Do they share latency / concurrency figures (aka the deployment shape that gets them that token/s/cluster? Would be interesting if they're using deepEP/WideEP + PP and TP