singletenant.ai

Guides

Azure OpenAI PTU explained: provisioned throughput vs pay-as-you-go, pricing model, global/data-zone/regional deployments

What an Azure OpenAI PTU buys, how it is billed against pay-as-you-go, what it costs at published rates, and where Global, Data Zone and Regional deployments process data.

For: Teams pricing or procuring Azure OpenAI capacity for production workloads

Last verified How we verify

Azure OpenAI offers two ways to pay for the same models: by the token on a Standard deployment, or by reserved capacity on a provisioned deployment measured in provisioned throughput units (PTUs). This guide explains the difference using Microsoft’s own documentation and price list, all checked on 22 October 2026. The vendor profile is at Azure OpenAI Provisioned Throughput.

What a PTU is

Microsoft describes PTUs as generic units of model processing capacity. When you create a provisioned deployment you choose how many PTUs to allocate; Microsoft reserves that capacity and charges for it hourly whether or not the deployment is handling requests. Each model has a minimum PTU count for a deployment.

Provisioned throughput vs pay-as-you-go

Microsoft’s comparison table sets the two side by side:

StandardProvisioned
BillingPay per tokenPer PTU per hour, or Azure Reservations
Latency SLANoneDefined latency target per model
Fit (Microsoft’s wording)Variable or unpredictable trafficGuaranteed throughput and consistent latency

Microsoft recommends provisioned throughput for predictable traffic, latency-sensitive applications, production-scale volume where per-token billing becomes expensive, and interactive scenarios such as chat and agents. Standard remains the better fit for development, testing, low volume and highly variable traffic.

What it costs at published rates

Microsoft’s pricing page lists per-PTU rates by model and deployment type. It labels them estimates in US dollars. For GPT-5.5 on 22 October 2026:

GPT-5.5 deploymentMinimum PTUsPer PTU per hourMonthly reservation per PTUYearly reservation per PTU
Global15$1$260$2,652
Data Zone15$1.10$286$2,916
Regional50$2$286$2,916

On the same page, GPT-5.5 on a Standard Global deployment cost $5 per million input tokens and $30 per million output tokens.

Worked from those figures: the smallest Global deployment (15 PTUs) costs $15 an hour on hourly billing, about $10,950 over a 730-hour month, or $3,900 a month on a monthly reservation. At Standard rates, $15 buys 3 million input tokens or 500,000 output tokens. Whether 15 PTUs serve more than that in an hour depends on the model’s throughput per PTU and your input-to-output mix, which is not verified here; size it with your own traffic before committing.

Hourly billing or a reservation

Provisioned deployments bill hourly, prorated (a deployment that exists for 15 minutes is charged a quarter of the hourly rate). Azure Reservations give a discounted effective rate in exchange for a 1-month or 1-year term. Two details matter in procurement. A reservation is a financial discount applied to PTU billing meters, not a capacity guarantee, and deployed PTUs beyond the reservation quantity are billed at the hourly rate. Microsoft also warns that unused quota does not guarantee capacity will be available when you scale back up.

Global, Data Zone or Regional

The deployment type decides where prompts are processed:

  • Global: may be processed in any Azure region.
  • Data Zone: processed only within a Microsoft-specified data zone: US, EU or Asia Pacific.
  • Regional: processed within the Azure geography you specify, possibly across regions inside it.

The table above shows the trade: Regional processing carries a higher hourly rate and a higher minimum than Global. Where data is processed is a residency question; it does not change the jurisdiction of the parent company. See the methodology.

Spillover

Microsoft documents spillover, an optional setting that routes overflow requests from a fully used provisioned deployment to a Standard deployment in the same resource, so peaks are billed per token rather than rejected. It is available for Azure OpenAI models that support provisioned throughput, not for other providers’ models in Foundry.

How PTUs compare elsewhere

Other providers reserve capacity in different units and terms; see provisioned throughput vs pay-per-token. Side by side: Bedrock vs Azure, Azure vs Vertex AI, Azure vs Together AI (which also sells a unit it calls a PTU) and Azure vs Modal.

Sources (4), each checked on the date shown