AWS cut ASR inference from 16 GPUs to 4 with CUDA MPS and no model change.

Running a clinical fine-tune of NVIDIA Parakeet TDT 0.6B V2, AWS partitioned each L40S into four MPS execution contexts at 25% SM each, layered on Triton dynamic batching and ONNX Runtime with TensorRT encoder optimization (AWS). On g7e.4xlarge at concurrency 32 it reached 92.1 requests per second per GPU with 352ms mean and 769ms p99, inside a sub-650ms mean SLA. If you serve small speech or embedding models, the concurrency sweep from 1 to 100 in that post is reproducible.