Azure AI Foundry
The services_ai_foundry role deploys an Azure AI Foundry account (a dedicated
resource group, virtual network and private endpoint) with one or more model
deployments. This page covers choosing models, controlling spend (the cost ceiling and the
key-rotation kill-switch), how output-token pricing is resolved, and change-detection alerting.
Finding deployable models
Each entry under services_ai_foundry.models must name a model, version and SKU that
Azure actually offers in your deployment region. Rather than guess, list what is deployable
with the Azure CLI and copy the values into your configuration.
List deployable models for a region
$ az cognitiveservices model list \
--location westeurope \
--query "[].{model:model.name, version:model.version, skus:join(',', model.skus[].name)}" \
-o table
The model column is the value for model:, version for version: and any of the
listed skus for sku: (commonly GlobalStandard, DataZoneStandard or Standard).
Prefer the current/default version. A version that has been retired for new deployments can still appear in this list, so a listed version is necessary but not always sufficient; if a deployment is rejected at apply time, pick the newest version the command reports.
Check what an existing account can deploy
Once an account exists, this shows exactly what that account can deploy in its region, which is the authoritative check:
$ az cognitiveservices account list-models \
-g <resource-group> -n <account-name> -o table
Example configuration
services_ai_foundry:
enabled: true
cost_ceiling_usd: 100 # monthly USD ceiling for the whole account
alert_at_percent: 80 # email heads-up
kill_at_percent: 90 # rotate the keys (hard stop)
guardian_interval_minutes: 1 # how often the kill-switch checks spend
models:
- name: gpt-4o
model: gpt-4o
version: "2024-11-20"
sku: GlobalStandard
- name: claude-sonnet-4
format: Anthropic
model: claude-sonnet-4
sku: GlobalStandard
# Not in the retail catalogue, so give the output-token price explicitly.
price_per_1k_tokens_usd: 0.015
Controlling spend: the cost ceiling and kill-switch
Spend is governed by a single number, cost_ceiling_usd — a monthly ceiling in USD for the
whole account. Two percentages act on it:
alert_at_percent(default 80)Emails the users tagged
ai_alertinusers.csv. A heads-up only, no enforcement.kill_at_percent(default 90)The hard stop. When month-to-date spend reaches this percentage the account’s API keys are regenerated, which immediately cuts off every agent still using the old key. Logs and data are left intact — it is a hard stop, not a teardown.
The 10 % gap between the alert and the kill, and the further gap up to the 100 % backstop below, is deliberate head-room so a lagging signal still stops spend before it runs away.
How the kill is enforced
Two independent mechanisms, by design:
- Near-real-time (primary)
A per-account guardian Logic App runs every
guardian_interval_minutesminutes (default 1). Using a shared read-only managed identity it reads each model deployment’s output tokens from the metrics API, multiplies by that model’s output-token price, sums the account total, and once it reacheskill_at_percentof the ceiling it triggers the pipeline that rotates the keys. Because it computes exact per-model cost it reacts within one interval of a real breach; any read error fails the run safely without firing, and the agent only spins up on an actual breach. Raiseguardian_interval_minutesto cut Logic App run cost in exchange for a slower reaction.- Budget backstop (fallback)
An Azure cost budget on the same ceiling posts the same kill at 100 % of the ceiling, and emails the
ai_alertusers at the same time. It is price-independent — it reads real billed USD, not token metrics — so it catches anything the token math misses: non-text models, a price the matcher could not resolve, or a stalled Logic App. It lags by hours, which is why it sits above the near-real-time kill rather than replacing it.
Why it doesn’t re-fire every minute
Month-to-date spend stays above the threshold after a kill, so a naive check would re-trigger
every minute and spin up an agent each time. Instead the guardian stamps a guardianResetAt
timestamp tag on the account when it fires, and the Logic App only counts tokens since
max(start-of-month, guardianResetAt). Rotating the keys stops new usage, so that window
stays empty and the switch goes quiet on its own. The tag is excluded from Terraform’s state
(lifecycle.ignore_changes) so a later deploy does not wipe it.
Recovering after a kill
Recovery is deliberate and manual: an operator regenerates (re-issues) the account keys and
hands the new key to the consumers. The reset timestamp is already in place, so the counter
simply resumes from the kill point — the Logic App needs no attention. A deliberate re-issue can
therefore spend up to another ceiling’s worth in the same month; permanently raising the cap
instead means editing cost_ceiling_usd and redeploying (or changing it in Azure by hand).
That friction is intentional — nobody bumps the ceiling by reflex.
Output-token pricing
The kill-switch turns token counts into USD, so it needs each model’s output-token price. That
price is pulled automatically from the Azure Retail Prices API
using the model and version you configured, and it resolves for most current OpenAI,
Meta Llama, Microsoft Phi and xAI Grok models.
Some models cannot be priced automatically: a few are billed under a name that differs from
their deployment id (for example DeepSeek), and some are not published in the retail catalog
at all (for example Anthropic Claude). In those cases the deploy stops with a clear error
naming the model, and you set price_per_1k_tokens_usd on that model to continue. Read the
output-token price from the pricing pages — Azure OpenAI
and Microsoft Foundry models —
which quote per 1,000,000 tokens; divide by 1000 for the per-1,000-token value this field expects.
Change-detection alerting
The role also adds an activity-log alert that emails the ai_alert users on administrative
changes to the Foundry account and its model deployments — a new or changed deployment, a
public-access or network-ACL change, or a delete. It is scoped to the account, not the whole
resource group, so the routine networking, private-endpoint and budget writes a deploy makes do
not generate mail. The guardian’s own key rotation is an account write, so a kill shows up here
too — the intended audit trail of an enforcement action.