Are we helping to destroy the planet every time we ask ChatGPT a question? Stephen Witt joins us to answer that and other burning questions about the future of artificial intelligence. Plus:

Caroline Mimbs Nyce
Newsletter editor

A single data center can use as much power as the city of Philadelphia. And they’re popping up everywhere. These sprawling buildings, filled with rows of computing equipment, are the factories of the A.I. economy; they power all those mundane chatbot searches, sucking up tons of energy in the process. As the OpenAI C.E.O. Sam Altman put it, “I do guess that a lot of the world gets covered in data centers over time.”

For our latest issue, the reporter Stephen Witt was invited (“after what felt like two hundred phone calls”) inside a Microsoft facility, still under construction. I caught up with Witt to discuss what he saw there—and what A.I.’s massive energy consumption means for our planet.

This conversation has been edited and condensed.

What does a data center actually look like?

It’s a barn. It’s a giant shed full of microchips. From the outside, they keep them as anonymous and boring-looking as possible, and then the inside is just racks and racks of computing equipment stretching off into the distance.

Was it totally crazy to be in there?

It does not feel like a place a human being should be inside. In fact, they try to limit the amount that people go into them. They’re totally clean, contamination-proof, humidity- and temperature-controlled. It feels like going into a bank vault almost. You’re inside the computer’s brain.

Talk to me a little bit about how these data centers are being built.

It’s one of the largest movements of capital in human history. You really have to go back to electrification, or maybe the building of the railroads or the adoption of the automobile, to see a similar event in terms of money deployed.

Jensen Huang, the co-founder of Nvidia, has called the data center the A.I. factory: data goes in and intelligence comes out. All of this is being built to develop neural networks, these little files of numbers that have extraordinary capabilities. That’s what all that computing equipment is in the shed doing. It’s fine-tuning your neural network until it has superhuman capabilities. It’s an extremely resource-intensive process.

Essentially, A.I. is a brute-force problem, and I don’t think anybody anticipated how much of a heavy industrial process the development of it would be.

Are we going to completely destroy the planet with A.I.?

Yes.

Great.

So, we’re already on track to cook the planet. It’s a huge problem, even before any of this happened. Now, having said that, I think the data center build-out is totally irresponsible from a climate perspective. But I don’t know what the answer is, other than building tons and tons of carbon-free energy. You just have to make so many nuclear power plants. And we have to do it at a scale that gets the cost down.

Are we contributing to this every time we use ChatGPT?

If you’re just asking A.I. questions, don’t worry about it. You use just as much electricity watching TV or turning on the light. Not a problem.

If you’re building a lot of short-form, A.I.-generated video content, that is like running a microwave all day. If you’re in pro-research mode, and A.I. goes and thinks for an hour before it gives you an answer, you know it’s using a lot of juice. The A.I. companies will not tell us how much power these things use. We had to back into an answer through open-source academic work, and then take a guess. But our equivalent for a three-thousand-word term paper was about three minutes of using your microwave.

That’s a lot.

Is it? When you microwave food for three minutes, are you, like, “Oh geez, I’m destroying the planet”? It’s an equivalent concern.

Is using A.I. driving up utility costs?

Yes. The grid does not have the capacity to support this right now. And a massive build-out is going to take years.

Electricity costs are going up anyway, due to inflation—but they’re way outpacing inflation. This is putting tremendous strain on America’s electrical infrastructure, and you, the rate payer, are picking up part of that.

And this is already happening?

Oh, yeah, it’s well under way. You’re paying. The grid is just a giant pool of electricity. When you connect the data center to the grid, it’s like someone coming and sticking a fire hose into a well. This big snaking thing is dipped into the pool, and starts draining it from everyone else. It makes everyone’s costs go up.

We’re essentially paying for A.I. companies to train their models.

In a way, yeah.

What do you say to people who feel stressed out about all this?

I am also stressed about this. I mean, I go back and forth. The end goal here is that most of what humans do becomes obsolete.

Do you really think that?

Yeah, one hundred per cent. I think that in the future, all forms of labor will at least be conceptually done by a computer. With the combined push for robotics and hyper-intelligent computing systems, what’s left? I guess we should all go to clown school—study live theatre, or something.