Product wants QPS to jump from 500 to 5,000 for a 100 m...

The decision depends on whether the bottleneck is memory or compute. Because the corpus size is constant, the memory-driven shard count remains valid. The increased QPS is a compute-load issue, which is best solved by scaling horizontally via replicas or vertically via more cores. Splitting into more shards is only the correct lever if the latency SLA tightens to the point where per-query cost must be reduced.