out without having to wait for parts to arrive.
Home Assistant is industry standard... it will just talk to your devices using the API provided by the supplier.
I'm not in any way suggesting modifications to the power electronics, or their controls. Just using the same remote access as the app.
https://www.home-assistant.io/integrations/siemens/
I use a ModBUS over TCP control for one of my applications - from your text it sounds like that might be appropriate
Home Assistant is industry standard... it will just talk to your devices using the API provided by the supplier.
Sure, but I'm not sure I want to, partly because the automation needs aren't high compared to reliability, security and not having to faff around with configuring and maintaining Pis, scripts ect. Plus I think if I wanted to go back to doing that kind of thing, it also has SNMP support. Having been looking at homes with varying degrees of automation, it seems to be one of those things that gets sold on lots of features that people rarely use. Like do I really want to be able to turn lights off & on with my phone? It's not like it'll be so large I'll need a golf cart or lazy boy on rails to reach light switches.
I'm hoping that the "AI" hucksters will suggest burning a copy of Fahrenheit 451 and then when they run this past ShatGPT it will blow up in a "DOES NOT COMPUTE!" irony machine explosion similar to those in 60s TV shows such as The Avengers.
Either that or they can all F*** off to F***ity F*** F*** land. I haven't got a preference.
I'm adding more chinese made solar panels!!! so I'm not going to get hit as hard although I still have to pay standing charge, which is odd as I cant charge utilkity companies the standard charge for the infrastructure I've put in andto be able to export to help the grid.
We might as well not charge any VAT or duty on Aviation fuel.. oh....
Even the most solid business model requires somebody willing to put up the cash (millions), willing to wait until payback (quite a few years) and most importantly not having a better investment option available.
Lower profit schemes are a winner if the payback time is short enough, plans with 100% profit after 10 years will die in a ditch if the alternative is 35% after three years.
When they were first produced they had generation costs that were higher than existing power plants. The comparison is flawed, of course, because the huge capital costs of existing stations have already been amortised. Anyway, early subsidies supported the innovation and economies of scale that now make total costs more favourable, so the subsidies should be discontinued.
OK - time for a little Devil's Advocacy here. I know I'll get visiously downvoted for what I'm about to say, but so be it.
Firstly, the article starts with: "...an IBM 360, a mainframe that was slower than a Raspberry Pi Zero W [and] used about 50kW". Yes, many moons have waxed and waned since then, but this DOES point out the trend that has been a continous staple in IT over the decades: a continued move toward more computer power for less energy consumption. Right now we're running AI on hardware that in the future will look like that old IBM clunker. In fact, in the days of the 360 it was a well-known fact that the Eniac used three times as much (it weighed in at about 150kW) and had far less computing power than a mid-1970s electronic pocket calculator.
Don't get me wrong, I'm NOT saying that the power consumption resulting from the current AI explosion is not a problem. It is. But it is not the end of everything. Right now the power and cooling requirements of AI are the technology's main threat, not in the least place because that is what takes the the biggest bite out of its ultimate profitabilty. For make no mistake: the companies behind AI intend to make money with it. So energy consumption, heat generation and cooling requirements are at the top of the list of cost centers that simply have to be dealt with. And necessity being the mother of invention, we can expect sigificant improvements in power consumption vs. computational output. Of course, some (probably a lot) will be undone when AI operators give in to the temptation to simply expand their systems when power becomes less of an issue, but the requirements of commercial viability will correct that sooner or later. They always do. If nothing else, money and profit will be the shore that turns the ship.
Secondly, and this is where the downvoting will really start, I can't help but wonder why AI as a whole receives such a harsh treatment by the Reg commentariat. Granted, at the moment it's overblown, overhyped, overbuzzed and overwhatnot, and the technology, being still in constant development, unsurprisingly has its shortcomings. That's all completely true. But at the same time, AI has its uses, real uses.
To use my own, personal and strictly subjective, experiences with ChatGPT as an example, it has save time by helping me debug code, during which process I learned more about coding and the limitations of AI. It has helped me to get a diagnosis for a decades old condition, simply when it gave me an answer that caused me to ask a proper medical professional the right questions. (Needless to say I'm NOT relying on AI as a doctor - I may not be a genius but I'm not quite that daft.) It has also got me out of writer's block by suggesting better structures for non-fiction articles, plot elements for some amateur novel writing that I'm dabbling in, generated illustrations for online blog posts that otherwise would have taken me days to draw. It has helped me sort out obscure issues with kernel drivers for an obsolete AMD GPU. And it's saved me oodles of time across the board because it hands me the info I need much, much quicker than Google searches or Stack Overflow posts ever did.
Yet I read a lot of comments here stating that according to the poster, AI can't go down in flames quickly enough. Yes, Microsoft Coprolite, the nonsense coming out of boardrooms and the overvaluation of AI companies (which is worse than the dotcom bubble of the late 1990s) are a PITA right now, and the reality check is coming, fear ye not! But this is nothing new. I remember the Java-on-a-thin-client hype quite vividly. Everything was going to be Java, remember? Write once, run everywhere, no more problems. Well, that landed with a thud. But today we use Java on Android devices every waking moment; I use Adobe Netbeans (written entirely in Java) almost daily, and in the financial world you can't get around it. The point here is that the hype might have been nonsense, but the technology and its benefits, if applied sensibly, were and are real.
AI is here to stay, people. The hype will pass. The technology will not. And I think that's a good thing.
Let the downvoting begin...
Two things:
- It's not AI.
You're talking about LLMs, and maybe some image diffusion, that's not Intelligence.
They are also prone to wildly confident hallucinations, and whilst they can produce useful output for a number of different tasks, there is an absolute requirement for the output to be checked and tested by a competent individual.
- It's operating on theft.
Many people object to theft, and current LLMs have been "trained" on vast amounts of stolen IP.
"It's not AI."
Than what is your definition of AI? According to the Brittannia, AI is "the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings". LLM based chat bots such as ChatGPT fit that definition.
"You're talking about LLMs, and maybe some image diffusion, that's not Intelligence.
I never claimed that AI, LLMs, image diffusion and what have you are intelligence. In fact, the term "Artificial Intelligence" is, in my opinion, a misnomer. "Simulated Intelligence" would be a better term.
But that's all beside the point I was trying to make.
"They are also prone to wildly confident hallucinations ..."
Yes. I did acknowledge that the technology in its current for has problems. The first cars were death traps by today's standards, the first attempts at powered flight were ludicrous, and the first applicatons of radioactive isotopes were lethal more often than not.
That doesn't make it bad.
"It's operating on theft."
That's not a problem with the technology in itself; it's a problem with the companies operating the systems.
Also, and I have tried to make this point in the past, machine learning and human learing are not all that different when it comes to the basic principle of perusing information and integrating that into a larger body of knowledge (and please don't say that AI's can't have "knowledge"; I'm talking about general principles here.) AI is disruptive technology, which means that the way we view IP has to change as the world changes.
You are probably right in that LLMs will find enduring useful niches as did expert systems in the 1980-90s [recall expert systems were going to make lawyers, doctors etc redundant] but those LLM applications might well be a rather small fraction of the total which also only require quite small models; both presumably reducing the projected energy demand by several orders of magnitude.
The tenuous and ineffable confection that is AI as spruiked by the snake oil peddlers is an abomination which is already causing long term harm and can only ever fail to live up to its accumulating posited but impossible promises being made daily.
The latter really is that bad.
I am guessing that any country not pursuing an AI infrastructure, will be dependent upon (lightly put) or at the mercy of countries that do have an AI infrastructure. These same countries pursuing the replacement of all ICE vehicles by EV's (Canada's current Liberal Government), will have to make a choice - Either pursue AI or EV's. There is no way to have both considering the cost and magnitude of electricity requirements described in this article.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
