In the late-stage aggregator paradigm that we're in, there's little room for anybody but the aggregators to participate.

AI's cost structure threatens to exacerbate that dynamic.

The web reduced the floor size and cost of the minimum viable experience, which allowed even small experiences to be viable.

The cost structure of AI messes with what we expect.

We've come to assume that consumer experiences will be free.

But LLMs have a non-trivial marginal cost.

Orders of magnitude below human labor.

But orders of magnitude above generic compute.

By default, this will lead to faster aggregation, as only aggregators can eat the costs by cross-subsidizing from their money makers.

The free model leads to one-size-fits-all, hard software.

Sometimes there are use cases where you want 30 cents of computation to be put into the task.

But sometimes there are important, challenging tasks where you want 30 dollars of computational work to go into the task.

With the "everything is free" model you get one-size-fits-all software and you can't choose.