Azure OpenAI

What Azure actually meters

Three of the four deployment types meter tokens in the familiar way, separated only by rate: standard, a priority tier that costs more and carries a latency target, and batch that costs less and returns results asynchronously.

The fourth meters capacity. A Provisioned Throughput Unit is a slice of dedicated model processing throughput held exclusively for your deployment. Microsoft is explicit that a provisioned deployment holds that capacity whether or not requests are being made, and bills at an hourly rate per PTU deployed regardless of the number of tokens consumed. The meter starts when the deployment is created and stops when it is deleted.

Each model publishes its own PTU-to-tokens-per-minute ratio and its own minimum PTU count, so the smallest viable deployment differs by model. PTU quota is a separate concept from capacity: quota is a policy ceiling on how many PTUs you may deploy per subscription, per region and per deployment type, and it carries no cost of its own. Having quota does not guarantee capacity is available in the region.

How reservations work

Provisioned deployments bill two ways. Hourly is the flexible mode, useful for benchmarking a model or scaling up for a short event. Microsoft's own documentation advises against treating hourly as a scaling strategy, for two stated reasons: capacity may not be available when you try to scale back up, and sustained hourly billing at high utilisation typically costs more than a reservation.

Reservations commit to one month or one year and bill matching usage at a discounted rate instead of the hourly one. Three constraints shape how they are bought:

• Reservations are not interchangeable across deployment types. Global, Data Zone and Regional each need their own, and a Global reservation will not cover Data Zone usage.

• Global reservations are not region-specific. One Global reservation can cover deployments across several regions provided you reserved enough units.

• Exchanges reset the term. You can swap region, deployment type, term or payment option, but the clock restarts.

Payment runs upfront or monthly.

What happens when you hit the limit

Nothing runs out, because nothing is allocated. Standard and batch deployments bill every token at list rate with no included allowance to exceed. Provisioned deployments have the opposite problem: the capacity is fixed, so exceeding it means requests queue or fail rather than costing more.

The real constraint is quota. If you need more PTUs than your subscription allows in a region, you request a quota increase, and separately confirm the region actually has capacity to give.