My thesis is very simple.
1. LLM models will become more and more efficient - needing less and less compute, memory and network requirements.
2. LLM models will also become narrow and deep in their context - reinforcing the low need for hardware resources even further.
3. LLM models will start getting deployed locally with a small condensed data set - and would ping the external network less - only when the embedded local data happens to not have any information relevant to the question
Most of the above is happening already so may be i just woke up from a deep slumber.
This would mean the compute demand will actually fall - reducing the capex for all ai provider hyperscalers ie google/aws/azure and dropping the demand on nebius/coreweave/applied digital.
This would cause hyperscaler stocks to go up and dc hardware/infra provider stocks to go down.
Or i am totally wrong and LLM models will just expand infinitely to fill up as much compute as provided to them.