An app calls a model deployment that is assigned 100,000 TPM. Load tests fail with HTTP 429 errors even though token throughput stays far below 100,000 tokens per minute. Each request is short, and the app sends thousands of requests per minute. What is the cause?