AI has a climate problem — but so does all of tech

Photo by Vjeran Pavic / The Verge

We’ve been talking about AI a lot on Decoder lately; it’s unavoidable. But there’s one piece of feedback we’ve gotten that I really wanted to spend some time on: how the lightning-fast explosion of AI tools affects the climate.

After all, to run AI at scale, we need to build a lot of data centers and pack them full of power-hungry GPUs. That takes a lot of energy, and whether using all that juice is worth it comes up frequently when we talk about AI. It’s both a matter of practical concern — “can our aging grid support all of this?” — and a moral objection — “we shouldn’t build these systems because they’ll wreck the planet.”

What’s particularly complicated is that big tech companies like Amazon, Google, and Microsoft have spent the last few years working with governments around the world to set ambitious goals around sustainable energy usage so that we might slow the rate of climate change to simply “bad” instead of “catastrophic.” But now, with AI, all of those companies are blowing past their emissions targets and are actually getting worse over time. That’s not great.

But putting a bunch of computers in a data center and running them at full tilt is how basically everything works now. If you have a moral objection to AI based on climate concerns, you might have a moral objection to TikTok and YouTube as well, which are constantly ingesting and encoding millions of hours of video. You might have a moral objection to video games, which run both on power-hungry GPUs in people’s homes and often require intense data center workloads for online multiplayer. And I’ll take a guess, but I feel pretty certain anyone with climate concerns about AI has a pretty harsh assessment of crypto, too.

I mean, think about it this way: the Nvidia H100, which is the gold standard for AI GPUs, is pretty similar to the gaming-focused Nvidia RTX 4090 in terms of power draw. What framework should we use to evaluate the climate impact of those cards and how we feel about how they’re used?

It’s messy and complicated, and there are a bunch of apparent contradictions along the way. So it’s perfect for Decoder. To help sort it out, I’ve invited Verge senior science reporter Justine Calma on the show to see if we can untangle this knot. Let us know how we did.

Posted from: this blog via Microsoft Power Automate.