(This is all written with low confidence.)

The usual story of the Baumol effect—or Baumol’s cost disease for the pessimists—is as follows. Sectors of the economy directly affected by technological progress (historically, anything like high-tech manufacturing) experience strong productivity growth. Other sectors (education, healthcare) experience little or no productivity growth. This manifests as the price of some goods (televisions, computers) falling, which is obvious, but also manifests as the price of other goods (education, healthcare) rising. This last effect is the more surprising one, and the one named after Baumol.

I long had it drilled into my head that this is something that raises the prices for human-intensive sectors of the economy: the service economy, essentially. But it seems that today, this shouldn’t be true. This observation is mostly inspired by Andreeson’s essay “It’s Time to Build” and related writings by just about everybody else.

For the last five decades, productivity in semiconductor production has grown faster than just about anywhere else. Because the infrastructure for software is so cheap today, productivity for software has also risen dramatically. It has risen so much, and consequently the price has dropped so much, that most software is free; in other sectors of the economy we might use the phrase “too cheap to meter”. Related goods are not quite too cheap to meter, but are close. I am particularly thinking of access to LLMs, and other computation-intensive cloud services.

Baumol’s effect then suggests that the price of everything that isn’t being powered by cheap silicon should rise (relative to the counterfactual in which silicon is expensive). But of course, the products that are affected by cheap computation have changed dramatically over the last several decades. Today, the most difficult areas for computation to affect are those that are intrinsically physical: building bridges, building ships. Infrastructure and heavy industry, in general. Meanwhile, service sectors (and particularly those service sectors where people spend most of their time interacting with a computer) are ripe for automation. So the Baumol effect should predict a relative rise in the share of the economy going to infrastructure and manufacturing, with a relative fall in the share going to the service industry.

Miscellaneous other observations:

• The optimal allocation of government resources probably tilts towards more investment in sectors whose price hasn’t dramatically dropped.
• Advanced economies tend to focus on producing higher-priced goods. Does that mean we should see re-shoring of manufacturing and infrastructure jobs, and off-shoring of service sectors and anything else that can be done by GPUs? Probably, unless protectionist laws (including data-protection laws) prevent this.
• Many jobs can’t be completely automated (if only for regulatory reasons), but will be made cheaper by the fact that the human is just acting as an interface to the LLM (or other AI). These can be off-shored to low-wage economies even in the presence of data-protectionist laws. Tech companies do this sort of thing for their support jobs, but lawyers and researchers (for example) could plausibly follow.