Source: Microsoft Foundry — Provisioned Throughput (PTU)

How Foundry's provisioned throughput units work: model-independent quota, separate capacity, sizing, spillover, hourly billing and reservations.

Contents

Summary

Concept page for dedicated-capacity deployments.

Key claims

  • A PTU is a fixed amount of model processing capacity held for your deployment whether or not it is used.
  • PTU quota is model-independent but scoped per subscription, region and deployment type (Global/Data Zone/Regional are separate pools). Each model has a minimum PTU count and its own TPM-per-PTU.
  • Quota is not capacity: deployments can fail for lack of regional capacity, and scaling down releases capacity you may not get back.
  • Sizing combines RPM, input/output token sizes, an output-to-input weighting and the cache rate. Cached tokens don’t consume PTU for models before GPT-6 Astra; for later models count cache reads and writes.
  • Spillover can route 429/non-200 overflow to a standard deployment in the same resource, globally or per request via the x-ms-spillover-deployment header. Only Azure OpenAI models support it, not DeepSeek or Llama.
  • Billing is hourly $/PTU/hr from creation to deletion, or discounted via 1-month or 1-year Azure Reservations. Reservations are decoupled from deployments and don’t guarantee capacity: deploy first, then reserve.
  • Priority processing is a pay-per-token alternative with a per-model latency target.

Feeds

Azure AI Foundry

Suggest changes →