Agents turn one request into dozens of inference calls, and that shows up fast on the AI bill. Vasanth Mohan from
@SambaNovaAI explained speed per request versus throughput per chip, splitting prefill from decode, and why you should test on your own workloads.