Every tech company now seems to have their own AI: Google Gemini. OpenAI’s ChatGPT. MetaAI. Spending for AI is reaching record highs, powering a big boom for the stock market. Even the White House wants in on the fun.

So are we in an AI bubble — an overblown investment period that’s bound to deflate? Yes, argues Paul Kedrosky, a partner with SK Ventures and a fellow at MIT’s Initiative on the Digital Economy. But not the bubble everyone thinks we are in. “AI is obviously a hugely important technology,” Kedrosky told Today, Explained co-host Noel King. So what, then?

It’s the money going into the AI infrastructure like data centers that concerns Kedrosky: “We’re spending this prodigious amount of money on the underlying infrastructure for AI with probably no likelihood of recovering most of that cost, and a significant likelihood that most of those assets become worthless because of the speed at which they depreciate.”

What happens when the bubble pops? And can past bubbles tell us anything about what is to come?

Below is an excerpt of their conversation, edited for length and clarity. There’s much more in the full episode, so listen to Today, Explained wherever you get podcasts, including Apple Podcasts, Pandora, and Spotify.

How much money is going into these data centers?

It’s going to be on the order of trillions now. Forecasts are in excess of $2 trillion in data center spending ahead. But an increasing fraction of the money that’s being spent on all of these things that allow us to distribute AI, like electricity, is coming from debt. And debt comes with obligations. You don’t get to just walk away from it. So that makes this moment even more perilous.

If AI is so important, why does it not make sense for trillions of dollars to be rushing in? Isn’t this what we should be doing?

We should be. But the problem, of course, is that there’s this idea of what’s called a rational bubble. Everybody thinks they’re doing the right thing, but when you add everybody’s “right thing” together, you end up with a prodigious amount of waste.

It’s no different than if you go back to the 19th-century railroad bubbles in both the UK and the US. There was simply too much track, too many enthusiastic railroad builders building almost adjacent tracks to the same locations. And this led to an incredible amount of waste. But it also led to company failures and various market crises across the 19th century in the US and repeatedly in the UK. It’s not as simple as saying, “Well, this is important, so we should build it and not care what it costs and not care about the consequences.”

If so many smart people think that we are in a bubble, why is money still flowing into data centers and other AI infrastructure at the rate that it is?

I’m not convinced that many people think it is a bubble. As I talk to people in technology, the most common response I get is not only is this not a bubble, but it’s probably the most important technology of our lifetime. We have an opportunity to build a super-intelligence, a god-like intelligence on top of all of these chips and buildings and this AI electricity thing we’re creating. And to say we should slow down at this point, according to the technology community, is just a huge error. But there are people outside of technology who say, “Oh, this is an incredible amount of spending.” The Bank of England said it. Other people are cautioning about it, but not inside of technology.

The United States and humanity broadly has had no shortage of bubbles throughout history. You mentioned the railroads; walk us through some famous American bubbles.

The railroad is probably among the most prominent in the US and that was, again, an enthusiasm for the idea. The same thing happened in the ’20s during electrification. In the 1920s we went from a single-digit percentage of rural areas having access to electricity, [to] by the end of the decade it was more or less ubiquitous. Everyone had access to electricity. But at the same time, that gave rise to this proliferation of utility companies, of ventures that were doing all kinds of questionable things in terms of overspending. You could argue that electrification and the frenzy around it gave rise to the stock market rise of the ’20s, which led to the crash of ’29 and helped precipitate the Great Depression.

People are pretty familiar with the telecom and dot-com bubbles, but the closest historical analogy to what’s happening now genuinely is railroads and electrification. In the same way that we don’t need to have two sets of tracks to Philadelphia, we probably don’t need the same number of companies delivering what are called these large language models, these AI models that people are using. These will naturally shrink.

How destructive are bubbles and what do they tend to destroy?

All of them do immense damage. It’s a question of how big the bubble is and where the damage goes.

So if you’re just holding an index fund and thinking you’re being very conservative, you’re actually soaking in AI right now. If everything reverses, goes 20 or 30 percent in the other direction, you’re much poorer than you were. That’ll change your spending. And that has implications for recessions.

Isn’t it always the case that the bubble bursts and then what it leaves behind is, maybe not something beautiful, but something workable?

That’s kind of a line of patter from the technology community. But the reality is almost every financial, every technology revolution has caused huge damage and can take decades before we get back to where we were before. And as the famous line in economics goes, in the long run, it may work out, but in the long run we’re also all dead.

A couple weeks ago, I went to the doctor to go over some test results. All was well — spectacularly average, even. But there was one part of the appointment that did take me by surprise. After my doctor gave me advice based on my health and age, she turned her computer monitor towards me and presented me with a colorful dashboard filled with numbers and percentages.

At first, I wasn’t quite sure what I was looking at. My doctor explained that she entered my information into a database with millions of other patients, just like me — and that database used AI to predict my most likely outcomes. So there it was: a snapshot of my potential health problems.

Usually I’m skeptical when it comes to AI. Most Americans are. But if our doctors trust these large language models, does that mean we should too?

Dr. Eric Topol thinks the answer is a resounding yes. He’s a physician scientist at Scripps Research who founded the Scripps Research Translational Institute, and he believes that AI has the potential to bridge the gap between doctors and their patients.

“There’s been tremendous erosion of this patient-doctor relationship,” he told Explain It to Me, Vox’s weekly call-in podcast.

The problem is that so much of a doctor’s day is taken up by administrative tasks. Physicians function as part-time data clerks, Topol says, “doing all the records and ordering of tests and prescriptions and preauthorizations that each doctor saddled with after the visit.”

“It’s a horrible situation because the reason we went into medicine was to care for patients, and you can’t care for patients if you don’t have enough time with them,” he said.

Topol explained how AI could make the health care experience more human on a recent episode of Explain It to Me. Below is an excerpt of our conversation, edited for length and clarity. You can listen to the full episode on Apple Podcasts, Spotify, or wherever you get podcasts. If you’d like to submit a question, send an email to askvox@vox.com or call 1-800-618-8545.

Why has there been this growing rift in the relationship between patient and doctor?

If I were to simplify it into three words, it would be the “business of medicine.” Basically, the squeeze to see more patients in less time to make the medical practice money. The way you can make more profit with lessening reimbursement was to see more patients do more tests.

You’ve literally written a book about how AI can transform health care, and you say this technology can make health care human again. Can you explain that idea? Because my first thought when I hear “AI in medicine” is not, “Oh, this will fix it and make it more intimate and personable.”

Who would have the audacity to say technology could make us more human? Well, that was me, and I think we are seeing it now. The gift of time will be given to us through technology. We can capture a conversation with patients through the AI ambient natural language processing, and we can make better notes from that whole conversation. Now, we’re seeing some really good products that do that in case there was any confusion or something forgotten during the discussion. They also do all these things to get rid of data clerk work.

Beyond that, patients are going to use AI tools to interpret their data, to help make a diagnosis, to get a second opinion, to clear up lots of questions. So, we’re seeing on both sides — the patient side and the clinician side. I think we can leverage this technology to make it much more efficient but also create more human to human bonding.

Do you worry at all that if that time gets freed up, administrators will say, “Alright, well then you need to see more patients in the same amount of time you’ve been given?”

I have been worried about that. If we don’t stand together for patients, that’s exactly what could happen. AI could make you more efficient and productive, so we have to stand up for patients and for this relationship. This is our best shot to get us back to where we were or even exceed that.

What about bias in health care? I wonder how you think of that factoring into AI?

Step No. 1 is to acknowledge that there’s a deep-seated bias. It’s a mirror of our culture and society.

However, we’ve seen so many great examples around the world where AI is being used in low socioeconomic, low access areas to give access and help promote better health outcomes, whether it be in Kenya for diabetic retinopathy, and people that never had that ability to be screened or mental health in the UK for underrepresented minorities. You can use AI if you want to deliberately help reduce inequities and try to do everything possible to interrogate a model about potential bias.

Let’s talk about the disparities that exist in our country. If you have a high income, you can get some of the best medical care in the world here. And if you do not have that high income, there’s a good chance that you’re not getting very good health care. Are you worried at all that AI could deepen that divide?

I am worried about that. We have a long history of not using technology to help people who need it the most. So many things we could have done with technology we haven’t done. Is this going be the time when we finally wake up and say, “It’s much better to give everyone these capabilities to reduce the burden that we have on the medical system to help care for patients?” That’s the only way that we should be using AI and making sure that the people who would benefit the most are getting it the most. But we’re not in a very good framework for that. I hope we’ll finally see the light.

What makes you so hopeful? I consider myself an optimistic person, but sometimes, it’s very hard to be optimistic about health care in America.

Remember, we have 12 million diagnostic errors a year that are serious, with 800,000 people dying or getting disabled. That’s a real problem. We need to fix that. So for those who are concerned about AI making mistakes, well guess what? We got a lot of mistakes right now that can be improved. I have tremendous optimism. We’re still in the early stages of all this, but I am confident we’ll get there.