Why it matters: model quality and latency must survive the actual runtime, concurrency and failure policy.
A strong answer should cover: quantisation, batching, prompt length, warm/cold behaviour, load distribution, timeouts and fallbacks. Use the inference engineering discussion as background; do not repeat vendor speed claims as a guarantee.
Follow-up: “What exact load trace and deployment configuration should a supplier deliver so we can reproduce the result?” Request the raw metric definitions alongside headline percentiles.
Interview status: Not yet interviewed.