Four horsemen of the AI-pocalypse line up capex bigger than Israel's GDP

Four tech megacorps intend to collectively fork out roughly $635 billion this year on capex, much of it for datacenters and AI infrastructure – more than the entire output of Israel's economy and well beyond all global cloud infrastructure services revenue generated last year. The tech titans are all trying to outdo each other …

[MachDiamond]

Profits? We don't need no stinking net profits

The amount of money being spent is frightening. There will be losers in this and really big ones at that. While I doubt I'll be affected directly, indirectly everybody is going to get hammered as money goes up in smoke and companies are reduced to puddles of superfund goo.

It's DotCom all over again with higher prices. When your local electricity hardware entity (not billing companies) that invested loads of money into substations and wires to feed these massive DC's suddenly have no customer, everybody's bills are going way up or government will have to step in with rescue money, or both. There will then be a giant building that's fit for little else with a massive power feed slowly being spray painted and vandalized. All the copper going first.

I'm going to pick up another 1kW of solar panels next week while they are still super cheap to match what I have. When the bell tolls, I don't want to get the bill.

[Doctor Syntax]

Re: Profits? We don't need no stinking net profits

"your local electricity hardware entity ... that invested loads of money into substations and wires to feed these massive DC's"

I'd hope that the substations and wires are paid for up-front by the DC owners.

[doublelayer]

Re: Profits? We don't need no stinking net profits

You'd hope but you'd often not get what you hoped for. DCs ask for a lot of subsidies for these things and sometimes get them, but even when they don't, they sometimes make agreements, agreements that take the form of electricity supplier builds the infrastructure which the supplier prefers so it's done well and can be managed in the same way, the DC commits to buying a certain large amount of power at a certain price that would make that profitable, sometimes very profitable, for the supplier, the supplier announces this to reassure regulators and average customers that bills won't be affected. If AI makes tons of money, that happens. If the DC company goes bankrupt any time before five years after the DC is completed and starts operating, the electricity supplier gets the bill and has to do something to cover it.

It doesn't always work that way, there are some examples of builds where, if it fails, the DC company is on the hook. Unfortunately, that's far from the norm, and often when someone says the DC is paying for all the new build, they mean "in the future if everything goes well".

[MachDiamond]

Re: Profits? We don't need no stinking net profits

"I'd hope that the substations and wires are paid for up-front by the DC owners."

That should be a requirement. If Government entities are going to pay for these facilities, as they do for big companies, there should be a requirement that the site design will support other sorts of businesses. Walmart leaves behind buildings only good for 15-20 years of life that are configured in a way that makes them too expensive to section off for other purposes. They'll also get covenants that prohibit a competing company from taking over the location which I don't think should be allowed. Walmart sells so many things, that any retail store can be argued to be a competitor. One in my area was turned into a giant uHaul self-store mega location.

Even if the DC companies pay for the upgrades, it still leaves behind a bunch of un/under used infrastructure that costs money to secure and keep maintained if they close down. It's a very unique set of requirements. Heavy industry could use a location with that level of power, but it's unlikely the zoning will allow that near to homes.

[Tim99]

Re: Profits? We don't need no stinking net profits

Privatise the profits, socialise the cost?

[ecofeco]

Re: Profits? We don't need no stinking net profits

As always.

[cyberdemon]

Re: Profits? We don't need no stinking net profits

> The amount of money being spent is frightening.

Yes indeed.. Stocks are down, gold is down, oil is down, crypto is down.. So where has all the money gone?? I fully expect a bank crash this year and I have no idea what anyone can really do to ride it out.. It is terrifying to think about.

> I'm going to pick up another 1kW of solar panels next week while they are still super cheap to match what I have.

I might do the same. I have 1kW of panels (realistically 600W, they are laid flat on a flat roof in sunny Britain). But from the grid's point of view, this is pure delinquency. It steals the lunch of the energy suppliers, while at the same time making blackouts much more likely. The inverters require a stable reference signal to operate, any transient disturbance will temporarily trip them off, and the combined effect of that would cause such a rapid change in load that any grid would struggle to cope with. On top of that, you have interesting resonant behaviours with a grid dominated by grid-following inverters causing oscillations in grid voltage and frequency.. Prime suspect in the Iberian blackout last year.

[retiredFool]

Re: Profits? We don't need no stinking net profits

Gold is not down.

[David Hicklin]

Re: Profits? We don't need no stinking net profits

https://www.bbc.co.uk/news/articles/c87r2700dq8o

Gold prices are fluctuating rather than simply rising - it lost 9% in one day but rebounded later - it all depends on what's in the news at any one time.

[MachDiamond]

Re: Profits? We don't need no stinking net profits

"The inverters require a stable reference signal to operate, "

Not really. You can get a system that isn't grid-connected or has the ability to island or can augment the solar power being generated with grid power when necessary. A solar system has to be synced with the grid if the two are interacting. My systems are independent. One feeds the chest freezer with a battery backup and the other feeds my swamp cooler with a battery back up. The next one will power things like my freeze dryer and dehydrator in the garage and I plane to use a Charger/Inverter that can use the grid for supplemental power as needed, but doesn't feed in to the grid. At some point I might combine everything and have a nice big reclaimed EV battery for nighttime use and the grid for backup (zero export). I've been going after the lowest hanging fruit first as I can afford it and find deals on hardware. Nearly all of my lighting is LED so I can turn everything on in the house and it might draw 200-300W. The ROI to cover that doesn't pencil out with solar/batteries unless there's spare capacity somewhere I can hang some of the lighting on. The welder can draw a bunch of power, but I don't use it so often that getting it off-grid is that important. Same with some of my machine tools and I have things I can be doing if the grid is down. I'm looking to save money, I'm not a doomsday prepper.

[cyberdemon]

Re: Profits? We don't need no stinking net profits

Nice.

That must involve some interesting non-standard wiring though. Here in the UK, the insurance company would probably condemn it.

For most people sadly, solar PV means having a grid-following inverter simply plugged in like a negative-power appliance.

[MachDiamond]

Re: Profits? We don't need no stinking net profits

"That must involve some interesting non-standard wiring though."

No, not really. The chest freezer power isn't running through the house wiring. It's a self-contained power supply. Same with the evaporative cooler. Things could get tricky if I get to a point where I want to join things up and I'd likely need to reconfigure everything.

I own my house outright so there's a lot of insurance kerfuffle I don't need to worry about. I do keep everything built to comply with codes and a bit further (the more strict commercial requirements over looser residential ones).

[Doctor Syntax]

Assume for the sake of argument there is scope for this to turn into something viable in the long run. Is there any way that there is sufficient income that could be raised from this to provide sufficient return to pay off all of their investments? If not does this mean that the least well funded fold, leaving just one or two standing or do they all fail anyway before they start to recoup?

[doublelayer]

That depends on how we assume it works. If we assume that AI becomes competent and can successfully do lots of jobs cost-efficiently, then there's plenty of money to be spent and the DC makers will be able to cover their investments. If any of that isn't true, and none of it is, then the DC companies will have more problems.

Likely a couple DCs will survive by being bought up by hyperscalers who have already committed to using it, they move their GPU workload machines into them and use more CPU workload machines in their old DCs, and the rest fail. It's also likely that many fail without getting built in the first place. The big cloud providers do have a lot of money and they'll likely have to pay big chunks of that to get out of obligations they've signed up to without thinking it through.

[Doctor Syntax]

"then there's plenty of money to be spent"

Is there? If lots of jobs are replaced there's less money to be spent by customers and that means that those who would have sold things have less money and they are the ones who would have been buying AI services. I think we saw a few weeks ago the amounts of money per head of the world population that would need to be spent to make the investments bandied about at that time recoupable and it was unrealistic.

[doublelayer]

We're talking about hypothetical consequences from events we know won't happen. Before we could start to estimate them, we'd need to decide or likely simply invent the conditions for that.

Most predictions from promoters of AI don't claim it will replace all or even most jobs, just certain subsets. If, somehow, that turned out to be possible, that wouldn't eliminate so many jobs that companies in general wouldn't have customers. It would be very hard for anyone whose job was lost, and quite possibly a net loss for the economy as a whole, but the economy as a whole wouldn't be paying for the AI, the people who hired the replaced workers would be, and they would probably enjoy some savings.

If a million people could be fired globally with an average saving of £100k, not an unreasonable figure for a few years of salary including some people above the average as frequently predicted, that would cover a £100B datacenter and AI investment expense, more than was actually spent in 2025. That doesn't cover everything that was promised, but most of that hasn't been built yet and might never be. A million lost jobs wouldn't be so many that nobody would be buying things anymore, especially when spread out over multiple continents. That would still be enough pain that we would all notice it to our detriment, especially if that number went higher, but it wouldn't do what you're suggesting.

[MachDiamond]

"A million lost jobs wouldn't be so many that nobody would be buying things anymore, especially when spread out over multiple continents."

If it doesn't cause a panic and 144pt news headlines. If those jobs are eliminated over several years, it would blend into the background. If it was over 6 months, that would be bad. It's also not going to be spread evenly since the vast majority of economic activity is concentrated amongst a billion people.

Money in a local economy has a magnifying effect as it passes from entity to entity. Goods and service are bought and sold. Rents/mortgages are paid. This is why exports are great and imports are significantly less so. The US government bought a contract load of toothpaste abroad as some beancounter "calculated" that it saved money. What didn't get taken into account (other than all of the melamine in the stuff) was the payroll/income taxes, corporate taxes and magnifying effect of the money being kept in the national economy. The local employees would also be paying fuel taxes, sales taxes, property taxes and on and on. Sometimes I think the government collects 98% on every dollar in taxes. Opening a hole in that by buying the toothpaste from China just let money flow out.

[doublelayer]

As I said, losing a million jobs would be painful if it happened quickly, but nobody said it would be wonderful. Even AI fans who think that could somehow happen don't all say it would be utopia, though the stupid ones sometimes do. Even if those million losses were packed into one month, that would probably mean a bad recession, but it wouldn't eliminate the economy by making it impossible for people to buy things which is what Doctor Syntax suggested would happen.

[MachDiamond]

"I think we saw a few weeks ago the amounts of money per head of the world population that would need to be spent to make the investments bandied about at that time recoupable and it was unrealistic."

And, that's just power/energy. What's being dumped into hardware and facilities is going to have to be updated on a moderately short schedule and that's going to be unrealistic on that front too when they run out of OPiuM.

[Anonymous Coward]

Yeah, $635B is like 1% of US GDP ('figures from the IMF' link) or $1800 per person ... Claude Max is $200/mo so if everybody (all ages) subscribe that's 9 months to recoup the investment, 8 to 10 years if 10% of folks subscribe, 100 years with 1% subscription. Adding running costs and the need for frequent hardware upgrades (to keep up) suggests the ROI will occur sometime between a great many years into the future, and ... never ever.

LSD is stronger than reason!

[Bitsminer]

Another factoid: without the AI build out the US economy would have shrunk in 2025.

Trumponomics indeed.

[MachDiamond]

"Another factoid: without the AI build out the US economy would have shrunk in 2025."

I think that's hard to say. Investment may have gone into other things and done even better as AI is losing massive amounts of money using real measures every month. It's a big system with lots of moving parts.

[retiredFool]

I do know a woman who should know better, was bored during the recent snow/winter event on the east coast used chat for fun for a few days (free). She liked it, for two weird reasons to me. It complimented her profusely, she said better than her therapist. And it remembered what she said. I was a little freaked out to be honest. This is not a dumb woman, but got tricked by some stupid compliments. Fortunately she is smart enough not to pay for it.

[MachDiamond]

"It complimented her profusely, she said better than her therapist."

A good therapist will be working towards getting somebody to build a solid self-image more than just erecting a flimsy structure of shallow compliments that make the patient feel good for a short period of time. Yes, it's nice to get compliments but a therapist being involved usually means there are more complex things to be solved.

Let's also look at the megalomaniacs that get told by everybody how clever they are when they are only just a wee bit more clever than their adoring fans that are not far above a box of rusty hand tools.

[Like a badger]

Is there any way that there is sufficient income that could be raised from this to provide sufficient return to pay off all of their investments?

Much depends on the lifetime and utilisation of the investment, and the definition of "paying off the investment". Relevant considerations are the cost of capital, how long will an LLM GPU be a viable asset for, and will it have a resale value when the asset owner replaces it. I'm too lazy to do the calculations, but I reckon that to break even they'd collectively need to raise an extra $130bn a year in respect of these specific investments. That could be doable, if companies are willing to pay.

[retiredFool]

Curious, any idea how long before a cloud cpu/gpu actually gets replaced? I tried searching but came up empty. While cpu improvements are not much lately, I think gpu still has a pretty fast upgrade cycle. So cpu maybe 4 year schedule and gpu maybe 2 seems reasonable. Although are people replacing x86 with arm, or just adding arm instead of x86. That time before replace is crucial to the calculation of capital cost. Normally the bean counters want accelerated depreciation to write stuff off quickly, but the ai guys are trying to slow it down to minimize the perceived losses. I think they were pushing for a 6 year schedule instead of I think 3 for computers.

[MachDiamond]

"Curious, any idea how long before a cloud cpu/gpu actually gets replaced?"

There won't be good data at this point. With no imperative to keep from splashing loads of red ink, the drive is for bigger, better, faster. If some new chip is 5% better, the old stuff has to go. It's also early days so there will be rapid improvements as hardware is optimized quickly. It has to hit something of a plateau before the replacement interval gets to years.

I get crazy deals on solar panels as commercial fields are re-paneled to take advantage of improvements in efficiency. There's loads nearby so a few resale companies that buy all of the panels and sell them off to turn a profit when that company has lots of fields to put pallets full of them.

A question I have is if these GPU's being used for AI DC's will have any afterlife. A PCI bus GPU card used for crypto mining can be used in a PC in some cases. They fetch less money as they live a tortured life. The GPU's can also be harvested to transplant onto non-specialized graphics cards if they come from headless cards as parts donors.

[Bebu sa Ware]

Clowns riding astride asses are horsemen ?

I am certain this ding·a·ling circus can sort out among themselves who is going to be War, Famine, Pollution^† and Death.

_{† following Good Omens but with the connivance of another clown show under RFKjr Pestilence is making comeback tour.}

[Roland6]

Shoe Event Horizon

For some strange reason, the scale of the AI expenditure and build-out discussed in the article brought this to mind...

[Anonymous Coward]

Re: Shoe Event Horizon

... expect a geological deposit of compressed AIs to form as we variously evolve into actual smarty pants, golf-playing cockatoos, and housekeeping cows (imho)!

[elsergiovolador]

Carousel

Imagine if you own all companies involved in these transactions.

Numbers are getting more ridiculous, general public is poorer and tax man is asleep.

[Filippo]

I don't understand much about economics. But this feels... bad. Wrong. The kind of thing that goes down in history books as "the period immediately preceding...". There is no way that sentence ends well.

[EricM]

> But this feels... bad. Wrong.

Let me add "surreal"

> The kind of thing that goes down in history books as "the period immediately preceding...". There is no way that sentence ends well.

True.

AI is still unreliable and non-deterministic 2 years after they pretended to sell "PhD-level"-AI. Making models larger so far only resulted in more subtle hallucinations.

Adding additional layers (agentic AI, orchestration) just adds non-deterministic control to already non-deterministic processes.

AI output still presents a bell curve from occasional perfect gems, over many quite good results with errors mostly correctable by a human expert to total fails.

Probabilistic results from a probabilistic process.

At present, the state of the AI industry reminds me at the period immediately preceding the sub-prime crisis ~2007, when many market participants already had understood the system would hit a wall soon, but continued to promote it anyway, as they themselves were heavily invested. Taking a cold, hard look at reality and shift their public stance would have devalued their own investment immediately.

So, pretending all was going really awesome as long as possible seemed like the only way forward.

Avoiding pain as long as possible is pretty human, after all.

[Roland6]

>” but continued to promote it anyway, as they themselves were heavily invested.”

You can see why so many believe and promote sunk costs even though we know it’s a fallacy..

[retiredFool]

And some what ironically, the 4 big guys are costing themselves money. By starting this feeding frenzy of must have now, not only are we paying more for memory and chips, so are they. The real profiteers much like in the gold rush days are the people selling the shovels. Nvid, TSMC, SK, Micron, ... The shovel makers will do ok I think even if the AI thing collapses, simply because we consumers will buy the shovels for PC's, cellphones, ... because we delayed purchases because shovel prices were so high. The shovel makers will experience reduced profits though, eventually their overproduction of shovels will cause oversupply after we recovered from our delayed purchases and the shovel makers will need to adjust production.

[MachDiamond]

"eventually their overproduction of shovels will cause oversupply after we recovered from our delayed purchases and the shovel makers will need to adjust production."

A shovel could find a new home. What the hell do you do with a warehouse full of used Nvidia H100's?

[Excused Boots]

"A shovel could find a new home. What the hell do you do with a warehouse full of used Nvidia H100’s?”

Could you possibly turn them in to shovels?

[retiredFool]

Not the shovel maker's problem. Those are bought and paid for by AI guys. No the shovel maker will start making gaming gpu's again, and will price high until demand is satiated. That is the point the shovel maker has a problem. Nvid probably has a bigger problem than TSMC/Sammy/SK/Micron. Their shovels are much more specific.

[MachDiamond]

"Not the shovel maker's problem. "

From that level, yes, correct. Moving up a few steps, policy makers need to consider if the landscape is going to be littered with handle-less shovels that nobody has a need for. It's not something they should be subsidizing or formulating policy that specifically benefits those endeavors. In a free market, those companies can sink or swim whether what they do has a long term benefit or not. My issue is taxpayer money being used to encourage the waste/risk.

[Snowy]

The more you spend

The more you save!

[Ken Shabby]

Re: The more you spend

Look at this great deal on a DC I got, saved you millions.

[Tron]

Railway mania for a digital age.

This will go pop. Nobody is going to spend enough money on AI for it to ever turn a profit. If you are an investor, get out now whilst you still can. As with a duff firework, everyone stand well back.

It's a pity they didn't spend the money on healthcare, water infrastructure and climate-proof homes, but stupid is as stupid does. It's their money. They can waste it on whatever they want.

Most of the data centre stuff won't happen. It will all go pop before it does. Big tech will ride it out with smaller cash piles and buy the wreckage of anyone daft/greedy enough to get in too deep.

Worry less, but insulate yourself from the inevitable damage as best you can. We have experienced the dot com bubble and sub prime mortgages. There is no excuse for getting caught up in a repeat of that.

[Alan Mackenzie]

Re: Railway mania for a digital age.

> It's their money. They can waste it on whatever they want.

If only that were the case. It's not. It's our money too.

It's our money when 64 MB of RAM costs 800€ and rising, rather than 250€, when our electricity bills go up "temporarily" to bridge over the time until the data centres will supposedly start paying.

It's our environment they will be polluting needlessly. It's our infrastructure that is going unmaintained, our healthcare and education which is barely adequate, due to not taxing these vast accumulations of wealth adequately.

And it's our money which will be being mis-invested in supporting these hallucinations.

When the crash comes, as it surely will, many of us will become poorer.

[MachDiamond]

Re: Railway mania for a digital age.

"It's a pity they didn't spend the money on healthcare"

It's been tried, but most of the money winds up with the lawyers.

[Pascal Monett]

$630+ billion

Could somebody please remind me how much they pay in taxes ?

Oh, right. Zero.

[TimMaher]

Ice cream van.

Where is the fifth horseman?

[MachDiamond]

Re: Ice cream van.

"Where is the fifth horseman?"

Ronnie?

He doesn't finished his delivery rounds until half 10.

[Boris the Cockroach]

i feel its like

this

1. build huge DCs running AI

2. ?????????

3. profit

of course when 3 doesnt happen and all these companies are on the hook for the huge amount of cash splurged on data centers and they look like they are going down, then the whining "too big to fail" starts, and magically, the costs of the overpriced DC build outs suddenly becomes the taxpayer's problem....

[David Hicklin]

Re: i feel its like

> "too big to fail" starts, and magically, the costs of the overpriced DC build outs suddenly becomes the taxpayer's problem.

Governments are broke and the taxpayers on the verge of open revolt if they get taxed more - so sorry chums, no bailout. And remember that they are not banks....

[spireite]

Bubble, bubble, toil and trouble...

I'm sure I speak for many when I say I look forward to the implosion of this bubble.

I use it in my role, and it's used elsewhere in my employer. It has its place, but boy - is it overrated.