Azeem Azhar on the Explosion of Hacking and Other AI Issues

. . .

TRANSCRIPT:
Paul Krugman in Conversation with Azeem Azhar

(recorded 10/29/26)

Paul Krugman: Hi everyone. I am talking today with Azeem Azhar, writer, researcher, investor with Exponential View, who is one of the people I turn to when I don’t understand what’s happening with AI, which is basically all the time. There was a really striking piece that Azeem wrote the other day about the PHP hack with the title, “The Wuhan Moment.”

Why don’t you explain to me what you mean by that, and let’s try to understand what’s going on here.

Azhar: Yeah. Well, that PHP hack was just one particular instance of a range of things that are now happening because AI models have become good enough. And the reason I called it the Wuhan moment is that there are infectious diseases all the time, but then came along one with slightly different characteristics, and it had a much bigger impact than all the other ones that were circulating.

So, when I say AI models have gotten more capable—we’ve had this experience as we use ChatGPT, it gets smarter and smarter—and AI models became good enough to be used in code, to be directed to find vulnerabilities and weaknesses and bugs in software. And because AI systems are tireless and indefatigable, and you can just spin more of them up and they don’t get exhausted, they’re not like human security researchers or human hackers. They can find all of these vulnerabilities, many vulnerabilities, quickly. They can research them in novel, non-human ways and chain them all together. And that’s a new regime that we find ourselves in. And it feels different to where we were a year or two ago when these weaknesses and vulnerabilities were discovered methodically by the occasional hacker, but generally by security teams. So that’s the moment.

Krugman: So my understanding is that basically the way the world works now is that there’s software everywhere. Everything is being run on software which is complicated and maybe not that carefully put together. Tell me if I’m wrong on this, but I think it’s sort of on stilts where you have newer software laid on top of older software, and God knows where the vulnerabilities are.

Azhar: Well, software and technology is this really big industry that somehow managed for 50 or 60 years to escape any part of safety testing.

Krugman: Right.

Azhar: And so, you know, that’s why you have beta releases and bug fixes, because the culture of software development for the longest period of time has been started from a tinkering mindset. And there’s no real security model required around software. It’s not an AI problem, it’s not a social media or an internet problem.

It was a problem in the ‘70s and the ‘80s. It’s just that we started to move from centralized mainframes to having personal computers and many, many more opportunities for that software to be developed in a decentralized way. And so here we are decades later, and it’s a miracle the internet works, given how many bugs and issues there are. But that has also been its benefit, because what it has by not being a fixed cathedral is agility. One software developer likened it to a crowded, buzzing bazaar. It has the ability to respond, adapt, and stay pretty resilient, which the internet has by and large done, except for a few major cyber attacks over the last 25 years. So what was a feature, and a flaw recognized by software engineers, that there’s no agreed security model for making software—has also been the reason why we’ve had a flourishing of software, and we are able to continue on our way even when bugs are found and there are problems that get discovered.

Krugman: Right. And a lot of it is based on PHP. You should probably tell people what it is. I mean, I actually had never heard of it until last week, but of course...

Azhar: Well, you’ve certainly used it. In reality, 70% of websites are running on PHP. It’s a programming language that happens to be very easy for people to use, and so it was cheap. If in the late ‘90s you wanted to write a piece of software, you used PHP. It’s simplistic; it doesn’t have the learning curve of C++ or something else. And it got extended and extended. Facebook used it a lot, and so it became part of the internet’s architecture. And like many other systems we use, like Python, which is another language, or BIND, which is a piece of software that turns friendly internet addresses like Substack.com to the numbers that the routers need. These are pieces of software that are actually open source as well, you know, they’re written by collaborations of volunteers who have their own governance systems, who share that code in repositories. And again, if you don’t think about it, you’ll be surprised. And it feels like it’s a miracle that it works, but we have evidence and proof that it does.

Krugman: Okay, and how relevant is it that it’s old? That PHP is a ‘90s thing. It’s all jerry-built, which was good enough until now we have these AI models that are tireless, right? And basically an army of gimlet-eyed experts in deception can go and find the holes, right?

Azhar: Right, yeah. I mean, the whole setup was maybe not quite on stilts because it all worked. It also has to be recognized that it’s really hard to write extremely precise, foolproof software in an enterprise—be it a corporate, consumer, customer environment—because standards change. If you want to understand why, think about the experience when you’re flying on an old plane, as you will across the Atlantic, and you pull up the in-flight entertainment, and the software is so slow compared to your modern smartphone because it had to be designed and locked down 18 years ago, or 10 years before the airline bought it, and it’s still in place. So all of these things were really part of the trade-off that we as customers and consumers made to have something that would adapt and be responsive to our lives. None of us really want to use the type of software that runs a nuclear power station, right? Because it’s stuck in the ‘70s or ‘80s. And that means that there are just lots of different ways of getting these pieces of code to crash, and once they crash, maybe to get access to the machine and figure out where to go next. And that’s what great hackers do.

AI models crossed a threshold at the start of this year when Anthropic had a model called Claude Mythos. It’s the best cybersecurity model in the world. It was not made widely available. In fact, only a small number of organizations get access to it. And they’ve been running around trying to patch or fix these bugs and close loopholes. The first part is finding them. But one of the things that happens with AI, is that the capabilities cascade down to cheaper models very, very quickly.

And so at some point this summer, the open-weight models, also known as open-source models, started to get towards the capabilities that Mythos had. And so they, too, crossed that uncanny valley into a point where they can be chained together and directed by either hackers or bad actors to find vulnerabilities before the good guys find them and then try to exploit them. And we’ve seen an uptick, a rapid increase in the number of those vulnerabilities that are being discovered. And this PHP one was one example of it.

And I guess what I was writing in my essay was to say, Look, what we know about these technologies is that they get cheaper, and as they get cheaper, they get used far, far more, right? We all take many more digital photos than we do film photos now from 40 years ago. The same thing would likely happen here. And that means if there are vulnerabilities or exploits, however obscure, it’s quite possible that they would be found and exploited. And I’ll just mention one data point that I’ve seen. There was one particular piece of software that had been tested for vulnerabilities trillions of times by humans running test suites. And yet a recent AI model was able, despite that, to find severe vulnerabilities that could be exploited. So even though we’ve looked very hard, we can never guarantee that these things are 100% robust.

Krugman: And this sounds like it’s a constant arms race. Now you’ll be using AI to try and patch or protect, but it sounds like inherently the advantage is going to be with the bad guys, right? It’s going to always be hard, or am I wrong?

Azhar: Well, I think the bad guys move more quickly than everyone else in general. Remember, the bad guys only have to find one vulnerability at one place. The good guys have to defend every vulnerability all the time. And so I think it would be fair to say that there are lots of private sector organizations and government organizations that are ahead of the curve, but there are also billions of computers connected to the internet and trillions and trillions of lines of code and systems that haven’t been updated for years that could be vulnerabilities. And that’s why the Deputy Chief Information Security Officer at Anthropic said it’s going to be quite a rough 18 months.

And I think what I alluded to in my essay was that there is that period where, you know, just as it was with the spread of what we then called, believe it or not, 2019-nCoV before we called it SARS-CoV-2, things can spread faster than we can figure out what we can do about it. And I think that that’s the risk that emerges in that shorter period of time.

Krugman: And what we did do to deal with the spread of the coronavirus, with extreme political consequences, was lockdowns. And you suggested that we were going to be seeing a lot of partial disconnect from the open web. So, how does a walled garden actually operate, and can we even do that now?

Azhar: Well, arguably, Paul, I think we may already; many of us spend a lot of time in walled gardens. If you live in Instagram or you live in Facebook, it’s not the open web that you and I remember. It’s not the open web I fell in love with in 1992, where you go from one place to the next. And you know, even so many things require password authentication to access. But I do think that what you’d expect to see is that the number of AI agents will rapidly exceed the number of humans. I think the amount of .browsing that is done by bots on the internet now is higher than humans. It’s always been high because search engines have to crawl, and people who are buying things on eBay will use price-sniping bots to keep an eye on the listings. But that will tilt enormously towards the machines over the coming few years. The internet and the companies on the internet will evolve to meet that demand.

The way that I liken it is that if you think about Fifth Avenue in 1895, there were more people strolling in the middle of Fifth Avenue than there were automobiles. But by 1920, people didn’t stroll down the middle of the street. You don’t go for an afternoon walk on the freeway. You might go for an afternoon drive and use the freeway to get you there. And so I think that the internet would need to start to filter out what these AI systems need. They move really quickly. They’re trigger-happy. They work in milliseconds. We work in seconds. They have completely different technical requirements. And that will mean that that would occupy a different patch of cyberspace to the patch that we want to be in.

And in particular, if we want to be really, really secure and robust about threats, we might choose to say, “I want to stay in an enclave that is safe.” And I think a good example of that is what Apple does with the Apple Store. In some countries you can use third-party stores—but mostly you have to get apps from the Apple Store, because Apple wants to say, “We’ve checked that this is safe.” Whereas in Android, you can go anywhere. And that’s why there are many more viruses and safety problems on Android. So I do see this potential journey where the internet needs to serve the kind of economic requirements of companies using agents, which means how fast it is, whether there’s transactions and pay-as-you-go and validation and so on. But that may not be the way that we will want to operate on the internet at all. It will feel it’s not for us in the same way as the 101 highway in California is not for humans to have a stroll down.

Krugman: So it’s going to be a world in which, instead of being able to let a bot go and make purchases, you will have to prove that, or you know, offer some evidence that you are you, or only certain things will be available; just a much pokier environment, is that it?

Azhar: I think you’ll end up having to validate your bot, and you’ll have to give your bot your approval, and the bot will have a set of rules. It’s allowed to spend $50 without sending you the approval link or whatever the system happens to be. And the truth is, I don’t think we know at all right now what that would look like. I mean, people are experimenting, and my sense would be that publishers in particular, and then retailers as well, will start to publish listings for consumption by machine agents in the way that they need to see the material, which is much more structured, much less emotive, just with more data that they can navigate around. And they’ll do that because you or I are quite likely to be using an agent on our iPhone or an agent from Meta to interact with wherever we know we’re buying or browsing. I mean, I really don’t think we’re there yet, but I think people in the industry generally think that that’s the direction that we would go.

And I have a little bit of experience myself because I have a bot that keeps track and looks out for concerts I might want to go to, because I don’t like going to venue websites because they’re all different from one another. So the bot knows my music tastes, it’s got my music catalogue, and it checks every week or so to see if there’s a ticket I might want to buy and tells me about it. So, I’m no longer visiting those venue websites, but I am going to be paying for the ticket.

Krugman: I remember how freaked out I was when I was a very early Amazon book buyer, meaning physical book buyer, and when Amazon started recommending music to me. I never actually did buy any music from them, but the recommendations were pretty good even then. And that was kinda shocking to realize that at some level, from the books I bought, they figured out what music I probably liked. And it’s just a very spooky world for someone my age.

Azhar: I think it is spooky; it’s spooky for all of us when you get those suggestions. I mean, you can probably do it in ChatGPT. If you took a photo of my bookcase and stuck it in ChatGPT and said to it, you know, “Recommend some books that this guy might like,” I think it’ll do quite a good job. I remind myself that it’s math, that’s what it is. It’s kind of traversing an enormous matrix and it’s finding some likelihoods.

Krugman: No one listening to this will know what I’m talking about, but it just shocks me that basically matrix multiplication with a few nonlinearities can do all of this. It’s just absolutely bizarre that it works.

Azhar: It’s so weird. It is so weird.

Krugman: Okay. But this is a little less apocalyptic than what I took from reading your piece. But I guess it still may be very, very scary. You think that we will eventually have worked out a reasonably seamless way of making it work?

Azhar: Look, what is an apocalypse? If we look at what happened with COVID, could COVID have been much, much worse? Yes. Was COVID really bad? Yes. Was it apocalyptic and awful for lots and lots of people individually and also for entire cohorts? Yes. I mean, imagine the kids who did their college in lockdown on their own away from home for the first time. I mean, that is just awful. So I think that those are bad outcomes. I think the thing that we’re dealing with here is that it feels like over 18 to 20 months people will have figured this out, and those guys are not going to get much sleep, but in the meantime really uncomfortable things could happen. You know, you could get a bank that just fails to get ahead of the ball, and your money in a bank is only an entry in a database. It’s not in Jimmy Stewart’s hands, right? And the cash is literally just a line in an Oracle financials database. And that would be pretty apocalyptic for people.

And I think that we can’t rule out what statisticians call tail risks, but we can’t see those tail risks, right? They’re Donald Rumsfeld’s “unknown unknowns.” And what we do know is that the internet with all the billions of machines connected to it is now much more porous and much more vulnerable than it was two or three months ago. And so does that mean that something quite bad could happen? Lots of things like power grids are sometimes connected or accessible to the internet. Well, it’s more likely now than it was a year ago because of this. And that’s where we have to be very, very proactive.

I’ve spoken to policymakers and a few months ago, I went off and I said, look, you need to come together and have a public-private agreement partnership where the state is actively encouraging small, medium, and large companies that are not tech companies to ask the right questions, to make sure these things are being proactively fixed because there will be a moment when you cross this uncanny valley, and the vulnerabilities will magnify hugely. So, I mean, it has to be taken really, really seriously.

If I don’t sound like I’m running around panicking, it’s because I model important things on Sully Sullenberger, the pilot of that plane that landed on the Hudson. You know, he had 115 souls on that plane and he didn’t panic. And there’s a great media interview with Katie Couric where she keeps asking him how he was feeling and was he panicking? And he is cool as a cucumber because at moments like this, when things matter, we want to use our evidence, we want to use data, we want to use our judgment, we want to think calmly. So I think it’s serious. People should act quickly.

Krugman: Okay, but how many people in major leadership positions in the world remind you of Sully Sullenberger?

Azhar: Well, I mean, you know, democracy is the best system we’ve got. Unfortunately, you know, that’s what it is.

Krugman: I know. Or as Churchill said, “The worst system except for all the others.” But anyway, yeah.

Okay. Before I read about the hack apocalypse or the possibility thereof, the big question we’ve all been asking is about an AI bubble. For many of us, our kind of formative experience on this was the IT bubble of the ‘90s, which more and more I’m starting to think is actually not a very good model for where we are right now. But you’ve been, you know, not a wild-eyed enthusiast, but you’ve been basically saying, “Calm down. There’s a lot of evidence that this is at least not that bad of a bubble.” Do you want to talk about why you’re coming down on the calmer side here?

Azhar: Yeah, not that bad of a bubble yet, might be the better way to phrase it. One thing that I have noticed in the debate over the last couple of years is that it splits between descriptive discussions—the facts of the world and normative discussion—and feeling: the oughts of the world. And people sometimes find it really hard to separate those. And I think the debate gets extremely heated, very passionate, but probably not particularly helpful. And I would say that everything that I say really is at this point descriptive. I’m not saying it ought to end one way because the world would be better or another, right? So I think that’s quite important.

Bubbles and this overexpansion of investments during any technology are really, really common, as you know from your own work. The real pain emerges when customers don’t buy enough to pay for the financing of whatever it is required to build the factories, the infrastructure. And that mismatch is what creates bubbles. It’s what created the railroad bubbles in the US. It’s what in some sense was happening with dot-com and with telecoms. Not every investment wave in a new technology ends in a disastrous bubble. Most of them do, and mostly it happens because the financing gets too exotic, it gets too circular, it gets too big. But not every one does. Automobiles in the US, for example, didn’t end up in a bubble. They ended up in lots of companies failing and there being consolidation, but generational companies came out of it, as well.

And so when I look at what’s going on with AI, the question that I’m asking—and we look at the data—is, what is the weight, the ballast? How is the complexity of the debt and the financing that is building up set against the way customers, mostly companies, are spending? And does one outweigh the other? And I think that a lot of that can be measured. A lot of that can be driven by fact rather than opinion. Some of it is unknown, and that’s why I’m calm. I have a view which is that, to put it simply, when I was starting out at work, my first mortgage was for 85,000 pounds, so like $110,000, this is 30 years ago, and that was really tough for me to pay. My mortgage today is substantially larger than that, and it’s quite easy for me to pay. And if I took today’s mortgage and gave it to me at age 24, I would have been crushed under the weight of it.

So the scale of the debt, the finance obligations, have to be put in proportion to how customers are spending and how they may spend. And that’s why I’m a little bit calmer, more sanguine right now, although I do think that the bubble metric is higher today than it was a year ago, you know, for the reasons I’m sure we’ll explore.

Krugman: Yeah. There was first a debate about how productive is this technology. It’s obviously awesome. It’s still hard to believe that basically matrix arithmetic can sound so much like a person, but there’s still some argument. There’s still some people saying that, okay, this stuff is amazing, but we can’t see it in the aggregate productivity numbers. Have you been taking that on at all?

Azhar: It’s very difficult to see it in aggregate numbers. You’re a far greater economist than I am so I’m not sure when you would expect to see it in aggregate numbers. And I’ve heard one economist in China—I was in Asia a few weeks ago—saying they’re starting to see some numbers in some sectors. I still think it’s very fair to say that they’re not showing up in aggregate numbers. They are showing up in small numbers of companies outside of the tech industry. Depending on which survey you believe, it’s between 5 and 10% of large companies seeing maybe a 6% benefit, increased sales, lower costs.

Krugman: Yeah.

Azhar: And that could increase. What we don’t know is whether the other 90% will turn into that 10% in one year, two years, five years, 10 years, or ever at all. And they may not. And we don’t know what percentage of that 10% will get more and more benefit and therefore spend more and more. But I think it would be—with something like this, it’s a little bit like a doctor trying to understand a complex disease. You don’t just take the patient’s pulse. You take lots of tests and blood tests and you try to paint a picture.

So certainly, there’s not much in the aggregate statistics, certainly not in the US. There’s some sector evidence, there’s some evidence from surveys from companies. I talk to executives quite a lot. What execs tell me is that this is harder than they thought. More of them are getting success. Everyone intends to continue to spend next year to try more. When does that show up in a number that an economist like you can say, “Here’s a number.” I mean, I think that’s a really important question. If it takes too long, then it’s probably the case that revenues are not going to grow fast enough to pay for the mortgage.

Krugman: Okay, so to get back to revenues in a second. Where I would be looking at the IT boom of the ‘90s, you know, by 2005 we were talking about total factor productivity. In aggregate, we were something like 10% above where the previous trend would have expected you to be, which is a lot, although it kind of stalled out after that. But we are, I would think in terms of the timing, something like in 1999.

Azhar: Mm-hmm.

Krugman: The productivity benefits were a lot more obvious in the first four or five years into the internet bubble. So I don’t know what to make of that.

Azhar: Yeah, my recollection is that it was even earlier with IT. I thought that the real rapid speed increase was in that ‘94 to ‘98 period. It happened to coincide, I think, with Bill Clinton as president. So he was able to do all sorts of things with the budget and so on. So it was pretty quick and sustained. And the question is, where are we with this ChatGPT thing? I think that the clock should start probably a year after ChatGPT, which is when Microsoft and Google released enterprise, corporate versions of what they were doing. So that’s been three years. And the question is, how fast would we expect it to go, and is IT a fair comparison? And with IT, people had to buy computers, for example, which is a much higher cost than paying for a ChatGPT subscription.

So I think that there’s a reasonable case to say that this is not yet provably faster than IT and could well be slower, right? I think that would be a reasonable position to hold today. And I have some ideas as to why that might be. Like what is the cause of the inertia within companies, and does that mean [the return on investment is] not as quick as we might’ve hoped or thought?

Krugman: But then the second question is, suppose this does great things for aggregate productivity growth. It really is great for the economy. The question has to be how much of that is captured by the companies, right? In the end, the internet was great, but you can look at those old Qwest ads from the ‘90s about this wonderful new world of the broadband internet, and everything in the ads has come true, but Qwest is no more, right? And so, how much should we be worried that Anthropic and OpenAI and maybe some of the others are busy doing all these great outlays and producing great stuff, but that everybody will end up buying Chinese products that are 90% as effective and 90% cheaper?

Azhar: Well, I think that we have to separate out the companies like OpenAI and Anthropic with the companies who are actually paying for the capital expenditures in the data centers, which are the infrastructure companies. And those companies will make money whether Anthropic is selling to you or whether you’re using a Chinese model, because the model still has to run on a computer.

Krugman: Okay.

Azhar: And that computer is a GPU probably owned by Microsoft or Amazon. So they will still make money at the infrastructure layer. And so we try to separate out how much money flows to the infrastructure layer versus how much goes to the model lab companies like OpenAI and Anthropic. Even if people do move to those cheaper models, I suspect that there’s enough capital flow, and enough business, at least for the next couple of years.

However, what we constantly hear from the industry is that there is a huge undercapacity of computing. Just before we recorded this, Microsoft and Google had their earnings announcements and they said the same thing again: we just don’t have enough compute capacity to serve all of our customers. So I separate out OpenAI and Anthropic. And I think when I look at Anthropic, the company published a document that’s called an S-1, which is the marketing document before it goes public. And we forensically went through it and deconstructed it. And I was quite surprised by things that I saw. For example, margins are a really important part of whether a business can survive or not. And Anthropic’s gross margin from over the compute costs that it pays is about 85%. That number was about minus 200% two years ago when lots of people were saying this company was going to die within a year or two. (And these people have not corrected themselves, by the way.)

And so I think that’s a surprise that the margins are as strong as that, that they are making profits in that way. Of course, your point remains, though. Like, there could be deflation or price competition from cheaper models, but in the end, the infrastructure bill mostly lies with infrastructure companies, and they will still make money serving open-weight models.

Krugman: So the open-weight—I am embarrassingly weak on exactly what we’re talking about here, but that’s okay, I think. Although the open-weight models are, at least as I understand, they basically just do a lot less computation than the full sauce, but nonetheless the demand for compute is still skyrocketing.

Azhar: So the reason open-weight models are cheaper for us is that they’re effectively put out for anyone to download like open-source software. So you’re not paying a profit margin to OpenAI or to Anthropic. You’re paying just the margin that the infrastructure company is going to put on that model for serving it. You’ve disintermediated one layer of margin. As to whether they’re actually cheaper in tasks, this is a hotly contested issue where they might be cheaper by the token, but you need more tokens to get a task done, so they end up being more expensive.

And my view again is a little bit sanguine about this. I’m pretty certain that there’ll be a lot of workloads moved onto open-weight models. But I also think that Anthropic and OpenAI resemble software companies of the past: there were cheaper open-source alternatives to Microsoft Office, like LibreOffice, but nobody used it. Most companies just went to the branded business, and so did that. And I think that there is a simple mental model that says that even if they lose lots of market share, they can still persist as businesses because there will be the typical reasons that people buy software to buy their products. I think the big question is how much market share would they lose and what happens to their pricing power? And the truth is, I don’t think anyone can look out over more than a few months to have any certainty. That’s why I go back to all the data that I look at and see all the data that can contradict the arguments that I’m making, to just make sure that I’m calibrating myself.

Krugman: But that’s the thought, that ChatGPT and Claude will end up being the Microsoft Words of the future. Basically inferior products that everybody uses because nobody gets in trouble for buying it.

Azhar: Well, yeah, it might be that, but I do think that they are still better products. In my team, we use a lot of open-weight models for the work that gets done because it’s often good enough, fast enough, cheap enough. Weirdly, I particularly use an open-weight model for one of my agents because the model is not as sort of machine-intelligent as the state-of-the-art models. So it does what I ask it more often. If I swapped it out for an expensive state-of-the-art ChatGPT model, it would often do more than I asked, right? And I would say, “No, but that’s not what I wanted. I just wanted the simple thing done. I wanted you to get me eggs. I didn’t want you to bake me a whole cake.” And so there are all these things, Paul, that we just don’t understand right now. We have to feel our way, and on the back of it is like $5 trillion sitting on top of it.

Krugman: Yeah, that’s right. I mean, there are some places apparently trying to use AI for air traffic control, and it’s like you might get the response: “I’m sorry, you were right, there was a plane already sitting on the runway. Would you like some advice on how to deal with mass casualty events?”

Azhar: Right. Well, fortunately, we haven’t got there at the mass level, but of course there have been kind of horrible stories of people who’ve been sort of taken down by AI psychosis, and to the end of their lives, which of course is a horrible outcome. But I think there is a risk in this debate about a bubble or a boom, and it tends to present a binary. And I think there are other outcomes. There is an outcome where there’s not a bubble that bursts, but a period of very, very low returns for the companies who’ve made these investments. So it’s not great for their investors. But you don’t see that kind of cataclysmic arrest as we saw in the housing market in the US when housing starts completely stopped after the global financial crisis.

One of the reasons I think it’s important to think about other scenarios than financial catastrophe is that if American and European companies do start to get results and do pay enough as customers of these businesses so that they can afford to pay this metaphorical mortgage they’ve taken out, where does the load of that lie? The load of that is quite likely to lie on the workforce. And we could blind ourselves—by getting so excited about Bloomberg’s terminals turning red as share prices fall—to starting to take proactive steps on the possibility that the technology actually does work, does deliver greater productivity, that companies as a result do cut costs, and some of that falls on the workforce. And I think that’s too important to leave to tabloid shouting over financial calamity.

Krugman: Yeah, I am mainly worried about how capital intensive this stuff is. There is definitely a competition for capital and these data centers essentially employ no people. You take eight billion dollars that might have been creating a bunch of jobs, and you end up with a hundred maintenance workers. But let me just ask you, since one of the things I’ve been hearing, and I have no ability to assess, is that presumably there will be a lot of money-losing investments. You know, there will be investments that companies are sorry they made or the company ceases to exist, which can be okay. We had that in the IT boom of the ‘90s because the infrastructure wasn’t there. But then it was. So we had a lot of fiber that stayed dark for years, but then eventually all became useful again. But the story that some people are telling me, and I have no idea whether it’s true, is that these investments will depreciate really fast. That they will depreciate financially because the chips get outmoded, but even perhaps more so because they literally, physically deteriorate. Is that a concern?

Azhar: This is another example where we have data to help us.

Krugman: Okay.

Azhar: So these chips do, of course, go obsolete just like any other chip, and they tend to be carried by infrastructure companies on depreciation for between five and six years. The evidence comes through public, regulatory statements that people have made, so we can kind of assume that they’re true. And what they say is that even eight- or nine-year-old chips are still being rented at premium prices because demand is so high. Will a 2026 GPU, one of these computer chips for AI be in use in 2046 the way an 1888 rail line is in use today? No, it will not be. I’m sure of that. But at the moment they pay themselves back within two or three years.

Krugman: Okay.

Azhar: Now what happens if one of these smaller infrastructure companies fails? We talked about the smaller ones who are not the big household names, and of course they’re on shakier financial footings. So if one of them fails, some of their investors, their shareholders, will obviously lose money. The people who’ve lent them money may lose money, but the chips will not be lost. They will immediately be taken over. In fact, a lot of the contracts have what’s known as an interrupted-in-service clause. So even if the legal entity that owns them fails, the customer still gets serviced and served. And as long as we’re in an undersupply of compute, which we appear to be, those things will get picked up. And I think we could expect that out of these hundreds of rinky-dink small infrastructure companies, there will be a lot of consolidation over the next few years because markets don’t support that many individual players.

Krugman: Right. So the dismal analogy that people have made is with fracking, where one of the big problems with fracking was that it turned out that fracked wells declined much faster than was in the accounting. But you’re saying that that does not look to you like it is really an issue with the chips. I mean, that’s a little bit reassuring.

Azhar: Well, it’s not an issue now, and it’s what the data says. But the thing about semiconductors and infrastructure like this is that we have undersupply right now, but the moment you have oversupply, prices can collapse. And what happens with chip companies generally, and has done over 50 years, is that they go into periods where demand grows, and so there’s undersupply, prices rise, and so they go and build factories, and by the time the factory is built, demand has cooled off, and now they’ve got overcapacity and they have to cut their prices and they lose all their money. So semiconductor investors are really aware of that cycle, and at the moment the question that the model people have to build in their heads—and there are lots of variables—is for how long will demand for computing of AI and these chips exceed the supply that’s provided and the supply that is available? And if that switches because there’s a new algorithm that’s much more efficient that everybody uses, and demand doesn’t pick up, then you could get into an oversupply situation which would have many of the effects that we saw in shale.

But I think there is a difference with shale, because what would happen with shale would be that you’ve got the external price of oil or gas that is setting a recovery price, but even if the price is high you also have the depletion curve of the well where every subsequent barrel—or whatever you’re getting out—is more expensive. So you have that kind of economic tapering off. And I’m not familiar enough with knowing how quickly a fracked well would pay off. But certainly right now, the GPUs that are being bought are being paid off within a couple of years or two and a half years.

Now, there is a bigger issue, right? AI is not some kind of airy-fairy thing that just lives out in the cloud. The cloud is physical. It is steel and sweat and energy and massive buildings. And as you say, buildings that don’t employ anyone. We can talk about what we could do about that. And so there is a scramble to get all the things other than the chips that you need for data centers, like land, like cooling, and most importantly, power. My team and I think that there is an emerging power gap. That is, the amount of power the chips will need set against what power the US grid and the European grids can provide. That’s not that important this year, it’s probably manageable next year, but it starts to become quite significant the year after that. And that raises all sorts of questions because no one’s going to buy $100 billion of chips in 2028 just to not plug them in for six months. So I think that there are these additional risks that make this question quite hard to come down with a definitive answer.

Krugman: Yeah, that’s where I’ve been coming in is to say, even more than the data centers and the connections to the grid or the other power sources, there’s a lot of political veto points. And there’s often local government and angry citizens where that becomes a real issue.

Azhar: Well, let’s talk about the angry citizens. So I think this is a really important thing that’s going on in the US. It is an act of democratic engagement and political engagement. Quite often people who observe the US go, “Nobody votes, nobody cares, everyone’s tuned into TV.” And this is not what we’re seeing now. I think data centers are less popular than ICE, which is one of your less popular government departments. This is an important part, I think, of how during a radical technology change, [there needs to be a] social contract, an agreement that gets hammered out.

As you rightly say, for every $100 million of investment in a data center, they employ 25 times fewer people than an advanced factory. So they’re not local employers. And a deal that got brokered in the UK in order to give permission was essentially for every megawatt of power—and a big data center might draw 100 or 200 or 500 megawatts—the data center operator has to pay $100,000 into the local community, and the total sum was, I’m just looking at my notes, about 500 million pounds, about $758 million. Now, I think that idea of constructing a local license to operate, it’s eminently affordable for both the infrastructure companies and the AI companies that they serve because the margins are so good. They can absolutely afford to pay this. Brokering at a local level, if necessary, these types of agreements I think goes some way. And it’s not just about the money. We have to come back to the point that you made: data centers employ a lot of people during the construction, then they don’t employ many people thereafter. So what can local communities then do in terms of development funds and access? I know the UK better than the US. But funds like that can stimulate the kind of local entrepreneurship and business building that is sort of generative and sustainable for a community.

Krugman: Yeah, one of the things that puzzles me a little bit about the public backlash is why the data center builders aren’t saying, “We will come to Pleasantville County and we will hold open town meetings and we will contract to pay a large sum of money into the county treasury.” I’m not aware of any examples of that, yet it seems such an obvious thing to do at this point.

Azhar: There are a few examples. I think there’s certainly a set of precedents out in Loudoun County which is the hotspot in Virginia. And I know that Microsoft’s campus in Washington state made commitments about what would happen to water bills and electricity prices and put in money for local redevelopment in different ways. So I think the money is there. And that means that half the battle has been addressed already, because it’s not as if you’re asking someone who’s going to make a lot of money to contribute some of that to a local community. And I think that the question is, what do you actually do with that? Do you do as we did in the UK with North Sea Oil, hand it back as tax cuts and God knows what happens 50 years later? Or are you able to use it to inject some commercial economic dynamism that can survive once the construction workers leave?

Krugman: Okay, last question: what scares you? What is the real fear that you have? Obviously, we know that the hacking is something that is looking pretty alarming, but further down, leaving aside maybe extinction of the human race, what do you really worry about here?

Azhar: The first is the extension of the hacking to a state-organized, multi-layer gray war, from polluting the information environment to widespread, increasingly significant hacks on what we call national infrastructure. I’m thinking of electricity grids, financial systems, water systems, where it’s not sort of individual criminal groups, but they are being mediated and managed by a state. I think that’s really, really serious, because we don’t see bombs going off. We don’t think we’re in a period of conflict. Especially in Europe, where there’s much more sabotage going on as well, I think we are getting there, and people need to understand that that risk is real and it exists. So I think that’s one of the really big things that I’m concerned with.

But the second thing, as we talked about, is jobs. And I think that what happens to the workforce, I think will be a very gentle turning up of the temperature. It won’t be a hundred thousand jobs going at once. It will be a reduction in wages, fewer new openings. And then before you know it, the workforce looks different to what it did. And we have to be alert to that.

And finally, I’m super, super worried about what’s going to happen to our brains with brain rot and just relying on, you know, “it’s not this, it’s that,” ChatGPT-isms, and losing critical faculties. I bought all of my team fountain pens. We all write in fountain pen. My desk is covered in pages of handwritten notes. I encourage them to work without a computer a few hours a day. I read really old books, a book from 1794, things that are predating ChatGPT. I use the tools, the AI tools, all the time, but at the same time, I’m really deliberate about how I try to protect my mind and try to pass my thinking and process ideas to my kids who, being the age they are—well, actually two of them barely use AI, but they’re not going to be that receptive. But I think that we have to really think about what is it about human thinking that we need to continue to cultivate and allow it to flourish alongside these technologies that, whether there’s a bubble that bursts and delays it for a few years, whether it’s grinding, whether it’s smooth, it will be with us. And we need to talk about that and actually do things about that sooner rather than later.

Krugman: Okay, I think that’s a good place to end. Boy, scary stuff. Fascinating, but I did not see this one coming.

Azhar: Paul, that’s the nature of AI. None of us see any of it coming. I mean, that’s a sad part.

Krugman: Maybe it turns out that all the people we were saying “learn to code,” should have been saying “learn cursive script.”

Azhar: Might not have been a bad idea. Cheers.