No. of Recommendations: 8
abromber wrote:
To give some perspective, the railroad boom of the 1870s–1880s was massive by any measure. Spending peaked in the early 1870s at nearly 5% of GDP, and railroads consumed 15–20% of total national investment through the 1870s and 1880s.
Today's AI buildout is accelerating but hasn't reached those levels. Krugman published an interview today with Azeem Azhar, a top tech blogger on Substack, in which they discuss the utility and economics of AI. Krugman playing a bit of the skeptic and Azeem more the believer.
https://paulkrugman.substack.com/p/talking-with-az...They have a lot of interesting things to say, including comparisons to past bubbles and tech revolutions like the railroad.
"
Krugman: Let’s move to more macro considerations. People have been worrying about a bubble. A lot of us still remember the nineties quite vividly and think about all of that. But you just aren’t seeing the bubble. You wanna talk about that?
Azhar: I remember what it was like in the nineties. I lived through that one and also the housing bubble, which frankly was far, far worse and much more terrifying. I have a really simple mantra here, which is that honest customer revenues tend to be the engine that gets you through this, right? You know, what caused the problems with the US railroads in the 1870s and 1880s? It was that the revenues didn’t materialize because the tracks were being laid in places where there were no towns. That was a problem. The same was true in the dot-com era. My team and I realized last year that it’s very hard to get good quality data on how much was actually being spent by American businesses and consumers on AI. So we’ve spent several months building systems and gathering data to give a deduplicated view of what that number is. And just to give you a sneak preview, is $150 billion per annum, annualized at the end of May 2026, and about 90 billion dollars in the previous 12 months, from May ‘25 to May ‘26...
It’s a much faster revenue growth rate than mobile or the internet. It’s also a small number because the US is a $32 trillion economy. And I think the thing is that at that level of spend, you are able to roughly cover the depreciation on the enormous capital expenditures that have gone into AI just this past year. But next year or the year after, you have to double your revenues again and again in order to cover these increasing commitments.
The thing that often pricks a bubble is when financing starts to get a bit smelly...So the other thing that we look at is how bad, poor, or strong or robust is the funding quality. And that funding quality measure is definitely getting worse. It’s worse now than it was nine months ago. But it doesn’t seem from the numbers to be at the level that it has been historically when these things have imploded. Nor does it seem to be the type of exposure that is really systemic, which is what we saw in the global financial crisis.
...
Krugman: ...People say, “I see the technology everywhere but in the productivity statistics.” Do you want to talk about that?
Azhar: It comes up all the time. I wonder if we need things to happen more quickly than we used to. We aren’t seeing it in the numbers yet. Erik Brynjolfsson at Stanford says he thinks it is showing up in the aggregate numbers. How quickly should we expect a technology like this to show up? At $90 billion a year, that’s not much of US GDP. These are early stages where companies are learning. The first $100 million you might spend on AI is about learning, and we’re in that mistake-making phase.
The model Paul David and William Devine talked about in electricity is helpful. In the first phases, you’re retrofitting your capital stock and processes with the new technology. It’s not until you depreciate existing capital and change processes—like Ford did at Highland Park—that you see productivity benefits. To put numbers to that, what would we expect to see in the ... equivalent of Highland Park in terms of output?
...
Chinese companies are using much less capital to build models that are nearly as good. So I think the harder part of your question is that if every model that OpenAI or Anthropic costs ten times as much to deploy and develop, but lasts only a couple of years before it’s defunct because of competition, what needs to be true for that to be sustainable for more than a year or two? To me, that is a really tricky question as well.
"
This content is free for subscribers of Krugman's Substack, so I think it's okay to copy this much here, with the link provided. He publishes prolifically for his free subscribers and it's mostly good stuff. Highly recommend signing up.
