Agreed with everything here. I think one thing that’s useful here is to make a distinction from the good ‘ol fashioned natural monopolies/oligopolies from economics (where extremely high fixed costs and exhaustable demand make it difficult/impossible to have many players in a market—so long as marginal costs are low relative to fixed, even if not low on an absolute scale), and aggregators. The classic version of natural monopolies/oligopolies are utilities.
Ben Thompson’s big thing with aggregators, who don’t always capture all the value in markets, is that they control demand with zero/very low marginal costs. Amazon is an aggregator mainly by virtue of the demand-control side. The classic version of an aggregator is Google or Facebook, on both the consumer demand side and (more importantly) aggregation of ad demand.
Everything you’re saying still works perfectly fine with natural monopolies/oligopolies, who have non-zero marginal costs. High costs for training frontier models and (relatively) low costs for serving them and finite demand for these chat interfaces mean that you can have only a few players win. That doesn’t necessarily mean they’re wildly profitable, even if they are entrenched, but it does give them market power. Morris Chang’s brilliance was recognizing that while semiconductor fabs were not natural monopolies yet… but as costs scaled, they would be. And thus, it’s beneficial to sink massive amounts of money into fixed costs in order to dominate it. That’s actually been something I’ve been thinking about for frontier models over time, which could be a way out of these companies’ problems.
It may seem like a distinction without a difference, but Anthropic has a good API business (arguably better than OpenAI). But they wouldn't care who sends customers (the demand side) to them, just as Waymo/other robotaxis might be totally ok with Uber controlling demand as an aggregator. ChatGPT is a consumer aggregator because it has the brand for it. THAT may be the saving grace for OpenAI, and they could (even if they don't/won't want to) be an aggregator for all the other frontier models as well. The reality of all of this is it makes the entire business side of things quite nuanced and fluid, without extraordinarily clear answers (yet), but interesting directions it can go.
Agreed with everything here. I think one thing that’s useful here is to make a distinction from the good ‘ol fashioned natural monopolies/oligopolies from economics (where extremely high fixed costs and exhaustable demand make it difficult/impossible to have many players in a market—so long as marginal costs are low relative to fixed, even if not low on an absolute scale), and aggregators. The classic version of natural monopolies/oligopolies are utilities.
Ben Thompson’s big thing with aggregators, who don’t always capture all the value in markets, is that they control demand with zero/very low marginal costs. Amazon is an aggregator mainly by virtue of the demand-control side. The classic version of an aggregator is Google or Facebook, on both the consumer demand side and (more importantly) aggregation of ad demand.
Everything you’re saying still works perfectly fine with natural monopolies/oligopolies, who have non-zero marginal costs. High costs for training frontier models and (relatively) low costs for serving them and finite demand for these chat interfaces mean that you can have only a few players win. That doesn’t necessarily mean they’re wildly profitable, even if they are entrenched, but it does give them market power. Morris Chang’s brilliance was recognizing that while semiconductor fabs were not natural monopolies yet… but as costs scaled, they would be. And thus, it’s beneficial to sink massive amounts of money into fixed costs in order to dominate it. That’s actually been something I’ve been thinking about for frontier models over time, which could be a way out of these companies’ problems.
It may seem like a distinction without a difference, but Anthropic has a good API business (arguably better than OpenAI). But they wouldn't care who sends customers (the demand side) to them, just as Waymo/other robotaxis might be totally ok with Uber controlling demand as an aggregator. ChatGPT is a consumer aggregator because it has the brand for it. THAT may be the saving grace for OpenAI, and they could (even if they don't/won't want to) be an aggregator for all the other frontier models as well. The reality of all of this is it makes the entire business side of things quite nuanced and fluid, without extraordinarily clear answers (yet), but interesting directions it can go.