AI-300 Practice Question 6: Provisioned Throughput Units
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 1 · question 6 of 10
A customer-facing copilot built on Microsoft Foundry runs at a steady, well-measured request rate all day. On its standard deployment, response times vary with overall demand and the pay-per-token bill has become the largest line item in the product's budget. Finance has approved a one-year commitment. What should you change?
Show answer and explanation
Answer:
A provisioned deployment holds a fixed amount of model processing capacity exclusively for your deployment, measured in provisioned throughput units (PTUs), and carries a defined latency target per model. Microsoft names this the right choice for predictable traffic, latency-sensitive requirements, and production-scale volume where per-token billing becomes expensive, all three of which describe this copilot. Azure reservations then apply a discounted $/PTU/hour rate in exchange for a one-month or one-year commitment (option C).
- A is wrong because quota is a policy limit with no cost and no guarantee. On a standard deployment, inference capacity is shared across customers and throughput still varies with demand, so latency stays unpredictable and per-token billing is unchanged. - B is wrong because batch deployments return results asynchronously and carry no latency SLA. They're for bulk processing, not an interactive copilot. - D is wrong because adding regions multiplies the operational surface without changing the billing model or removing the shared-capacity latency variability.
Reference: Provisioned throughput for Foundry Models - Microsoft Foundry