Resources
The units, and what
they actually bill for
Three abstractions cause most of the confusion in AI and data platform cost. This page explains what each one charges for, the mechanics teams most often get wrong, and what changed in the check set recently.
The units
Three abstractions over one question
Provisioned throughput unit
A reservation of Foundry capacity billed by the hour whether or not it serves a request. Utilization counts input tokens plus your max_tokens setting — not the tokens the model returned — so a generous default sizes the commitment for capacity you never use.
Databricks unit
A normalised measure of compute consumption, priced differently by workload class. The same hardware running the same job costs roughly two to three times more on All-Purpose than on Jobs compute, which is why cluster choice is a bigger lever than instance size.
Fabric capacity unit
Capacity billed continuously against an F-SKU ceiling, with smoothing that hides short peaks and throttling that absorbs sustained ones. Pausing stops billing on pay-as-you-go and does not on reserved capacity — the same action, opposite outcomes.
Mechanics
The things that catch teams out
- A stopped VM still bills for its disks; a deallocated one does not bill for compute. The distinction is invisible in most cost views and costs real money
- Reservation scope decides which spend a commitment can absorb — a subscription-scoped reservation cannot cover the workload that moved next door
- Tags on a resource group are not inherited by the resources inside it unless policy enforces it, so container-level coverage overstates allocation
- A provider advisor recommends against a usage lookback, so it will happily recommend a commitment for a machine you deleted last week
- Prompt caching is not automatic: prefixes have to be stable and aligned to a breakpoint, and on current model families a cache write can itself carry a charge
Check set v2026.08
What changed recently
Every report states the version it ran under, so a later run can separate estate change from platform change.
Foundry geography premiums
Microsoft Foundry raised EU Data Zone pricing 9% above Global and non-US Regional deployments 7–16%, and launched an APAC Data Zone at a 20% premium, effective 1 September. Added AIF-019: deployments sitting in a premium geography with no data-residency requirement behind them.
Prompt cache breakpoints
GPT-5.6 and later families support explicit cache breakpoints on Standard deployments, and cache writes can now carry a charge. Revised AIF-014 to test prefix stability against the new breakpoint model rather than assuming caching is free and automatic.
Databricks Standard tier sunset
Azure blocked new Standard workspaces in April and auto-upgrades every remaining one to Premium on 1 October 2026, a rate increase north of 35%. Added DBX-015: Standard workspaces with no migration plan and no budget line for the change.
Fabric pause and reservation mechanics
Pausing a capacity only stops billing on pay-as-you-go; reserved capacity bills through the pause. Fabric still has no native capacity autoscale. Split FAB-009 into separate pay-as-you-go and reserved checks, because the correct action is opposite in each case.
Want this against your own estate?
The first assessment is free and takes one subscription or workspace.