This is all "perfectly spherical cow" economic theory which does not apply in reality. Do we really have a free and fair marketplace? Where's the competition that's shown up and been able to get any kind of substantial piece of the outsized profit of Facebook, Apple, Microsoft, Google?
What you should have learned is that in an idealized, perfectly rational, perfectly competitive market this is true, and you should have also learned that such markets do not exist in reality, not least of all because rational choice theory does not accurately model human behavior, but also because perfect competition is not an attainable condition.
However, the oversized profits are a significant source of incentive for investment in alternatives. Kharma will catch up.
And this wouldn't even be very difficult to enforce. Running SOTA models at useful speeds (which you'd need to actually compete - a local setup like you describe where time to first token is measured in minutes for a decent sized prompt is not going to cut it) requires a lot of compute. Which is to say, hardware, and energy to power it. Both things that can be tracked.
We're not talking about goods produced in bulk, but rather about a new intelligence that helps produce job shop type goods. That intelligence itself isn't a physical good though, and it can be copied freely and run elsewhere at almost no cost, and copyright won't save it. It's like toothpaste out of the tube once it's on the internet.
If we want to keep on track with the analogy to frame weaving machines, the difference is that capital and the availability of power were required to use those machines, which provided the moat required to protect profits.[Econ 101]
Models can be copied, and all of the investment in training used freely by anyone possessing a copy. This might have been how DeepSeek was trained, we'll never know. This is clearly different than in the case of weaving machines.
So instead of the cost of cloth, the cost of thought is going to drop considerably, and the quality of that artificial thought is (likely) at the lowest point it'll ever be. I can already run quantized versions of Deepseek on my CPU with 32 Gb of RAM. The future really doesn't seem to have an upper limit on what can be run locally.
With the ability to run models locally, the costs of running models will always have competition to lower prices. The other force is the capture of prompts and responses, whether covert or in the open, for other use, which must always be kept in mind.
The ability to build completely new things, will, I think, help to usher in a new era of innovation and allow anyone, anywhere, to start up companies to service the world.
[Econ 101] - In Economics 101, I learned that in a free and fair marketplace, it was impossible to maintain an outsized profit, because competition would always show up to take part of that profit.