This conversation on Paul Krugman's Substack (Nobel Prize economy winner) is just really interesting.
But most of inference is not you and I. This is a mistake we make all the time. People think that we are the story. With respect to inference, most of the global inference from consumers—from you and I and others—could be satisfied by a single data center in northern Virginia. That’s how small a fraction of the total load we are with respect to inference worldwide.
So 60%, let’s say, is training. We’re maybe 5 or 6% of the total workload of data centers. That bit in the middle, a huge chunk of that, is software itself—is coding, which turns out to be a huge profligate use of tokens.
# Dec 6, 2025