I have apparently broken LinkedIn because everything it shows me now is about data centers. Somebody is building one, somebody else doesn't want one built near them, somebody is worried about how much electricity they use, somebody is building a nuclear plant to power one, and there seems to be an endless stream of announcements about another enormous piece of land being turned into a building full of computers. Maybe this is just what LinkedIn thinks I want to read because I run a data infrastructure company, but after seeing enough of these stories I started thinking about Malthus.
Thomas Malthus, writing more than 200 years ago, had a fairly simple concern. Population could grow faster than our ability to produce food, and eventually there would be too many people for the resources available to support them. The logic wasn't ridiculous. In fact, looking at the world available to him, it probably seemed obvious. What he couldn't see was how much we would change the equation. We got dramatically better at agriculture, fertilizer, transportation, refrigeration, energy and a thousand other things. The population grew enormously and somehow the carrying capacity of the system grew with it.
I wonder if there is a version of that happening with data right now. There is too much data. There isn't enough compute. There isn't enough electricity. There aren't enough data centers. We need more chips, more power plants, more transmission, more cooling and more giant buildings to hold all of it. Some of that is unquestionably true. The IEA thinks electricity consumption from data centers could more than double between 2024 and 2030, so this isn't an imaginary constraint.
But every time I read another story about the need for another gigantic data center, I find myself wondering about the other side of the equation. Are we actually good at using all this data and compute?
I'm not sure we are. Think about what happens inside a large company when somebody wants to answer a new question. The data might exist, but it is sitting in five different places. Somebody has to find it, get permission to use it, move some of it, clean it, reconcile it with something else, build a pipeline, add the business logic and finally run the analysis. Then another person asks a somewhat different question and a surprising amount of that work happens again. We have gotten incredibly good at creating, storing and processing data without necessarily getting equally good at reusing what we already know about it.
This is where something I've been calling gold dollars comes into it for me. I've found that most companies place their expenses into two meta-categories: blue dollars and green dollars. Blue dollars are easy to understand: spend money on technology and reduce costs. Green dollars are also easy: spend money and create more revenue. Gold dollars are harder to put into a spreadsheet because they come from being able to ask questions that weren't practical to ask before. A company discovers something it couldn't see, understands a relationship it didn't know existed, or makes a decision it simply couldn't have made with the information available in the old way.
That seems particularly relevant if we are about to spend hundreds of billions of dollars building the physical infrastructure for AI. Of course we should care how much compute we can produce and how cheaply we can produce it. But I think we should care just as much about what we get from each additional unit of it. If I double the amount of compute available to an organization and all I do is allow it to process twice as much data in essentially the same way, that's useful. If I make the data it already has interoperable, preserve the context around it and make what the organization learned answering one question reusable for the next question, something more interesting happens. The number of questions it can practically ask starts to expand.
I don't think this means we don't need the data centers. We probably need a staggering number of them. I just think we're spending an enormous amount of time talking about how to increase the supply of compute and not nearly enough time talking about the productivity of compute. At some point, endlessly copying, moving, transforming and reprocessing data because our systems don't understand how things relate to one another starts to look less like an unavoidable consequence of AI and more like an infrastructure problem we haven't solved yet.
Malthus looked at population growth and assumed that eventually we would run out of the things required to support it. What he underestimated was our ability to get radically more productive with the resources we had. I don't know if the analogy completely holds, but every time LinkedIn tells me we need another several-billion-dollar building full of computers because the world is producing too much data, I can't help wondering whether we're making a similar assumption.
Maybe we don't just need more data centers. Maybe we need to get a lot smarter about what we're doing inside them.