Three Fixes for the Slow Requests

A photo-tagging endpoint serves 1,000,000 requests a day. Profiling splits them into three kinds and times every stage (all times in ms):

kind of request	share of traffic	network	preprocessing	model	postprocessing
typical	97%	40	30	90	10
slow connection	2.5%	500	30	90	10
huge upload	0.5%	40	800	90	10

The team tracks the slowest-1% latency: the smallest time $T$ such that no more than 1% of requests take longer than $T$ .

There is budget for exactly one fix:

Plan X: quantize and prune the model so it runs three times faster. The model stage falls from 90 ms to 30 ms, for every request.
Plan Y: a faster image decoder. Preprocessing for huge uploads falls from 800 ms to 100 ms.
Plan Z: servers closer to users. Network time for slow connections falls from 500 ms to 150 ms.

Which plan leaves the lowest slowest-1% latency, and what is that latency?