The cloud has a material footprint

“Cloud” is a clever name. It gives computation an image without weight, machines or an address.

Most of the time I’m happy with that. I don’t need to inspect a power station before I send a file. But once we start making claims about the future of production, the infrastructure has to come back into the picture.

In April 2025 the International Energy Agency published its Energy and AI report. In its base case, global electricity consumption of data centres reaches around 945 TWh in 2030, more than double the 2024 level. That’s a projection for all data centres. It doesn’t tell you how much energy one prompt uses, and not all of the increase is AI.

A large global number can’t tell me whether a specific use is worth it. And an efficient single operation can’t tell me what happens when it’s repeated at a much larger scale.

I thought about this with GrasShopper. One buyer’s guide meant several agents and many model calls so that one person could choose sneakers. Was that better than three hours with 15 browser tabs open? I’d guess yes, but it stays a guess until somebody measures it.

The same goes for physical production. UNEP’s Global Resources Outlook 2024 describes how resource extraction keeps growing and how unequally its impacts are spread, so the cost of making something often lands far from the person enjoying it.

I don’t think every new capability is an ecological mistake. A better simulation can save a failed physical prototype (I run a simulation company, so I like this argument a lot). But the benefit has to be shown against the alternative. Being digital doesn’t prove it.

Take two ways of designing a component. One uses more computation and less material, the other less computation and a heavier part. To compare them I need to know how many parts will be made, how long they last, what powers the computation and what happens to the part at the end.

Efficiency doesn’t settle it either. A cheaper process can free resources for something else, and it can also encourage a lot more production. Usually it does both.

Sutton’s The Bitter Lesson puts growing computation at the center of long-term AI progress. Next to the numbers above, I’d add a question his essay doesn’t deal with: available to whom, and at what total cost?

What I try to do is name the boundary before celebrating the result: - Am I counting only operation, or equipment and construction too? - What am I comparing with: the previous method, a different solution, or doing nothing? - Who gets the benefit and who gets the dependency?

None of this takes the fun out of making things. A design that does something worthwhile with less is an achievement I’d love to understand in detail.