3.1.2. Capacity Planning: PTU, TPM Quota, and the Quota-vs-Capacity Distinction
💡 First Principle: Two different sizing units exist for two different worlds. TPM (tokens-per-minute) quota governs standard deployments and is granted per subscription and region, then allocated per deployment. PTU (provisioned throughput units) governs provisioned deployments, is model-independent, and is purchased as reserved capacity. And critically, quota is not capacity — having quota means you're allowed to deploy, not that the hardware is guaranteed.
For standard deployments, you allocate TPM quota across deployments in a subscription/region; one app monopolizing quota starves another, so you rebalance quota or split deployments. For provisioned, you acquire PTU quota per subscription/region, then deploy, then buy a matching reservation for the discount. PTUs are normalized units — the same PTUs can back different supported models — whereas TPM is tied to the specific deployment. The exam tests the subtle point that quota grants permission, deployment-time allocation grants capacity: you deploy first, and capacity is held only while the deployment exists.
| Concept | Applies to | Scope | Key trait |
|---|---|---|---|
| TPM quota | Standard deployments | Per deployment, per subscription/region | Tied to the deployment |
| PTU | Provisioned deployments | Per subscription/region | Model-independent, reserved hourly |
| Quota | Both | Subscription/region | Permission to deploy, not guaranteed capacity |
| Capacity | Both | At deployment time | Allocated and held while deployment exists |
⚠️ Exam Trap: "We have PTU quota, so capacity is guaranteed." No — quota is a ceiling on what you may deploy, not a reservation of hardware. A scenario where a deployment fails despite available quota is the capacity-vs-quota distinction in action; the fix is to deploy (and reserve) in a region with actual capacity, not to request more quota.
Reflection Question: Why can two teams both hold PTU quota in the same region yet have one fail to deploy, and what does that reveal about what quota actually promises?
