Beauty of scale. Saving ten milliseconds a hundred times is just a second. But do it a billion times and you've shaved off ~4 months.
In most real jobs those ten milliseconds will add up to what, 5 seconds to a minute?
Those 10 ms may quite well mean the difference between success and failure... or they may be completely irrelevant. I don't know if this is knowable.
* ... if they're at ClosedAI or Facebook or something. If they're at some startup selling "AI" solutions that has 10 customers, it may be wishful thinking that they'll reach ClosedAI levels of usage.
If it's about saving the users time, it's very non-linear. And if it's not a scalable read-only service, the costs will be very non-linear too.