I can’t see why, because as far as I know, nobody’s trading derivatives on AI real estate. That’s not to say securities don’t have a worrisome exposure to AI, but as far as people building out a bunch of compute capacity that doesn’t end up getting used, that’s simply not happening. Software is eating the world, it’s unthinkable that there would exist a glut of compute capacity that’s not gobbled up instantly if the price drops a bit. If someone passed a law to outlaw AI tomorrow (unrealistic, and easily circumvented), that capacity would be snapped up by other companies and consumers who are currently starved for it.
AI is being used for a lot of things. Some good, some bad, and some pointless. But there’s never going to be a point where it’s determined to be worthless and dies on the vine. Even if you resist and reject it with all your soul, it will be doing backend work outside your awareness. I know it’s an annoying cliche at this point, but it’s here to stay, because the core use case of reducing labor costs is an eternal feature of the human condition. We want to work less, we’d rather others work for us. There’s nothing to be hashed out at this point except who gets paid and who loses out.
Did you (the collective you) ever read any of the science fiction stories that involve a computer achieving consciousness? (The ones that inspired the Terminator films, for instance.) Some of them involve a computer achieving consciousness once it becomes sufficiently complex. So I almost wonder if the people behind huge data centers like this new half-trillion-dollar one think the same thing will happen in the real world.
I don’t think this is true. The benefits of more model parameters have already plateaued. Many experts believe they need a fundamentally different architecture to make a step change in capability. Nobody knows for sure what that architecture is yet.
I heard some commentator on TV mention that each time you get a google AI summary for a search, that’s $10 in compute power. Not sure how true that is, but if it’s even close to that, it gives an idea of the scale involved.
Another comment was that AI construction is sucking up more and more of the money in the financial world available for lending. One pundit said, this was enough to create the scarcity to contribute to driving up the rate for long term treasury bonds (and for multiple countries). Before the war started, the AI companes were at the point of hitting up the Gulf states, getting loans from them in return for building data centers there. (I assume uncertainty and lack of cargo shipping has impeded these plans)
The problem is not the demand going away, it is that the demand is not going to keep up with the supply, and these data centers are not going to be viable at prices that can pay off the loans made to build them.
I saw the end of the bubble from the inside. My company built servers, and we were selling as many as we could build. When the bubble burst, the demand did not vanish, but there were tons of companies that went bust with a supply of up to date servers which they tried to sell, driving down the prices and cratering our sales.
There was an analogy to the data center builders and runners also. No one built boards back then, we went to contract manufacturers who built our boards for us, which we then assembled into systems. The deal was that they bought the components and we’d pay for finished product. When the bubble burst there was no demand for product and the CMs wound up eating it.
Today is worse, since we didn’t finance the CMs and so we weren’t as affected by their problems. And the magnitude of the investment was much smaller. Plus, as mentioned, the economy in general was in better shape than it is today.
I suspect that when the bubble bursts tokens will get cheaper. Already some of the companies with unlimited token use budgets are cutting back. Demand might increase slightly, but probably not enough to make the revenue of most AI companies come anywhere close to paying their debt off.
I saw an article in the Times today about bond yields soaring for this very reason.
I think you misread the reference. That’s not total investment, that is one data center. If the total investment was $500 billion we wouldn’t be having this conversation.
The difference is that back then we were building for anticipated demand in servers. I was in telco at the time, we had built out something like 10 years worth of lead capacity that never materialized.
The difference is that AI demand cannot keep pace right now, before we even get around to forecasting future demand for it. Prices are rising, consumers are fighting a losing battle to figure out how to demand less from it.
All of the AI datacenter demand currently being planned out… it might meet needs for the next 2-3 years. Maybe. There is nothing on the horizon suggesting why any demand destruction ought to take place. Web servers ultimately are a marketing and delivery tool, AI is a labor reduction tool. The former waxes and wanes in demand, but demand for labor reduction is always high. The economics aren’t comparable.
Do you think there is an infinite demand for AI, a demand which will pay enough? I wonder what you’d get if you add up all the market share percentages of the big players, even with very optimistic views of the market. Way over 100%, I suspect. The crash will come when the players who are behind have to face reality and shut down. And back in 2000 we had a hard time keeping up with current demand also.
Demand for data centers needed for training is generated by the AI companies, but that doesn’t generate any revenue. Maybe the top companies will generate adequate revenue, but the second and third tier companies need to do just as much training for probably less reward. When they give up, that increases the hardware supply which hurts the revenue of Nvidia etc., which is just what happened 25 years ago.
But labor reduction has to pay. If AI was free or near-free, no problem, but if it gets priced to pay off the investment it might be cheaper to hire people. It happens. A big company used people in its SE Asia plants because their salaries were less than the costs of automation.
We also have the problem of getting rid of so many people that demand for the products of all these companies craters.
When the smoke clears, just like after the bubble, there will be AI companies left, for sure. But it is going to get very painful.
I live in SW Ohio. Rarely a day goes by without an oversize tractor-trailer with its escorts groaning across the tri-state as it heads East. Huge metal boxes full of …what? computers? servers? some of them are huge machines looking,like chillers to cool water or A.C. heat exchangers. This has been going on for months
I have a few things I don’t hear being discussed in regards to data centres. One is, if they can ship all our data to a US data centre, (this is the reason they’re pushing in Canada, we must keep our data at home!) via cables or whatever, then why not move the data centres away from inhabited areas?
Especially in Canada. As I understand it, the wild electric and water demand is due to needing constant cooling. Wouldn’t it make more sense to build these things in the far North, where cooling could be more easily accessed?
I realize what they want is to avoid building their own power station and water
infrastructure. Instead using existing resources. Which the taxes of the inhabitants paid for. Which will, regardless of what they tell us, most definitely drive costs for locals way up.
My second issue is this is pretty clearly emergent technology. It’s seems pretty certain, improvements will be made. A newer better version WILL appear. I’d like to live somewhere smart enough to wait it out, to be honest.
Then there is the constant hum. Such things have a measurable ill affect on people suffering with ADHA and other mental health issues. How’s that going to go I wonder?
I think politicians and industrialists are really failing to fully understand exactly how much people DON’T want these near their town. People are literally rising up in big numbers against these, in many different places in the world.
Modern society has slowly sucked up more and more dystopian seeming things, slowly, one small step at a time. I feel like this could be the tipping point, where people feel it’s a step too far. And may feel forced into action, to get the attention of deaf elected officials.
This is misleading at best. There are two steps involved in AI: First you have to create the models, and then you have to run the models. Running the models is really cheap. You can take a standard consumer-level PC, download an AI model onto it, and run it yourself. A query that takes it a quarter-second to respond takes about the same amount of computing resources as it takes to run a high-end video game for a quarter-second.
What’s expensive is creating the models. That takes a lot of computing resources, and requires huge data farms and a lot of money, and can only be done by big companies like Google or Anthropic.
And so, if you want to figure out the cost of a single AI query, you have to decide how to pro-rate the cost of that initial development of the model. You can ignore the development cost altogether, and just give the (very low) marginal cost, which is the sensible way to look at it in some contexts. You can take the development cost and divide it by the number of queries you expect to ever be fed to that model, and assign each query that share of the development cost. You can say that you don’t know how long into the future the model will continue to be used, and so just divide by the number of times that model has been queried so far. You can (but shouldn’t) even take the entire development cost and pin it onto the cost of a single query.
These data centers are being built with the expectation that the AI companies will constantly be building new and improved models. And so far, building new (larger, more complicated) models has resulted in models capable of doing more things (and hence making more money). But eventually (and some think that that’s already happening), that progression will slow or even stop, and the new, more advanced model will only be slightly more capable, or not at all, than the pre-existing model. And once that happens, all of these massive data centers will suddenly have almost no use. They will still be able to be repurposed for other computing demands, but there just isn’t enough demand for all of them.
And even if all data center construction were halted right now, and even if most of the existing AI data centers were shut down, we could still use the AI models we already have to do all of the things we already use AI to do. What we have now isn’t going away.
I understand its just one. Supposedly the world will spend about 7-10 trillion on AI investment between now and the end of 2030. Globally we will probably spend tens of trillions on AI investments in the 2030s too.
My point is that AI enhances problem solving across the board. In the long run that multi trillion dollar investment will be worth it just like investing large amounts of money into electricity or paved roads was a good investment in the long run.
I would push back on computing power and better models (is that parameters/architecture?) being interchangeable. In my mind, computing power is the energy to run it. The parameters are how many connections can we make to make AI even better. Even with a plateaud parameter, you still need more computing power and AI can still get “smarter”.
For example, you need to train this same model on more and more data. That increases it’s knowledge. That takes more computing power. Or, I prefer Claude to use more effort for my answers. That makes them more accurate. That takes more computing power.
Same level / parameters. More computing power will make it better (or at least needed to do that extra work).
Or simply, using more AI, requires more computing power, that in itself can make it better.
I wonder if this is where the tech “subsidized with venture capital, gain market share, enshittify” process is done on such a scale that it actually wrecks our economy. Because people here are right - it’s not like we’re building unused data centers. People are using the capacity. People are using the AI. But… they’re using AI at subsidized prices. Are they going to keep using AI when the companies start to enshittify and/or raise the price on their AI?
Almost all tech companies work this way. Let’s use Uber as an example. Uber had some genuine advantages over traditional taxis, like their app, but the fact that uber was cheaper was an illusion. They were using billions in venture capital to subsidize every ride so that more users would pick uber over taking a taxi. Consumers were convinced it was both more convenient AND cheaper, but it wasn’t. A sane society wouldn’t even allow this sort of deliberate market distortion.
Anyway, uber drives out their competition, weakens the cab industry, finally achieves a dominant market position, so of course now they need to turn a profit and the subsidies go away, and suddenly everyone is paying twice as much for their uber ride. Well, suck it up, this was the real cost all along, it was just hidden from you.
I wonder if we’re going to do the same thing with AI. AI is being used, but losing massive amounts of money on subsidized compute. When they have to switch to the profitability stage (which I think is at hand), is everyone going to be so eager to use it when they have to pay the real costs? Both in terms of actual money cost, but also in terms of companies running dumber models to save money, and finding alternate ways of monetization like integrating ads into AI output and collecting and collating massive data on users.
While your overall point that training models and running models are separate with separate costs is true, I think you are wildly understating the costs to run the models. Most queries are not something a consumer PC can run in a quarter of a second, and for that matter, I’m sure you’re aware that frontier models are orders of magnitude more difficult to run than the small models you’re referring to, in both hardware infrastructure and energy costs per query.
But not only that, but most people deploying AI at scale are doing something like evaluating entire code bases and massive documents. Every query reprocesses the entire context of the converation, including data like documents and code. There are techniques for reducing the cost of that like KV caching, but that costs a lot of money in infrastructure for storage too. And when AI queries do exist at the small scale, it’s often done in the billions of queries per day, like how search engines automatically give you an AI summary to the billions of searches per day. Training them is very costly, but running them is not cheap.
Training was 2/3 of the cost in 2023 but it has been trending down. In 2026 2/3 of the energy spent on AI is for inference rather than training.
From scratch, yes. But Same model, I’d say it’s a variant of that same model (eg, 4.1, 4.2, etc).
But really…I don’t know. This stuff is completely to me every couple months. It’s wild. I’m more curious about AI than knowledgeable about how it works.
Google processes about 5 trillion searches per year [1]and about half have an AI summary[2]. Google total spending for 2025 was $278B[3]. So even if all of Google spending was on AI summaries, it couldn’t cost them more than 11 cents per query. In reality, it’s less than a penny per query as most of the spend is for future data centers, not current demand.
Any commentator who is off by at least 2, probably 3 orders of magnitude is suspect on every other aspect of their analysis.
(Apologies for the terrible reputational quality of these cites, I was trying to use pure, non-AI Google to generate them to forstall accusations of using AI and Google search quality has gone so downhill these are the crappy cites it is generating).
That can’t possibly be even remotely true in the literal sense, because that would mean that Google spends 10 dollars for something like ever other search. In other words, they would be spending 500,000 dollars every second just for that summary.
My point here is that no-one should blindly trust reporting about the AI economy, even when it’s from a respectable outlet. The entire thing - from announcements to construction to reporting - is hopelessly hyped up.
The big question is, who is subsidizing it, and how?
I don’t buy that AI is being given away at far below cost right now, because those economics would’ve already failed. You can’t throw away trillions of dollars on a product that’s not at least approaching paying for itself.
Even if this were all VC pouring its own cash into it, that’s a good thing for the economy. It means people with too much money are unwinding their hoarded cash and putting it back into the economy, which is a good thing in any sense you look at it. If the investments go bust, fake value is either destroyed, or real value goes into the hands of the people who are developing them. In this way new technology ventures are successful even when they fail.
Back to the server glut and crash of the 2000’s. This was just a blip in investment, directionally correct but with flawed timing. What people were calling “surplus” supply in 2000 was entirely inadequate to supply demand even 10 years ago. Yes it’s possible that AI will have some financing hysteresis, most businesses do, especially hot ones. It would be surprising if there weren’t some temporary blip in associated economic activity.
But as things stand, the present book demand is unsatisfiable, let alone projected growth. Prices are rising. As providers switch to metered billing to capture the value of what otherwise would be approaching a free service, our customers are trying (and failing) to cut back on their token demand. The only real concern (at least to me) is that knowledge workers like software engineers are already reporting that they’re forgetting how to produce without AI. It’s debatable whether that’s a real problem that matters, but it’s unambiguously good for AI vendors.
So if you’re projecting a massive permanent crash in AI, you need a thesis for why you see sudden demand destruction on the horizon. And there aren’t any. People will inject their personal feelings about why AI content isn’t as good as human content, or why datacenters are problematic, or invent some fantasy that nobody’s finding real value in it right now. There are real externalities that need to be addressed. Datacenters need to pay full freight as far as taxes and utilities. But otherwise, these objections are just wishful thinking about why someone can and must halt progress that seems frightening.
The reality is that AI is as revolutionary as the microprocessor in the 1970’s. Which has seen its own peaks and troughs, but if you bet against its longer term success, you’d have been embarrassed many times over.