I've been thinking about the environmental conversation around AI, and I've realized that I've been struggling to articulate something that feels obvious to me but isn't quite so simple when you start pulling it apart. AI and data centers are inseparable. Without enormous amounts of compute, we don't get today's frontier models, and training those models requires enormous amounts of electricity, cooling, water, chips, land, and infrastructure. So when someone points at the environmental cost of AI, they're not wrong. The part I'm struggling with is whether we're really talking about the environmental impact of AI itself, or whether we're talking about the consequences of the economic race we've built around it.

I think those are two different things. AI is a technology. Data centers are the infrastructure required to run that technology at scale. The explosion in data centers is largely an economic phenomenon driven by the enormous amount of money companies believe can be made by being the company with the best AI. Those things are connected, but they aren't the same thing, and I think that distinction matters.

I keep coming back to smartphones because I think the comparison is surprisingly useful. My first cell phone was a Cingular flip phone, back when T9 keyboards were still a thing and text messages cost ten cents apiece. The phone itself was basically free with a contract, and the service was something a high school kid with a part-time job could afford. Then came the Nokias, with Snake and those ridiculous batteries that seemed like they could run forever. After that came BlackBerry and the first wave of touchscreen smartphones, and that really was a revolution. The phone stopped being primarily a phone and started becoming a computer that happened to fit in your pocket.

For a while, every new generation felt like a meaningful leap. I bought a Samsung Note 4 when it came out, partly because the giant screen and stylus actually felt like a different kind of device. Eventually I bought a Samsung S22. But somewhere along the way, I stopped caring about having a flagship phone. Not because I couldn't afford one, but because I couldn't justify one. A $100 Samsung does almost everything I need. The screen is good enough, the processor is good enough, the camera is good enough, and it runs the apps I use every day. Eventually the battery will wear out or I'll drop it and break the screen anyway, so spending another thousand dollars for the latest flagship doesn't really solve a problem I have.

I suspect the smartphone manufacturers are dealing with that same problem now. Phones haven't stopped getting better, but the improvements have become increasingly difficult for most people to notice. The difference between the first generation of smartphones and the phones that came before them was enormous. The difference between this year's flagship and last year's flagship is much harder to feel.

AI may be approaching that same inflection point.

I've used models ranging from GPT-4 through the latest and greatest, GPT-5.6-Sol.^[1] They're all very good, and there are certainly measurable differences between them. I'm not arguing that the models haven't improved. What I am saying is that, as a software engineer who uses these models constantly, I haven't experienced another smartphone-2007 moment. I haven't suddenly discovered that I can do something that was fundamentally impossible a few months ago. The improvements I've experienced have increasingly felt incremental rather than revolutionary.

That may just be my particular use case, and I could be completely wrong. There may be domains where each new generation represents an enormous leap in capability, and there may be another architectural breakthrough around the corner that makes today's models look primitive. But if the returns really are beginning to diminish, it raises an uncomfortable question about how much compute we're willing to spend chasing the next few percentage points of improvement.

Because the economic incentive to keep chasing them is enormous.

This is where capitalism enters the story, and I don't think we should dance around that. The same system that gave companies an enormous incentive to make AI useful also gives them an enormous incentive to keep making it better, whether or not the incremental improvement is particularly valuable to the rest of us. If OpenAI can build a model that's better than Google's, there's an enormous amount of money at stake. If Google can build one that's better than OpenAI's, the same is true in reverse. If another billion dollars of compute might give one company an advantage over another, someone is going to spend the billion dollars. Then someone else spends two billion, and eventually we're talking about tens of billions of dollars and entire new data centers being built to support the race.

The problem isn't that these companies are trying to make money. That's what they're supposed to do. The problem is that the profit motive doesn't have a natural stopping point. The market asks whether the next model is worth more to the company than it costs to build. It doesn't necessarily ask whether the electricity, water, land, transmission infrastructure, semiconductor manufacturing, and environmental impact are worth the incremental improvement. Those costs can become externalities, spread across communities and ecosystems that aren't part of the decision to train the next model.

That's what makes me wonder if the environmental problem with AI isn't really AI at all. Maybe it's the race we've created around AI.

We don't necessarily build another enormous data center because humanity desperately needs a model that's five percent better. We build it because our competitor might build one first. They build theirs because they don't want us to get ahead. The cycle feeds itself, and the technology becomes inseparable from the economic incentives surrounding it.

I don't think that means we should stop developing AI. There are too many potentially transformative applications, and we're still early enough that we don't know what breakthroughs are ahead of us. Maybe the next generation of models really will unlock scientific discoveries that justify extraordinary amounts of compute. Maybe we're not anywhere close to the point of diminishing returns.

But maybe we are.

And if we are, perhaps the next phase of AI shouldn't be about building ever-larger models. Perhaps it should be about figuring out what we can do with the models we already have. Better agents, better memory, better tools, better context, better integrations, and better software could make today's models dramatically more useful without requiring every improvement to come from another enormous training run.

That's essentially what happened with the smartphone. We didn't stop improving phones because we reached some magical point where there was nothing left to do. We reached a point where making the phone itself dramatically better became less important than figuring out what we could do with it. The revolution moved from the hardware to the experience.

Maybe AI is approaching the same transition.

The difficult part is that "good enough" isn't a particularly attractive concept in a competitive market. If your competitor is still chasing something better, voluntarily deciding that you've reached the point of diminishing returns feels a lot like surrender.

That's where I keep thinking about Ahab.

Ahab didn't hate the whale because whales were bad. He became obsessed with catching one particular whale, and eventually catching it became more important than what the pursuit was costing everyone aboard the ship.

Maybe that's the question we should be asking about AI.

The problem may not be that we invented artificial intelligence. The problem may be that we've attached an economic incentive to chase an ever-better version of it without necessarily having a point at which we can say, "This is good enough."

I don't know where that line is. I'm not even sure we've reached it. But I think we're going to have to start looking for it.

Because AI may be the technology. Data centers may be the infrastructure. But the race between them is something we created.

And unlike the technology itself, that race is something we can choose to change.