Is the AI genie stuck in a bottleneck
27 claims analyzed
Transcript(27 claims)
Money talks from The Economist is sponsored by IDA Ireland. Ireland is a global hub for innovation and business with the talent and connectivity to grow. Explore how Ireland can support your company's global growth and investment at idaisland.com.
The Economist.
Artificial intelligence labs have a big problem.
AI demand is so white hot hot that the frontier labs cannot build data centers and find compute fast enough to serve their users.
Demand for AI is booming, partly because of the growing use of coding tools. But the industry just can't keep up. Firms like Google are increasingly supply constrained.
Obviously, we are compute constrained in the near term. As an example, our cloud revenue would have been higher if we were able to meet that demand.
This year, five hyperscalers, Alphabet, Amazon, Meta, Microsoft, and Oracle, will together spend more than $750,000,000,000 on capital expenditure. But adding AI capacity quickly is hard. Chips are scarce, and local opposition to energy hungry data centers is rising.
Protesters rallying today outside of a proposed data center in Saline Township. Data center is going to increase traffic, pollution, and everyone's electric bill is gonna go up. It's not okay. So
today, just how serious are the new bottlenecks for artificial intelligence? You're listening to money talks from the economist, our weekly podcast on the markets, the economy, and the world of business. In Washington DC, I'm Alice Forward.
In New York, I'm Mike Bird.
And in today's show, how the AI rush is hitting a bottleneck. Mike, hello.
Hey, Alice.
Have you come up against any AI bottlenecks?
I haven't really. I've been doing a little bit of vibe coding. I've been enjoying Claude Code. I think ultimately, I've come to the conclusion that it's fun, and I'll probably use it for small things. But I'm starting to think it's a bit like three d printing.
Do you remember the big three d printing craze in the twenty tens?
Basically, everyone was like, there's gonna be three d printing guns everywhere. That was, like Yeah. The the main vibe.
There was a more optimistic side of it than that. It wasn't just guns. But the idea was, you know, that, oh, you're you're gonna have a three d printing outfit in your garage, and you're gonna make things for yourself. And as it turns out, I don't I don't want to make things for myself. I want cheap, high quality products made by other people and I want to purchase them.
And I think that's probably true for now of of coding related stuff as well.
Yeah. I I haven't played around with the coding tools that much, but I am a big fan of co work, which I think is basically Claude code for like dummies who don't understand how computers work. And I'm a big fan of using this for lots of different tasks while writing stories. And I was indeed intensively using it the day that there was some sort of Claude outage when they sort of miscalibrated how many tokens people were using. And I did have the experience of going, oh, well, maybe I'll just increase my monthly spend limit a little bit.
You know, Claude, here's my credit card and it did feel rather like casino esque that I was like, no, give me more Claude, give me more. It turns out that was all kind of a mistake. They have refunded me my token money. Maybe I'm not as much of an addict as I feared, but yeah, I experienced the AI bottleneck in real time. So I'm very excited to talk about it more seriously on today's show.
So to help us dig into all of this, I'm delighted to say we have with us Shailesh Chitnis, The Economist Global Business Writer. Shailesh, welcome to the show.
Good to be here.
So Shailesh, you just written a big piece of the paper on AI bottlenecks. What prompted you to dig into this right now?
I think in recent times, has been a steady stream of news. Like you mentioned, people are experiencing outages in their favorite chatbot, in this case, Claude. It's either slow or they've changed their terms of service. So there is that kind of anecdotal side of things, but also some very real statements that you see from companies about a compute shortage. OpenAI, for example, recently shut down Sora, which is their video platform, quite abruptly to the surprise of many because they wanted to allocate resources towards more lucrative ventures as it were.
And a lot of tech companies have been talking about, quote, unquote, a compute crunch, which has limited their ability to grow as fast as they can. So I think that was quite interesting, and I decided to look into it.
I'm really enjoying the fact that the AI general narrative has flipped fairly rapidly from, I think, a totally reasonable discussion we were having not very long ago at all about whether the sort of AI build out might be too much. There might be too much capacity for the actual demand. This is the opposite of that story. Right? There's not enough capacity.
People aren't taking it seriously enough. They're not building enough to meet this sort of rampant demand. You can see this all over the place. Someone I was looking at the memory chip producing stocks this morning, which are on yet another tear. SK Hynix, one of the high bandwidth memory producers in South Korea is up, as of this point, slightly more than a 130% this year.
You know, it's still May. It's still May. We haven't done very much of this year. Tell me a little bit about what's happening here. What's driving the shift from worries that capacity was too much not that long ago to amazing capacity constraints now?
I think demand has really taken off in a way that people did not expect. So for example, you have these different coding tools that are now being used in a very big way in tech companies, and that has been one big driver behind the usage. And a good measure of this is look at the growth of tokens, which are the smallest units of text that an AI model processes, typically around three to four characters long. And that has been the metric that the tech industry is now using to measure AI demand. And so if you look at the growth in tokens, that has almost quadrupled this year between January and March.
And here are examples where tech companies most famously recently, Meta, had a leaderboard, which they since disbanded, where they were tracking employees who were the highest consumers of tokens within the company as a gauge of their AI usage. And so this term token maxing is now firmly in kind of Silicon Valley lingo in terms of how many tokens you use as a measure of how good of an AI user you are.
I'm really hoping we've hit max max and that we will stop applying this term to everything. I'm not hopeful, but yeah. Can I ask, is the advancement in the models also a part of this? Know, as the models get more sophisticated, more powerful, you burn through your theoretical tokens more quickly. Right?
Absolutely. I mean, as models become more efficient, the immediate reaction is you're gonna need less processing power. But in fact, the opposite happens pretty much the next day. And so as you get more powerful models, models that use memory more efficiently, whatever you have, the amount of usage that people are driving from it just continues to increase day on day.
Yeah. I guess it's like the better the model gets, the more induced demand there is. And, actually, that's outstripping any gains made by making the models more efficient. So this whole idea of issuing these tokens and letting people max them out or burn through them, this suggests that the AI labs are in some way trying to ration demand. They're sort of limiting people's access at certain times or certain amounts.
Why are they doing that rather than just increasing their prices?
So that's a good question. I think in the past three years, what you've seen is for a lot of these AI labs, their mode of growth has been to continue to drive the price of inference down. So over the past two years, the price of inference, which is when an AI model actually gets used, that has dropped by something of the order of 90 to 95%, and the trend is going to continue. You also have cases where, for example, I was in India. There, the prices that are offered for some of the AI models like Charge GPT and Claude are super, super cheap.
I mean, those prices don't exist anywhere else in the world. So the entire focus is on driving as many users as possible to drive as many usage as possible. And so when the demand is increasing at this space, the only other lever that is left for them is to actually dial up or dial down the token usage either by encouraging users in off peak times or burning through their tokens a lot more quickly so that they can actually constrict demand.
So this is putting a lot of pressure on the hyperscalers, big cloud computing companies, basically AWS, Google, Microsoft. How are they able to respond to this? I feel like there's a moving target. It's a number that's constantly updated upwards in terms of their CapEx. Where do things stand at the moment?
I think last week, most of the big tech companies came out with their earnings. And what is not surprising is they updated again how much they plan to spend just this year. So when we started 2026, there were estimates that the five big hyperscalers, which include Meta, Microsoft, Amazon, Google, they plan to spend roughly on the order of six hundred or six fifty billion dollars on capital expenditure. Last week, that number was up to 725,000,000,000, and I'm pretty sure by the end of the year, it's gonna approach 800,000,000,000. So their approach is to pour a lot of money in building out AI infrastructure to actually meet this demand.
Now even as the hyperscalers are pouring money, they're running into some real physical shortages. Two in particular stand out. So the first is actually building out these data centers takes time in terms of acquiring the land, getting the power into data centers, the whole physical infrastructure. And the second one, which has been quite a surprise, is they're just running out of chips. These are GPUs, the kind of chips used to run the AI models.
Even CPUs, which used to be this old workhorse of the tech industry, which have forgotten but have now suddenly come back in vogue, memory chips, as Mike alluded to. Basically, the tech world is running out of enough chips to power these data centers.
Okay. Let's take this in turn, I guess. We can start with chips. What are the chip companies doing? Are they expanding their own capacity to meet supply?
Do you think they'll be able to do that?
So the companies are expanding now. It's tricky simply because you have a few companies that are quite dominant across various parts of the stack. So for example, NVIDIA is the dominant player in making the AI chips. For memory chips, you have predominantly the two big South Korean memory makers, SK Hynix and Samsung and Micron of America. So there are really few companies that make most of these chips, so that's one bottleneck.
But the biggest bottleneck in in producing enough chips is really making them, and everything pretty much ends up at TSMC, which is a Taiwanese chipmaker, which produces, I think, most of the AI chips that are sold in the world at the leading edge. Now the challenge is you can't expand capacity quickly enough when making chips. A typical fab takes anywhere between two to three years to fully build and equip, and it costs around 20 to $25,000,000,000, and the costs just keep going up every generation. Now TSMC has committed to around $65,000,000,000 in capital expenditure this year, which is almost 30% more than what their figure was last year. So it's it's a pretty big sum from their perspective, but it's not large enough in terms of what the software makers want.
They actually want TSMC to do a lot more, and currently, that's where a lot of the bottleneck is.
This is a fascinating conversation because anyone looking at the stock chart for a company like TSMC might be surprised to hear that this is a sort of relatively financially conservative company. You know, despite the extremely rapid growth, because it's at the center of the supply chain, you invest incorrectly once. You mentioned $2,025,000,000,000 dollars there. You don't get to make that mistake that many times before investors start to get quite upset. Is that sort of financial conservatism at play here?
Is this a reason that there's less of maybe of an acceleration to ramp up supply?
Yes. I mean, some have uncharitably called TSMC a, quote, unquote, natural break on the AI boom. I don't think they see themselves as quite that way. But the semiconductor industry has been through cycles. So during time of booms, they build a lot, and then there's a bust, and that just happens with regular precision.
And and a typical cycle is between four to five years when you go between a boom to a bust. And so TSMC is very cautious because they've been through that cycle. You don't want to be left in a situation like you mentioned, Mike, where you build for the forecast demand. Probably only 50% of that shows up, and then your fabs are lying idle, that severely hits their margins.
Let's talk about the other big constraint on the boom in AI, which is data centers, these massive warehouses filled with all of these chips. If you can get your hands on them, they need a huge amount of power, they need huge amount of water, you know, some of them the size of small cities. People don't always want them to be built in their neighborhoods. So this is not necessarily the easiest task to construct these things. How much of a constraint is all of this posing on the AI labs?
Yeah. So that's a pretty important question. And to find out more about this, I spoke with someone who knows The US energy market inside out. His name is Jigar Shah. He's an energy entrepreneur and an investor.
And previously, he worked with the US Department of Energy under Biden. Jigar Shah, welcome to Money Talks.
Thanks for having me.
Jigar, to start out with, everybody must have seen the announcements from, big tech companies last week where they upped CapEx for this year, and I think four of the hyperscalers are going to spend almost $725,000,000,000 this year, with most of it in The US. Do you think there is enough electricity capacity to be able to service the kind of plans that hyperscalers have in terms of the AI infrastructure?
When you think about where we are from a CapEx standpoint, it's roughly $50,000,000,000 for one gigawatt of capacity. So if someone is spending $725,000,000,000, that's about 14 gigawatts worth of power. And that's not all in one year. Right? They're buying stuff now and they're gonna deploy it over two or three years.
Right? And so it's pretty obvious to anyone who's done the math that there's no chance that we go above 50 gigawatts of incremental new capacity in the data center space between now and 2030. That is a number that we can easily accommodate within our existing grid. But for whatever reason, people are amping up the hype cycle by saying it's gonna be a 100 gigawatts, 200 gigawatts, 300 gigawatts, when that is clearly not true and there's not enough money in the world even from these large hyperscalers to provide that much cash.
Yet, I think at the same time, we keep reading about the fact that in the actual data center construction itself, there are delays. So for example, the actual electrical equipment that are needed to build this data center stuff like transformers or switches, There seems to be a lot of delay in either procuring this equipment or in building more of them.
Yeah. It's the same thing we went through during COVID. We ran out of toilet paper, and we ran out of all sorts of things. And the question was, were people using more? And the answer is, no.
They weren't using more. They were just hoarding it. And so you see a lot of hoarding. Right? So if you look at companies who have no chance of moving forward, they have pre purchased a lot of this equipment for data centers that are never gonna get a tenant.
Right? Like, they're never gonna be able to close. They are increasing the cost of equipment for everyone else by 50. And there isn't a lack of supply. Right?
There's just a lack of allocation to people who are actually gonna move forward with their projects.
And what are the kind of lead times that we're seeing for some of these things?
Oh, yeah. For large transformers, I think we're up to three years, three and a half years. Right? But I think just to give you an example of where the capacity could come from, India alone has a 100 GVA of excess capacity of large transformers. That capacity alone could solve the entire supply demand mismatch within The US.
Some companies are talking about going, quote, behind the meter, meaning they want to build their own dedicated off grid power supply. What do you make of that?
So there are some people who are saying the grid itself is what's holding me back. You know, the interconnection process is taking too long, and so I would like to find an empty piece of land. I'll put in a data center. I'll connect it directly to a power plant, natural gas in the case of most of these data centers, and we will just not be connected to the grid at all. There's a number of problems with this.
One is is that it doesn't work. Right? So data centers are the most unforgiving loads in the country. Their load shape changes 10 times a minute. So then you have to have a battery buffer behind there to deal with that.
But, like, if for whatever reason, the programming is not done correctly and the data center stops computing, the stress that it puts on the natural gas power plants is so high that you're seeing premature failure of the power plant. Now when you put it onto the grid, right, well, now several things occur. One is is that that natural gas power plant that they've installed is now a resource of the entire grid. So if something goes down on one part of the grid, the grid can ask the hyperscaler, hey, can you spin up your natural gas engines because that will make it easier for us to run the grid. So when the hyperscaler says, I'm gonna go behind the meter, they're saying we don't wanna be part of a collective good.
We don't want to be backup for other consumers. We don't wanna be helpful to our neighbors. We want to just be left alone to do our own thing. Right? Which is the same reason why data center companies wanna go to space.
They're like, just leave us alone. We wanna go to space. Right? But the second thing is that when you connect to the grid, a lot of these transients that would actually, like, destroy a natural gas behind the meter plant could be absorbed by the grid. So there are lots of reasons why you'd want to be plugged into the grid.
And, again, if you're only at 50 gigawatts by twenty thirty, you can easily accommodate it within the existing grid.
And so we've talked about hyperscalers, particularly building out these data centers. In your view, what has been the position of the actual utility companies to this kind of build out? Are they in favor of it? Are they cautious about it? What has been their position in this?
Well, it's been a fascinating set of changes in their mood. Right? I mean, I was serving in government in 2023. Right? So when ChatGPT came out at the 2022, everybody was caught flat footed.
At the time, the utilities were basically like, we hate data centers. We don't want all of this load growth. We don't want all this issue. Right? Because they're old and stodgy.
Then when the Trump administration came in, the Trump administration basically said, you are gonna love data centers and we are gonna force you to love data centers. So by 2025, the utilities were starting to say, hey. This load growth thing is something we should plan for. Weirdly, now it's becoming apparent to everybody that 50 gigawatts is the number, not 300 Gigawatts. So now there's actually not enough load growth for all of the empty promises that the utilities have made to their shareholders.
And so now they're fighting for load growth. Right? So now they want the data centers to come in because they've realized that their load queue that they're seeing all of the empty people who are just asking for load without any financing in place or any hyperscalers signed up. Those are all phantom. And so they were inundated and now they're realizing that most of that is phantom and they actually need some of it to meet some of their load growth requirements.
So a lot of companies have been putting in requests for data centers that are speculative in the sense that many of these data centers just won't get built. But hyperscalers, which are these large software companies, are in a position to actually build out these data centers and are doing so, but they are running into another problem, which is opposition to data centers at the local level. Around 180 different groups have formed all around The US opposing data centers in that locality. Between them, they managed to stall or block around $150,000,000,000 of data center expansion last year. Now these are people who have concerns around data centers being built in their backyard because of land use, because of environmental concerns, or simply because they are linking data centers in their backyard to higher electricity bills.
How big of a political issue is this becoming in The US?
I mean, there's 36 governor races in The United States today, and I'd say every single governor's race for both the Republicans and the Democrats have been, what is your position on data centers and what is your position on their impact on electricity rates? Such that many Republicans will vote for a Democrat if the Republican is too business friendly and not saying enough tough things about the data center. And so I'm hoping that everyone comes to the table and says, hey, instead of being lazy on all these topics, let's be a lot smarter about how this can benefit consumers.
So this sounds like a coordination problem between utilities, between the private companies, and the government itself. Right? And it's not just one government, it's state governments, it's at the federal level. What in your view should policymakers actually be thinking about or how should they approach this?
Yeah. So this should be run by the federal government. But unfortunately, we don't have a federal government that actually wants to coordinate these things today. And so we're in a situation where that's been left to The States. I think there's some best practices coming out, and one is that every utility has to put data centers on an interruptible tariff.
That means that, yes, we will connect you in places where there's enough capacity for you. But if something goes wrong, right, a generator goes down unexpectedly, a line goes down, whatever, then we're gonna shut you down for a hundred to two hundred hours a year. That's point number one. Now the question is, once you put a data center on an interruptible tariff, well, what are the mechanisms by which they can run without interrupting? One is obviously behind the meter gas, which we talked about.
The second would be Emerald AI and some of those kinds of companies where they can shift some of the compute from that data center to other data centers. That is not really proven. It's really in the pilot phase, but folks are talking about it. The third is that the data centers can invest into the community. Right?
Remember, there's a lot of anti data center sentiment in the community. And so what you're finding is that data centers can actually give people free batteries or they can pay for half of the cost of the batteries that people wanna put into their homes anyway. And then when the data center is asked to turn down their data center from, let's say, a 100 megawatts to 50 megawatts, they can instead say, well, why don't we have these residential customers and commercial customers run off of their batteries instead of running off of the grid, and that frees up the 50 megawatts of capacity so that the data center can run at the full 100 megawatts. And so you're starting to see the data centers understand that there's a path forward where it's a win win for everybody, and I think you're starting to see these 36 governors races forcing these best practices in every one of those 36 states.
So far, we've been talking a lot about The US market. In your view, how does this change when you look at the picture globally? I know Middle East is one big area where there are data centers being built out. But overall, do you see these issues very US centric or the other common patterns if you take a global lens as well?
Well, I think The US is much maligned on its electric utility progress. Right? I think there's a lot of people who believe that China is way ahead of The US. But on this issue, which is like behind the meter gas, virtual power plants, batteries, The US is 10 times ahead of China. China just announced this week, for instance, that it is mandating that all the utilities use all of this, like, next generation technology that The US has already integrated to run distributed solar.
So you're starting to see a lot of people understand that this technology wave around how to run the utility better, how to get better grid utilization out of the stuff we've already paid for is something that's required. Right? So you're starting to see that in China, but you're also starting to see that in Europe where they think they have a lot of the technology, but the balkanization of the 27 member countries has left them where they don't really orchestrate across the 27 member countries very well. They run each grid individually. Now moving to The Middle East, I'd say today, people are very reticent to invest heavy dollars into The Middle East.
And even The Middle East is very reticent to spend money that it doesn't have. Right? I mean, it's not selling as much oil every month as it used to. So you're starting to see a lot of pullback around how committed they really are to building out all the data center capacity. And so the other big thing that The US is innovating on is on the edge compute, and so you don't need these 1,000 megawatt data centers.
You can actually live with 50 megawatt data centers or five megawatt data centers, is fantastic for the African Continent, Southeast Asia, other places where inference is gonna be very important for the AI race. Right? Is when you use your AI, you don't want that signal going all the way to The United States to get processed and back to the country. You want that compute to happen in country and you've even seen Jensen Huang talk about how inference is going to the edge.
Jigarh, thank you so much for joining me.
Of course. Thanks for having me.
Thank you, Shailesh, for talking to Jigar for us. I enjoyed how he poured a dose of cold water on lots of the fears around data center capacity in particular. What were some of your big takeaways from your conversation with him?
I think there were two. The first is, as you mentioned, just the disconnect between the electricity demand that was forecast based on the announced projects and actually how much you need. It seems a lot more manageable. That is not to say that energy is not an issue, but at least it's not an immediate issue. So that was number one.
Number two is looking at the hyperscalers and the AI supply chain, it's just striking the disconnect in terms of how fast software is evolving and the time it takes for the hardware supply chain to catch up. I think that is a fundamental disconnect that I'm surprised that the software industry is surprised about this because it is always known. But the fact that two to three year timelines are normal to build out a new fab, to build out additional capacity, all these things, whereas software is improving in a matter of months, demand is improving in a matter of months. And I'm just surprised at how much of a surprise this is for the model makers.
I enjoy that sense of childlike wonder of people at the very bleeding edge of the tech industry coming into much more conservative, often much more slow moving industries and being like, wow, does it really work like this? You know, if you think TSMC is conservative, wait until you meet a power utility company that really isn't used to moving fast or breaking things. Breaking things in power utility generally considered to be a a bad thing. I wanted to ask you one sort of cynical question, Shailesh, which is occasionally I speak to people about AI build out issues and they raise the issue of circularity. So they'll say, listen, the cloud computing companies are investing in AI labs, and the AI labs then make a commitment to buy enormous amounts from the cloud computing companies.
And they say this is all vaporware. The demand is fake, basically. It's a circular relationship. How do you address that sort of argument?
I would say two things. First is I don't think we should conflate announced projects with actual projects on the ground. I think at any point in time, there is always a lot more announcements than actual investment on the ground, so that always exists. The second part is a lot of the build out is being driven by the hyperscalers. So the number we looked at around 750,000,000,000 this year out of a total of maybe a trillion dollars potentially this year, most of it is being built out by hyperscalers.
And as you've seen from their earnings report, their demand is real, and they are also claiming that they're seeing revenue gains from AI. So I do believe on the edges, there will be a lot of these deals that will actually end pretty badly. But for most cases, I think the fact that the demand is being driven by these large companies, that should be some comfort.
Yeah. It's interesting. I definitely was probably on the side of the ledger that thought people weren't necessarily finding enough ways to use AI to potentially sort of match all of the fantastically large investments that people were making in capacity. I do think my sense of that has has shifted a little. I mean, in particular, everyone who works in software is just using this all the time.
I have actually spent a couple days this week at the SelectUSA Investment Summit Conference, which is a sort of big shindig hosted by the Department of Commerce to try and attract foreign direct investments into The US. Basically, it's just like a pony show between all the states who are trying to attract businesses to come and set up there. And you know, the businesses that everyone wants to build are data centers. People are really not interested in having them. And it's really obvious why.
Data centers provide basically no jobs. You know, you maybe need like a couple dozen security people. They are hugely resource intensive and they're ugly and loud and, yeah, nobody wants them. But it does seem like the real constraint, Shailesh, correct me if you think this is wrong, is the chip side of things, not the data center side.
It is. And and as you were talking, I just realized there's there's one additional data point to point at the shift in, quote, unquote, vibe, right, which is the which is the term we're all using now. A few months ago or maybe six months ago, there's a lot of talk about depreciating assets on chips, and there was a concern that a lot of chips that these companies buy will be obsolete in two years. Well, today, the situation is such that the h 100, which is a very old chip made by NVIDIA, and by very old, I mean, it it was launched in October 2022, so very old in tech terms. The rental price of that chip in the last three months has gone up by 30%.
Right? So even though you have had two successive models that have come out, which are potentially a lot more powerful, more efficient, blah blah blah, the fact that you can't get access to enough of those means people are going back to the so called older models. And that's actually another indicator of just how much of a supply crunch there is on these processes.
Yeah. I think, like Alice, I've definitely sort of long since passed the point of thinking, oh, maybe this won't work. Maybe maybe this sort of demand isn't there. I mean, I I do find the discussions about monitoring employees' token use as almost like a productivity measure I find to be absolutely bonkers in the sense that it's like measuring a a truck driver's productivity by how much oil they've consumed. Right?
It's like, this is an input. You shouldn't think about it this way. But it's fair enough. The local politics stuff fascinates me. It fascinates me because we often think about the winners of AI, the countries or places that benefit most to be a technological question, or at least a top level government question.
You know, are these industries incentivized? Are they working with government well? That sort of thing. And a lot of it is actually gonna come down to whose local politics works best. And the opposition to these things in The US now is absolutely bonkers.
I tweeted something sort of as a joke about how can we make data centers look more beautiful and marked one up as an AI image of like a a castle. It's a joke, but I think there's probably something to they don't have to be the ugliest buildings in the world. But the number of replies you get, where it's like, I don't want my electricity bills to go up. It's about the local water. It's about the land use, blah blah blah.
Alice, you mentioned the lack of jobs they create. You can't win on this because if they created loads of jobs, people would complain about the traffic, and they would complain about the pressure on local facilities and the area doesn't have enough doctors for all these new employees and blah blah blah. There's always something to whine about. So, yeah, I am interested as to how this plays out. Can people come up with some sort of deal?
There has to be some sort of win win here of better incentivizing data center development so that local areas will be encouraged to sort of invite them in. It doesn't seem like anyone's quite there yet, but you've got to get there eventually.
Yeah. I mean, I take your point that there's always something to whine about, but my impression is local people do in general like it when there are jobs in their community. That seems to be the vibe that I'm getting from all of the politicians I've been meeting this week.
Because this
Local people just don't worry about anything. I I no. I take the point. I take the point. It's probably, all things being equal, better if if there are jobs there.
There's a really interesting piece of research by judge Glock at Manhattan Institute about Loudoun County in Virginia, which is like data center heaven, basically. It's like the most built out place for data centers in Northern Virginia. And he notes that basically half of the property tax in Loudoun County for the next couple of years, the projection is that it's gonna be paid for by data centers, which allows them to keep property taxes lower on people's homes and other businesses than they otherwise would. It seems like there is a version where you can get this right, But at the moment, people do not seem to like it very much.
Yeah. I do think that tax is gonna be the great solver of this equation. I guess it's interesting that's working in in Virginia. We should pivot to our stats of the week. Shailesh, as our esteemed guest, why don't you go first?
What have you got for us?
A big number, 1,100,000,000,000. This is the estimate by Morgan Stanley for how much the hyperscalers that we've been talking about are going to spend next year on capital expenditure.
It's a lot of money.
I think once you get past 1,000,000,000, then I think everything is is the same.
I just I guess, when we talk about trillion dollar sums, usually, we're talking about, frankly, like invented money. We're either talking about something that the US government does where, you know, the money amounts get very large, or we're talking about market capitalization, which is a bit of a sort of flim flam figure extrapolation from, you know, the last share of something sold, not a real realizable sum that anyone is actually exchanging.
Speaking of big numbers, my set of the week this week is a $114,000,000,000. That is the hypothetical value of the venture capital investments made by Sam Bankman Fried through FTX If he still held those positions, they were liquidated in the bankruptcy of the firm. And the major contributor to this is the massive stake that he took in Anthropic. He invested $500,000,000 in Anthropic in 2022. Obviously, that's worth a lot more money now.
It's worth about $82,000,000,000 or a 165 times his original investment. Very unfortunate that he chose to make these investments using commingled funds from his exchange that he shouldn't have been using to do it, but they were pretty good investments.
I've said it before. I've said it again. If I was that good at investing, the one thing I wouldn't do is conspire to commit fraud. I would just do the investment and make the money and be really rich, and I wouldn't do any fraud at all.
Use you could use from Mike.
Yeah. Yeah. My stat of the week concerns billionaires as well. It is 41%. This is according to research by the California Tax Foundation.
The proportion of California's billionaire wealth that is already pledged to leave the state, that's just nine tech billionaires. If the California billionaire tax act is approved by California voters, this is a one time, supposedly one time, wealth tax of 5% on billionaire net worth. Lots of tech billionaires are already saying they'll leave, including Mark Zuckerberg. I think it's fair to say the California Tax Foundation is against this tax to sort of reveal the bias here. But yeah, that's quite a lot of billionaire wealth heading for other states.
Maybe they'll come back if it doesn't pass. I don't know.
Or I guess you do. What good is it to those other states if you can't tax it? You know? Well well, you know,
they spend money and employ people and stuff like that. I don't wanna sound like too much of a bootlicker, but there are some benefits to having billionaires live in your state even if you're not taxing their will.
Yeah. I mean, I think we have better and worse ways to try and go after the money of rich people if that's what you're trying to do.
Mhmm.
And it's possibly not the most cleverly designed text anyone has ever thought of. Well, with that, I think we are just about out of time. So all there is left to do is to give a big thank you to Jigar Shah for joining us.
And thank you again to Shailesh for joining us. Always a pleasure.
Thank you for having me.
And thank you for listening to Money Talks. If you know someone who might enjoy this podcast who is not already an Economist subscriber, there is a way to share it with them. Find the episode on The Economist app, click the share button, and select give as a gift.
You can always write to us at podcasts@economist.com.
Today's show was produced by Tom Wolfenden and Lauren Snight.
Our sound engineer is Mark Burrows.
Dan Asher is our editor, and the producer is Hannah Marino.
I'm Alice Hallwood.
I'm Mike Bird.
And this is The Economist.
No claim selected
Click an underlined phrase in the transcript to see the fact-check here.