The tricky thing about compute is you can plow thousands of...

The tricky thing about compute is you can plow thousands of dollars into top-spec hardware to run within your organisation, but an office machine used by an individual will be at 1% capacity most of the work day, and 0% outside office hours when it’s switched off. So your IT might run some rack-mounted services to normalise loads, but those are still idle outside 9-to-5 in a domestic org.

If you self-host your hardware your compute demand is highly volatile (within office hours), inelastic, and will never get close to the capacity you actually paid for over any 24 hour span.

If you host an international SaaS service for others, the economics change though, because your demand curve can be flatter (across time zones), and demand can be elastic because you can lower prices at times when demand is low, thus increasing demand and revenue, at very low marginal cost.

For these cost reasons I think hosted services will dominate the AI space for some time, except in cases where privacy, security, or legislation requires they be run locally.