But fab is extremely capital intensive and complex to setup, I don’t think this is the realm of startups anymore.
Then a completely new way of steel production came about. Nothing iterative about it, a completely different set up that came with its downsides but had better economies of scale.
If there is a change to come, it's not by building a bigger fab that has higher capital costs. It's by making the idea of a fab entirely obsolete by finding a way to bypass all the steps involved in the current manufacturing process and producing a slightly inferior product in some way but isn't a multi-month long process with capital costs measured in $10s of billions. Fabless manufacturing was a great boon in its time by changing the business model around fabs but it didn't change the underlying engineering. Somebody, ultimately, owned a fab in the chain.
The real trick will be something that bypasses the whole idea of pure crystal ingots as the core of the process that has to be processed in batches with 100s of steps. Something more continuous and simplified would be revolutionary. Hypothetically, imagine instead of growing a perfect crystal and etching or doping it, you grew an imperfect crystal but that growth was controlled in the direction that got you 80% of the way to the layout of the final design. Or imperfections were embraced and utilised as a feature for randomised bias in AI cores than a flaw that wasn't tolerated.
All of this, hypothetical of course, but my point is even if we are at the end of this fabbing game, the real trick is creating a completely new game entirely.
I'm convinced that there are plenty of design space that is not explored. In particular, the fabrication process today is mainly 2 dimensional with a very high necessity of control.
If you relax one of this strong constraint and build the system bottom-up instead of top-down. It reduces to a search for the chemical compound, which can self-assemble into an interesting structure. It should look like a biological machine but running at computer frequency.
Plenty of nano-computing units interacting. For example today we can already build some silicon-nanorobots ( https://www.electronicdesign.com/industrial-automation/artic... ). Give each one a small memory unit, processor, and radio-link, and you can build tons of them on imperfect crystal using standard process. Then your processor is a 3d volume of these bots instead of a 2d-surface. You will still be constrained by heat dissipation but if you put them inside a liquid, it would be less problematic, but you'll have to compensate for nano-processor brownian drift.
https://www.gammabeyond.com/wp-content/uploads/2019/05/Why-C...
Page 19 talks about the history of the mini mill in steel production.