And when you have enough spending to account for 1%+ of revenue for the AI hardware companies?

You can get the engineers from those very hardware companies to do bespoke optimizations for your specific high load use cases. That's something a startup would struggle to match.