If google has 10 million servers then a 2% improvement would be like freeing up 200,000 servers. That's a significant amount of value added!

It usually makes more sense to think in terms of cores rather than servers, because server density has always been increasing :)

If Google has 100 data centers they can sell 2 of them.

In this day and age, that's a pretty large compiled code performance improvement from just one feature.

Across the board improvements from changing register allocators is surprisingly difficult to get. Couple of percent sounds great.

It's usually done heuristically in a fashion that scales badly with how many instructions you're considering at a time so it's really easy to overfit to today's benchmark.