The news is full of stories about new data centers being built, and local residents complaining.
And I have some simple questions:
Who is building all these centers? And why?
How does each new proposed one differ from others? How many are currently proposed, both in the US and the rest of the world?
I understand that different AI’s will be trained on different data sets.
But what is the end game for each one?
What Is the specific need for it, which makes investors willing to risk the money? Is each one aiming for a specific niche, to be the leader in a specific industry? Or Is everybody just jumping in blindly, hoping to become the Google or Windows of AI, take over the world, and kill off the competitors?
Are all the data centers roughly the same size (measured by the amount of data they can process)?
How do they define their data sets, and what are the sources? I assume it mostly comes from the “regular” internet,(i.e. info freely available), so how does one center limit itself to relevant data, and know what to ignore from the wider internet.?
And ( IMHO territory ahead): do you think all these data centers are going to be necessary, or are a whole lot of people going to lose a whole lot of money?
Here in Oz, one of the mooted AI data centres is being promoted from IREN, which already has 5 in the US. They may be involved in other proposals in Australian. They are promoting themselves as providing AI compute for companies that want to do AI. They don’t do the AI themselves.
So whether this is a rain follows the plough business model or they already have major customers already waiting with money in hand is an open question.
They say Train powerful AI models and run inference at scale, fast.
So we are seeing a repetition of the business model that made Amazon’s AWS make more money than selling stuff. Provide services to companies with scalability. Companies can buy in small, and if things go well, it all seamlessly scales up and down as needed. A startup is insulated from capital expenditure and everything is lovely. So long as they come
Whether there is going to be the demand is really unknown. IMHO it is unlikely. The lead time for actually getting a centre built is such that if things don’t pan out, in many places all that will happen is the plans will fail though and nothing gets built.
Everything is currently paced by manufacturing. And that doesn’t scale easily. AMSL is making stepper machines as fast as they can, and the lead time to scale up is measured in years. Worse, history is replete with boom and bust cycles in the industry. There is very little incentive to be dropping billions on what is probably another mirage that could bankrupt them if they got greedy. The various tech bros moan about this.
The problem with the model IREN has is that it assumes that there are yet unknown killer applications that smart young startups will discover and make zillions on. The sheer scale of the build out is essentially saying that these yet unknown killer AI apps will be making hundreds of billions in only a few years. No doubt there are going to be some interesting and important breakthroughs powered by machine learning systems. But something that is of the same scale as the mobile phone? Difficult to believe.
One thing that does fuel data centre build outs in countries outside the US is the increasing perception that sovereign capability and control is becoming ever more important. Governments in a few countries have already been dismayed to discover their data is held and managed in other countries with unclear guarantees about privacy and access by foreign governments.
It’s difficult to say who’s building them, since a lot of data centers are shrouded in NDAs and multiple layers of shell companies. They’re deliberately making it hard to tell who’s building what. My guess is that that means that they’re mostly built by a small number of big players, who are trying to hide their monopoly or near-monopoly status.
Not sure who exactly is building them, but I’m pretty sure it’s private money with Gov’t incentives. You can build your own data center and use it all to yourself for your own AI (vertical integration; like Microsoft), or rent out the individual servers to several different AI companies (horizontal integration, like Softlayer).
Why? The main limitation from AI being better tomorrow is computing power. Computing power comes from hardware/Nvidia chips in servers. Servers take up space and needs to be maintained as it produces heat / requires cooling with water / needs electricity. AI needs enormous computing power. More. More. More. I’d imagine we’ll need to double the world’s computing power in a few years. So a massive amount. So, to get more computing power to make better AI, private companies are building a lot of data centers.
A limitation on data centers themselves is electricity, or more broadly, energy. Residents near data centers complain of them because they require so much of (their) energy and water.
A super basic understanding, and you might know all this, but anyways…it’s easy to think of AI as a software technology/problem, but it’s probably better understood as a hardware technology/problem.
I didn’t realize this. I thought each data center was owned by a company that wants to issue its own unique version of AI (like ChatGPT, Claude, Perplexity,) and then take over a huge market.
I can see how certain markets will need a specialized AI trained on data relevant to their market, and not available on the regular internet. (say, a medical AI trained on patient records, or a legal AI trained on court documents.) So whichever company becomes the unrivaled world leader in that niche will become the Google for that profession. Then all other companies which had tried to conquer that profession will fade away and die.
Which means that a huge amount of invested money will die, too.
Compute services to a large degree are fungible – not 100% exactly because the exact hardware can be a better match for some tasks than others, but “cloud” services and resources can be used pretty much interchangeably between hosting a video streaming service, web page hosting, running giant databases, hosting internal services for companies, etc. They’re also used by AI, though those require closer to specialized hardware (high end GPUs) but even then they can be used almost interchangably for any particular AI system. A prominent example of this is that xAI (Grok) built some of the largest computation clusters on the planet and then rented them out to anthropic who needed the extra computing power for several billion dollars a month. Those clusters don’t care if they’re doing computation for Grok or Claude. (Google is an exception here because they build their own hardware designed to work with their own AI).
So these are not monolithic. There’s no company trying to become the next google here. Rather, they’re just putting together a lot of computing and selling it to whoever may want to use it. Given how high demand is for both AI and traditional types of internet hosting and processing, it’s a pretty good bet that if you build a data center someone is going to rent out all of your services. Maybe you can lease it to one company or maybe a thousand different companies share your resources.
So it’s really just computation as a service rather than any specific mysterious new company.
It’s possible entire data centers are dedicated to something like amazon’s AWS or microsoft’s azure (I’m not sure how the actual contracts work in that case), but even then, those servers are designed to be resold/leased to other companies that need “cloud” computing. What’s really being sold is computation and hosting as a service and that’s pretty agnostic towards any particular corporation.
Maybe think of it like: normal CPU storage / computing (uploading pictures to the cloud, etc); Training AI; Using AI. You’re example is just talking about training AI.
Training probably needs a specialized data center, so sure, you build a data center to help get that AI model trained/better. I’d guess small/tight is better than spread out for that purpose. And it uses a lot of computer power to do that.
But also data centers are also (still) needed for all normal internet/computing stuff; and more relevant, to now be able to actually use the (now trained) AI on a computer, iphone, business, Gov’t etc. and the other millions of those around the world. That needs massive computing power too and the data centers needed to provide it. For these tasks, you want the computer power/data centers spread out for redundancy and reliability purposes.
I suppose the question is will we build too much computing power / data centers; more than we need. Unlikely. AI will never not need to be trained to get better; it will be used more and more as it integrates into everything. Etc. Etc. But that’s a guess on my part and I suppose the gamble of investing in data centers.
You first have to fundamentally understand the AI Boom story as a finance story. Certain large technology companies can only justify their valuation if there’s going to be an unprecedented AI boom happening. For an AI boom to happen, you need to be investing in the data centers now. If companies aren’t investing in data centers now, that’s a signal that they don’t believe in the AI boom and a cratering in their stock price so it’s more financially rational to build data centers they know they won’t use than not build data centers.
But the structure of how all this is happening is telling and deliberately confusing to the public. There’s lots of splashy announcements and giant dollar figures thrown around but the deals are carefully structured in a way that doesn’t bind long term dollars to a lot of projects and allows the private companies to walk away and leave local governments/power companies etc holding the bag if the AI bust happens.
IMO, the most important people to watch out for are the memory companies and what they decide to do. They’re the ones structurally holding the bag if the AI bust happens. They have long institutional memories of the boom and bust cycle of RAM pricing and how devastating it is to them to bring on new capacity just when the market starts to crater. They will never release any of these findings to the public but they have internal teams doing the proper stress tests of just how much AI usage will there in in 2029 - 2035.
It’s not like we haven’t seen this before, this was largely the set of circumstances that caused the original dotcom bubble. The difference is the dotcom bubble was overheating what was already a decently strong economy. The AI bubble is about the only bright spot on an otherwise dismal economy right now so there’s a lot of vested interests in propping it up just a few more years so the right people can cash out and pawn off the resultant crash onto the next generation.
Launched in 2007, Data Center Map was the first research tool of its kind. We operate a global data center directory, mapping data center locations worldwide. Our intention is to make it easier for buyers, sellers, investors, regulators and other professionals working with the industry to gain insights into the markets of their interest.
Click on Explore the map. There are dots all over the world. Estimates are about 12,000 worldwide, half in the U.S., with a thousand being large enough to be called “hyperscale.”
Hyperscalers and enterprises alike are discovering that raw compute capacity alone is no longer the differentiator. Instead, the focus has shifted decisively toward the speed of deployment.
In this new era, the primary metric for success is Time to Token - the end-to-end duration from initial planning and site preparation to the moment an AI cluster powers up and begins generating its first output tokens. …
Modern AI deployments demand deep, partnership-based orchestration that brings power, cooling, and hardware vendors together from day one. The power train and thermal chain should be co-designed alongside compute as an integrated stack.
This collaborative approach compresses deployment timelines from years to months with industry leaders increasingly designing infrastructure to be “silicon-ready,” with facilities prepared and waiting for graphics processing unit (GPU) shipments rather than the reverse.
We have quite a few data centers here in Oregon, mostly because they can take advantage of some liberal tax breaks. In eastern Oregon, they mostly seem to be owned by large corporations, rather than obscure ones you never heard of. Google has several in The Dalles, Apple and Facebook have them in Prineville, Amazon (AWS) has them in Morrow County. I expect there’s some over there not owned by big corps, but you never hear about them.
Here in Hillsboro is the largest concentration in western Oregon. They seems to be more obscure companies, although we do have some owned by Microsoft. For example, there’s 4 or 5 of them along a street I frequently bicycle on owned by QTS. I had never heard of them, but Google says they’re Korean company.
My amateurish understanding is that there are at least 3 major reasons
One is emergent properties. As AI gets more advanced, certain abilities will spontaneously arise. Its impossible to tell if some emergent property that has a massive financial or military benefit will arise or not. So there is a race to find these new emergent properties.
Another is recursive self improvement. As AI gets better, it is more able to contribute to the creation of the next generation of AI. Supposedly when AI is as competent as the best humans ever, and starts to surpass them, then there could be dramatic growth in AI competence in a very short window. They are all chasing AI advanced enough to play a more meaningful role in building and optimizing itself.
Another reason is the cold war with China. The US and China are locked into a cold war for the world superpower, and AI is a major factor in that. China is releasing open source AI models to compete with paid US models, to help slow down the US in the AI race by making private AI models unsustainable. Also the US is limiting chips and chip building technology to China. So AI is a major part of the cold war with China just like shipping lanes or proxy wars in Venezuela and Iran, or Trump’s rhetoric about Greenland and Panama are a big part of the proxy war with China.
A truly competent AI will provide massive military and economic benefits to the nation that has it. I’m guessing eventually all of NATO as well as other enemies of China (Japan, South Korea, Australia, New Zealand, etc) will work together to beat China in the AI race.
AI has now become a meaningless scare word, just as computers and robots used to be. Both of them were going to destroy jobs, run our lives, and take over the world. Today both are so intertwined with everyday reality we no longer even think about them. It’s hard to buy an led lamp that doesn’t have a hundred possible settings of warmth and brightness; robots flip burgers.
The implication is that AI will also become background in every application in every business, profession, school, and household. Lawyers will use AI trained on law, doctors will have AI trained on medicine, chefs will have AI trained on cooking and nutrition, mechanics will have AI trained on vehicles, and on and on and on.
The most mundane example I know came recently when I had a glitch on a WordPress page that wouldn’t properly generate a table of contents. I typed about five words into the AI assistant. It went through the table line by line, came to the glitch, corrected it, and went to the end. Somebody must have fed basic css and html and WordPress idiosyncrasies into a more general AI.
@Wesley_Clark gave a good high-level answer above. There’s also a mundane low-level answer to the end game: billions of AIs individually trained for tiny specific purposes, all generating tokens and needing access to data centers. The world will not function properly without them, just as it now would now function properly without computers.
AI Oprah: You can have an AI, and you can have an AI, and you can have an AI…
The usual riposte to this is that AI can be dangerous and destructive. Yes, it most certainly can. That is also true for every major technology in the history of mankind. We might have been better off as a species without the introduction of farming or the printing press. Good that we have myriads of examples from the past to alert us to take possible steps before the destruction begins. And provide some answers to the critical questions: who gets to define destruction, who gets to define good?
The first AI problem is money. Estimates I’ve read say that about $1.5T has been spent building data centers so far, and it’s not getting less over time. As mentioned above with RAM chips, it’s sucking up a lot of the current production of computer hardware. The effect is to create the aura of vibrant economic activity while burning through investor money and loans.
The big question is - who will pay back all this? What revenue stream will support the payback for all these data centers? Allegedly they will replace employees - which has its own downside - and then, if that’s the case, if you are paying the same or more for AI where’s the gain? Meanwhile we see more stories about how AI simply does not live up to the hype, or is costing too much. Where is the massive revenue stream to pay for it all?
It’s all one big bubble and likely to come crasing down soon. What happens to all these venture start-ups when the market for their facilities collapse, they can’t pay the electricity bill or loan payments, investor money dries up? Abandoned data centers sitting unused where the dissolved ownership cannot even afford security to keep them intact?
The only saving grace is that some are owned by otherwise stable companies with alternate revenue streams like Google and Microsoft and Amazon. However, unless the revenue is there, owning and running these centers is not financially viable so not worth taking on the excess burden to buy up the rest. And some real companies with real production - like those RAM chip makers, and NVidia - are going to hurt when the music stops.
The other question is how quickly do these data centers become obsolete as technology marches on? Unlike 30 years ago, your home computer does not become outdated in a year or two, but there is still technological progress to worry about. There’s a risk that when the bubble bursts, these data centers become the modern equivalent of all those abandoned rail lines across the country. Because, this has all the hallmarks of a financial bubble, the over-the-top hype and the scramble to get in on the action with larger and larger expenditures. Like other bubbles, burning through investor income is being mistaken for business activity.