A retail company deploys a small text-classification model in Azure AI Foundry to tag support emails by topic. Traffic is sporadic, sometimes zero calls for hours and sometimes a short burst during a sale. The team wants to minimize cost without committing to reserved capacity they might not use. Which deployment approach best fits this pattern?
Select an answer to reveal the explanation.
Short Explanation
Think about how you pay for electricity versus a gym membership. A gym membership charges the same amount every month whether you show up once or every day, which only makes sense if you use it constantly. Metered electricity charges you for exactly what you use, which fits usage that jumps around unpredictably. This team's workload is the electricity case: quiet for hours, then a short burst. Paying per token processed means the quiet hours cost almost nothing and the burst is simply absorbed and billed for what it actually consumes. Reserving fixed capacity ahead of time means paying for that gym membership every idle hour, which does not fit a bursty, low-average workload. Retraining the model does not touch billing at all, and standing up your own cluster to run it just adds a permanent baseline cost and operational burden on top of a problem that a metered, serverless option already solves cleanly.
Full Explanation
The correct answer is B. A pay-as-you-go serverless API deployment charges only for tokens actually processed, which matches sporadic, unpredictable traffic where the workload sits idle for long stretches and spikes briefly. There is no upfront capacity commitment, so the company pays close to nothing during quiet periods and scales automatically during the sale-day burst. Option A is incorrect because provisioned throughput units reserve fixed capacity that is billed whether or not it is used, which is wasteful for a workload that is mostly idle rather than steady and high-volume. Option C is incorrect because fine-tuning changes what the model has learned, not how it is billed or scaled, and would not address the cost concern raised in the scenario. Option D is incorrect because standing up and operating a dedicated Kubernetes cluster adds infrastructure management overhead and a baseline compute cost that runs regardless of traffic, which is the opposite of what a sporadic workload needs.