1. An operations team is reviewing cost and latency for a high-volume Azure AI Foundry workload that processes approximately 50 million tokens per day with very predictable, uniform request patterns. They are currently using a serverless API endpoint. Which deployment option would most reduce per-token cost and provide guaranteed throughput for this workload?
- A. Switch to a managed compute endpoint backed by a GPU cluster with auto-scaling enabled
- B. Remain on the serverless API endpoint but enable batch deployment mode
- C. Purchase provisioned throughput units (PTUs) and deploy the model on a provisioned managed compute endpoint✓ Correct
- D. Increase the serverless API endpoint's token-per-minute quota via a support request
Explanation
Provisioned throughput units (PTUs) are specifically designed for high-volume, predictable workloads. They offer a reserved capacity model that reduces per-token cost compared to pay-as-you-go serverless pricing and guarantee a consistent level of throughput. A GPU-backed managed compute endpoint with auto-scaling (option A) is suited for custom fine-tuned models, not foundation model deployments, and does not offer PTU economics. Staying on serverless with batch mode (option B) does not reduce per-token cost for real-time traffic and is intended for offline processing. Raising the quota (option D) removes rate limiting but does not reduce cost or provide guaranteed throughput.