LLMs, Competition, and the Government
Reflexively, I recoil, but there might not be better options if it do end up being like that
There’s an inverted U-shaped relationship between competition and innovation.
With no competition (including no serious threat of entry), there’s little incentive to innovate because you just extract monopoly rents, and there’s little extra profit to make by making a better product, past a certain quality level.
Under perfect (intense) competition, any product improvement gets competed away, driving profits back down to zero, so you get no returns on the investment you made to improve the product.
This is the justification for patents in, say, drugs. Any new drug can be cloned cheaply, creating intense competition post-innovation, so there’s no reward for investing all that dough in developing the drug. So, we reduce competition with patents.
You need just the right amount of competition to get strong innovation. Not too much. Not too little.
I wonder if this isn’t an argument that the LLMs will end up getting government money.
It sounds silly—and I tend to be reflexively against this kind of industrial policy—but there’s essentially zero property rights. No lasting product differentiation. The lagging models catch up to the frontier models in a matter of months, which is not material for current use cases. They produce (significant) economic value, of course, but I think there’s a real doubt the LLM producers can capture it, given the market structure.
For now, folks are still betting that there will be a winner of the race. But if this market structure sticks around too long, without any shakeout, investors won’t continue to believe that. They’ll start to believe there are no profits out there to offset the scale of the investment.
There is some return to packaging, software, harnesses you might bundle with the LLM. I might like the ergonomics of Claude or Codex more, but that’s small-fry differentiation compared to actual model performance. I can’t imagine the returns to this kind of product differentiation are enough to encourage the investment needed to make Silicon Gods.
When products are valuable, but the ex-post profits will be competed away, we tend to do… something. Patents or other restrictions on entry are the most common approaches, but they won’t work here—at least not so far as the US Government is concerned. China will keep producing these models, economic returns be damned. The Chinese models can’t be easily excluded from valuable markets, and the models are valuable…
And so, if someone can’t break away and competitive profits are all that await Anthropic and OpenAI, I’m betting we’ll end up subsidizing investors to encourage future investment. Yes, it’ll make this whole corner of the industry a little more… mucky with the usual rent-seeking, and the incentive problem of encouraging riskier investments, but I could see the government reasonably deciding it’s less bad than the alternative.
Thanks for reading!
Zach
Connect at: https://linkedin.com/in/zlflynn