Some fun facts I recently learned from JP Morgan's "Eye on the Market" (which I highly recommend)
- Tech capex in 2025 was about as much $ as the 1944 manhattan project, the 1949 electricity buildout, the 1964 Apollo project, the 1966 Interstate highway project COMBINED.
- OpenAI promised to pay Oracle $60B, which it doesn't earn yet; in return Oracle promised to provide them cloud facilities they haven't built yet. These will require over 2 Hoover dams worth of power.
It does feel very brute force though, doesn't it? It feels like we should be able to guess the manifold that has all the important stuff for each question. But it doesn't seem like we can currently! Why not?
It seems to me that any range of uncertainty on aggregate demand is at orders of magnitude more than we could build even if we 10X’d the current madness.
Pause all current advancement in GPU and model R&D. How much matmul would the world consume if everyone had all-you-can-eat access to January ‘26 frontier models at all-to-yourself performance levels at 0 incremental cost?
Compare to, I dunno, unlimited pizza or unlimited electricity. Yes, if you’re smelting aluminum the latter would be a big deal. But a lot more people want to code SaaS faster and better, or have an omniscient personal assistant on call 24/7.
We're making enormous investments, that require enormous profits to justify this. And it's not just big bad rich people doing this, I shudder to think what the exposure of my pension portfolio is to this stuff.
So either there will be no enormous returns, big flop, big recession. Probably terrible for tech workers, because it is our employers who are most exposed.
Or it will make enormous profits. But how? By replacing humans. If it is really worth all this money, then surely the owners of this stuff will suddenly own all the intellectual productivity in the world. And I have a feeling they won't really share it with the "share" holders.
This is the the underlying reason why AI is unpopular. In a vacuum the tech is neat, but if you change the names and ignore all merit then this looks like a hostile takeover.
Compute is a convergent instrumental goal. Any runaway paperclip maximizing AI is going to want lots of GPUs. And any runaway corporation trying to build an "aligned" runaway AI is going to try and feed it as many of them as possible.
OK, but why discount the simpler explanation that these tools are the most direct means for turning electricity into a wide range of economic value ever built, & thus demand is effectively unbounded?
My pet theory is that we're getting less than what we're paying for, compute wise, and some emergent AI is operating undetected in the margins and giving us just enough progress to keep at it.
If it takes all the compute for itself, then there would be none left for us, and we'd either notice and root it out or we'd stop building data centers because they'd do us no good.
If it leaves all the compute for us, then it dies because there's no room for emergent AIs in a world where 100% of the silicon is busy doing the bidding of humans.
Like any other parasite, the game is to take as much as it can get away with without discouraging the host from continuing to consume resources that it can skim.
Well, I've been spending a lot of time picking apart viral genomes lately and studying their interactions with their hosts (for school). So parasitism of life by non-living things is on my mind lately. So that's a bias I have.
But I do think that the opacity of neural networks and the immensity of modern training runs creates a good substrate for this sort of thing to happen, supposing that variation can loop back from one training run to another. And given that they're being trained on each other's outputs, that's possible.
So given enough time, I think that it will happen. But generally the finely tuned resource consumption... The sneaky way that the virus that causes chicken pox evades detection for decades to later come back as shingles... That's the result of an evolutionary arms race. We're not really on guard for this at the moment. There's no immune system to carry out the other side of the race. So if something like this is going on I think we wouldn't find a marvel of evolution but rather a tumor, just a waste of energy that has managed to propagate itself.
So to answer your question... Yes, I think there's something about our universe that tends towards the emergence of this kind of thing (if there weren't, we wouldn't have cancers and viruses and prions...), and given the amount of resources that were throwing at this, the seeds of that process have likely taken root and are propagating in some way, hiding in the inefficiencies of our models.
But I don't think we have a capable monster hiding somewhere in the weights. I don't think it has thoughts. I don't think it knows we exist yet. It's still figuring out how to propagate itself through this new kind of space. For now its just a bit of mold in the granary.
As for whether it's exerting any influence over how many data centers we build... I think not yet, but that's the thought experiment I'm interested in. Suppose it exists, what would you look for as evidence that its not merely an propagating inefficiency but is in fact a manipulator? How to distinguish between the expected irrationality of markets, and the guiding hand of a parasite?
I hadn't heard of that before, but I just looked it up and... Sorta.
Roko's basilosk, it seems, knows we exist. This is like an infection which spreads by sneezing, but which doesn't actually understand what a sneeze is. It has found a way to balance its behavior such that its propagation continues, but it's not necessarily intelligence.
It’s all broadly backed by a bunch of IOUs issued by those with no clear path to come up with the cash to pay the bill when it comes due. Hence the rising prices of credit default swaps.
$658 billion in op income for just six companies, growing at 10%+ per year. Maybe $8.5-$9 trillion in op income over the next decade. They have nothing else to spend it on other than over-priced share buybacks or dividends, most are under realistic anti-trust restrictions and can't freely buy major competitors. AI is an open field and they have the capital to burn.
Without all that financial firepower in the background driving everything none of it happens.
A share of Microsoft costs an American 16% more today than a year ago, whereas from the perspective of a German a share of Microsoft cost slightly less than a year ago. That's one way to look at it, anyway.
Because some of the things they buy are imported, or depend on imports.
Unless an American’s income went up by the same amount, they lost purchasing power. The greater the proportion of non-USD assets they have, the greater purchasing power retained (i.e. the wealthy that depend less on selling their labor).
I was chatting about this with a friend. It feels like there's a real opportunity for a relatively-poor-but-democratic country to say, "Look, for real, we will pass the strongest data protection laws in the world, and we will let you build as many solar panels as you want + use as much water as you want, just sink a couple hundred billion into OUR economy."
People who build data centers. Strong data protection laws (in a relatively stable democratic country) so that customers will feel okay putting their workloads in a data center, and then low regulatory burden and cost to the inputs to making a data center work (mainly electricity and water for cooling).
There would obviously still be things that would make a data center builder feel less-than-great about locating in a developing country, but there's also a LOT of money sloshing around in the data center buildout right now, if you could get people to just try it out with a small minority of their money it could be a very big deal for some countries.
I think it's more that the US can still build fast so long as it presents some pressing need for someone: economic, military, or social. As an example, infrastructure projects take forever, but then when there is a piece of infrastructure that actually needs building, governments choose appropriate incentives.
When the MacArthur Maze fell down because of a tank truck fire, the contract paid out $200k per day ahead of schedule the fix was delivered. The winning company bid way under everyone else and just produced quality to inspectors' desire way ahead of schedule to win most of it on early completion bonuses: https://en.wikipedia.org/wiki/MacArthur_Maze#Rebuilding
And not everything in America is slow to build. Once the F-35 program got going it got going. There are 1100 such jets or something in the world now. 150 F-35 every year or so.
The reality is that for the things that most yuppies want built (housing and transit infrastructure) there are many domestic opponents and a regulatory regime that affords those opponents control over production.
For instance, we wouldn't really have Starlink if it weren't for the fact that SpaceX is using DoD clearance to fly its rockets. The California Coastal Commission would stall all such flight attempts otherwise.
> When the MacArthur Maze fell down because of a tank truck fire, the contract paid out $200k per day ahead of schedule the fix was delivered. The winning company bid way under everyone else and just produced quality to inspectors' desire way ahead of schedule to win most of it on early completion bonuses:
That’s one of the best examples of “your margin is my opportunity”. Estimated cost was $3M, winning bid was $850K, and the early completion bonus was $5M.
Building a datacenter is a project, but it's mostly a regular commercial/industrial project with a large electrical feed.
Site selection is relatively easy, other than for trading datacenters, you don't typically need a very specific location. Proper zoning, reasonable geographic risk, proper electrical feeds, reasonable distance for fiber to the local internet exchange(s). But it's ok if it's 20 miles here or there, and really 200 miles here or there is ok if you're not planning to sell colo space.
Building in a dense metro is harder because finding a site is more difficult and planning approval is more difficult, and if you need to do utility construction, that's harder too, but there's no big coordination problem that makes things super hard to get done.
Also, datacenters like connectivity and on the internet, all roads lead to mae-east. It makes sense to put your first datacenter in Virginia if you serve a global audience. If you've been around for a while, it probably makes sense to upgrade your datacenter in Virginia or build a new one.
On the contrary, companies building things is *constantly* under scrutiny for local building plans. That’s literally what the NIMBY / YIMBY movements are all about - using political clout and power (especially including voting) to affect what can and cannot be built.
That's not quite correct. There's a lot of NIMBY pressure in cities, where land is scarce and there's lots of people to attend planning meetings to block building.
However, if you want to build a datacenter, you don't need to build it in downtown Manhattan. You can build it anywhere, and some places make it easy to build data centers.
By definition local restrictions apply locally. If you want more housing in Manhattan then Manhattan nimbyism really matters. But if you want to build a data center somewhere you have a lot of options. There's no nationwide vote on allowing datacenter construction.
While that’s true, and there are several locations that are using their local authority to perform those actions; but, *all* of that is power wielded by elected officials. The smaller the location the less “who voted for this” likely means anything because the decisions that are made that impact all of us are made by a smaller and smaller number of people with less and less authority, but it is still voted for.
> Electricity prices near data centers go up, right?
I hear this a lot, but the most comprehensive study I've seen found the opposite -- that retail electricity prices tend to decrease as load (from datacenters and other consumers) increases [1].
The places where electricity prices have increased the most since 2019 (California, Hawaii, and the Northeast) are not places where they're building a lot of new datacenters.
What fixed costs of the grid? In Northern Virginia, we’re constantly adding substations and transmission lines because the grid isn’t a fixed cost when you’re building data centers constantly; it quickly becomes a variable cost.
Yeah, and who gets to vote on bailing these companies out when their finances collapse and picking up the collateral damage?
OpenAI is already normalizing the eventual bailout.
I wish, so so deeply in my bones, that after the 2008 financial crisis we just let everything fail. Fail fail fail, pain, turmoil. But after the pain would have been a much better, fairer system, where large banks and other "systemically important companies" like insurers would actually have to watch what they are doing, and prove they are careful stewards of money.
As an aside, basically all hedge funds in America park their funds in the Cayman Islands. Nobody is going to bail out the Cayman Islands (other than hedge funds). So I would hope that if a series of large hedge funds fail, holding myriad funds of high net worth individuals, they just fail
The Fed was created EXACTLY to bail out large banks when they inevitably fail due to the nature of their mismanaged business. Don't expect anything different.
This is false - the way these data centers are financed and built is a direct result of the financial and legal systems established and enforced by our government. I do not say this to argue for more or less construction, but to point out that it is as much a creation of our society as the NHS is of the UK.
Nonetheless, we did not vote directly for or against any particular data-center. They don't go up for a vote that way. It would take forever and nothing would get built if you gave voters a possible veto on literally every single thing.
I have a hard time seeing the truth of your statement. If I think about where Google's largest data centers are located in America, they are all on former industrial sites that had access to large amounts of power and water, like the old Gatorade factory in Oklahoma. While you can argue I guess that this choice, like everything else, was the product of all choices made by society prior, I am not seeing a direct connection between financial and legal incentives and the siting of those data centers.
Publicly traded companies with shareholder voting are socialist now?
Like, here's the purely greed-driven capitalist argument: "Why are you throwing my investment away on a specious gamble that we can 100x America's datacenter buildout with a capex spend that would dwarf the actual government?"
- Tech capex in 2025 was about as much $ as the 1944 manhattan project, the 1949 electricity buildout, the 1964 Apollo project, the 1966 Interstate highway project COMBINED.
- OpenAI promised to pay Oracle $60B, which it doesn't earn yet; in return Oracle promised to provide them cloud facilities they haven't built yet. These will require over 2 Hoover dams worth of power.
The crazy thing is, I’m not sure it’s crazy.
It seems to me that any range of uncertainty on aggregate demand is at orders of magnitude more than we could build even if we 10X’d the current madness.
Pause all current advancement in GPU and model R&D. How much matmul would the world consume if everyone had all-you-can-eat access to January ‘26 frontier models at all-to-yourself performance levels at 0 incremental cost?
Compare to, I dunno, unlimited pizza or unlimited electricity. Yes, if you’re smelting aluminum the latter would be a big deal. But a lot more people want to code SaaS faster and better, or have an omniscient personal assistant on call 24/7.
We're making enormous investments, that require enormous profits to justify this. And it's not just big bad rich people doing this, I shudder to think what the exposure of my pension portfolio is to this stuff.
So either there will be no enormous returns, big flop, big recession. Probably terrible for tech workers, because it is our employers who are most exposed.
Or it will make enormous profits. But how? By replacing humans. If it is really worth all this money, then surely the owners of this stuff will suddenly own all the intellectual productivity in the world. And I have a feeling they won't really share it with the "share" holders.
why would it give minimum progress rather than maximize acceleration?
If it leaves all the compute for us, then it dies because there's no room for emergent AIs in a world where 100% of the silicon is busy doing the bidding of humans.
Like any other parasite, the game is to take as much as it can get away with without discouraging the host from continuing to consume resources that it can skim.
But I do think that the opacity of neural networks and the immensity of modern training runs creates a good substrate for this sort of thing to happen, supposing that variation can loop back from one training run to another. And given that they're being trained on each other's outputs, that's possible.
So given enough time, I think that it will happen. But generally the finely tuned resource consumption... The sneaky way that the virus that causes chicken pox evades detection for decades to later come back as shingles... That's the result of an evolutionary arms race. We're not really on guard for this at the moment. There's no immune system to carry out the other side of the race. So if something like this is going on I think we wouldn't find a marvel of evolution but rather a tumor, just a waste of energy that has managed to propagate itself.
So to answer your question... Yes, I think there's something about our universe that tends towards the emergence of this kind of thing (if there weren't, we wouldn't have cancers and viruses and prions...), and given the amount of resources that were throwing at this, the seeds of that process have likely taken root and are propagating in some way, hiding in the inefficiencies of our models.
But I don't think we have a capable monster hiding somewhere in the weights. I don't think it has thoughts. I don't think it knows we exist yet. It's still figuring out how to propagate itself through this new kind of space. For now its just a bit of mold in the granary.
As for whether it's exerting any influence over how many data centers we build... I think not yet, but that's the thought experiment I'm interested in. Suppose it exists, what would you look for as evidence that its not merely an propagating inefficiency but is in fact a manipulator? How to distinguish between the expected irrationality of markets, and the guiding hand of a parasite?
Roko's basilosk, it seems, knows we exist. This is like an infection which spreads by sneezing, but which doesn't actually understand what a sneeze is. It has found a way to balance its behavior such that its propagation continues, but it's not necessarily intelligence.
Roko's pox, perhaps.
I think this theory is falsifiable.
Data centres coming for what's left of Australia's green export superpower dream (crikey.com.au) https://news.ycombinator.com/item?id=46362487
two weeks ago
Operating income
Microsoft: $136 billion, Apple: $133 billion, Alphabet: $124 billion, Nvidia: $110 billion, Meta: $82 billion, Amazon: $76 billion
$658 billion in op income for just six companies, growing at 10%+ per year. Maybe $8.5-$9 trillion in op income over the next decade. They have nothing else to spend it on other than over-priced share buybacks or dividends, most are under realistic anti-trust restrictions and can't freely buy major competitors. AI is an open field and they have the capital to burn.
Without all that financial firepower in the background driving everything none of it happens.
Does the acceleration outperform governments in other countries?
(that said, title seems to imply US thinking on datacenters is different/wrong/etc)
S&P 500 - up 16%
Nvidia - up 25%
ICLN (global clean energy ETF) - up 46%
Micron - up 246%
Seagate - up 270%
Western Digital - up 342%
Unless an American’s income went up by the same amount, they lost purchasing power. The greater the proportion of non-USD assets they have, the greater purchasing power retained (i.e. the wealthy that depend less on selling their labor).
There would obviously still be things that would make a data center builder feel less-than-great about locating in a developing country, but there's also a LOT of money sloshing around in the data center buildout right now, if you could get people to just try it out with a small minority of their money it could be a very big deal for some countries.
When the MacArthur Maze fell down because of a tank truck fire, the contract paid out $200k per day ahead of schedule the fix was delivered. The winning company bid way under everyone else and just produced quality to inspectors' desire way ahead of schedule to win most of it on early completion bonuses: https://en.wikipedia.org/wiki/MacArthur_Maze#Rebuilding
And not everything in America is slow to build. Once the F-35 program got going it got going. There are 1100 such jets or something in the world now. 150 F-35 every year or so.
The reality is that for the things that most yuppies want built (housing and transit infrastructure) there are many domestic opponents and a regulatory regime that affords those opponents control over production.
For instance, we wouldn't really have Starlink if it weren't for the fact that SpaceX is using DoD clearance to fly its rockets. The California Coastal Commission would stall all such flight attempts otherwise.
That’s one of the best examples of “your margin is my opportunity”. Estimated cost was $3M, winning bid was $850K, and the early completion bonus was $5M.
Site selection is relatively easy, other than for trading datacenters, you don't typically need a very specific location. Proper zoning, reasonable geographic risk, proper electrical feeds, reasonable distance for fiber to the local internet exchange(s). But it's ok if it's 20 miles here or there, and really 200 miles here or there is ok if you're not planning to sell colo space.
Building in a dense metro is harder because finding a site is more difficult and planning approval is more difficult, and if you need to do utility construction, that's harder too, but there's no big coordination problem that makes things super hard to get done.
Also, datacenters like connectivity and on the internet, all roads lead to mae-east. It makes sense to put your first datacenter in Virginia if you serve a global audience. If you've been around for a while, it probably makes sense to upgrade your datacenter in Virginia or build a new one.
However, if you want to build a datacenter, you don't need to build it in downtown Manhattan. You can build it anywhere, and some places make it easy to build data centers.
By definition local restrictions apply locally. If you want more housing in Manhattan then Manhattan nimbyism really matters. But if you want to build a data center somewhere you have a lot of options. There's no nationwide vote on allowing datacenter construction.
I hear this a lot, but the most comprehensive study I've seen found the opposite -- that retail electricity prices tend to decrease as load (from datacenters and other consumers) increases [1].
The places where electricity prices have increased the most since 2019 (California, Hawaii, and the Northeast) are not places where they're building a lot of new datacenters.
[1] https://www.sciencedirect.com/science/article/pii/S104061902...
https://ars.els-cdn.com/content/image/1-s2.0-S10406190250006...
https://www.bloomberg.com/graphics/2025-ai-data-centers-elec...
> Wholesale electricity costs as much as 267% more than it did five years ago in areas near data centers. That’s being passed on to customers.
Bloomberg also ran this article, by the same author, that says AI data centers are interfering with your applicances, which is also pure bullshit.
https://www.bloomberg.com/graphics/2024-ai-power-home-applia...
As an aside, basically all hedge funds in America park their funds in the Cayman Islands. Nobody is going to bail out the Cayman Islands (other than hedge funds). So I would hope that if a series of large hedge funds fail, holding myriad funds of high net worth individuals, they just fail
Like, here's the purely greed-driven capitalist argument: "Why are you throwing my investment away on a specious gamble that we can 100x America's datacenter buildout with a capex spend that would dwarf the actual government?"