Thinking a lot about Fly.io's "We Were Wrong About GPUs" and Oxide and Friends' episode with Cerebras.
Us true nerds love learning and developing all we can at the lowest possible level, but when you need to sell a product that people need to buy in order to build and sell their product, then you need to provide an offering at the highest level of abstraction that permits them to do their job.
If you provide fast easy raw GPUs via an API... If you provide the worlds biggest chip with the most integrated memory... All is lost if no one wants to work at that level of abstraction. Maybe all your users just want an LLM API.
That's unfortunate!