AI-300 Practice Question 50: Fine Tuned Deployment Hosting Cost
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 5 · question 10 of 10
A team fine-tuned five Azure OpenAI candidate models and deployed all of them so that reviewers could compare outputs. Two weeks later, one model was chosen for production, and the other four receive almost no traffic. Yet the monthly bill still shows a large recurring charge for the fine-tuned models. What should the team do to stop the unnecessary charge?
Show answer and explanation
Answer:
Azure OpenAI fine-tuned models are charged in three ways: *training*, *hosting* (an *"hourly cost per deployed model (applies even if the model is unused)"*), and *inference* per 1,000 tokens. The Foundry cost documentation warns that *"fine-tuned deployments incur hosting charges while deployed, even during low usage periods. Remove or scale down deployments that you don't need."* Deleting the four unused deployments stops their hourly hosting cost (option A).
- B is wrong because a token limit affects the size of individual responses. It does not change the hourly hosting charge of a deployed model. - C is wrong because hosting is billed by the hour while a deployment exists. Not calling it only avoids inference charges, which were already close to zero. - D is wrong because the processing option controls where inference runs. Switching a Standard or Global Standard fine-tuned deployment to another processing option does not remove its hourly hosting charge.
Reference: Plan and Manage Costs - Microsoft Foundry