This week in Azure

A quiet week on announcements, a loud one on numbers. Azure Firewall Premium’s published throughput with TLS inspection and IDPS in deny mode is now up to 22 Gbps, and the single-connection ceiling is 600 Mbps. If you inspect egress, those two figures decide your topology more than any feature on this list.

The other item worth your time is Markdown for Agents in App Service: your web app returns markdown instead of HTML when an agent asks for it, and Microsoft’s own testing puts the median response at 97 percent smaller. That is a token bill, not a nicety.

Three items from John’s update did not make it into this post. Two I could not verify against a Microsoft source, and one I already covered last week — details at the end.

Networking

Azure Firewall Premium performance numbers

The Firewall performance table is the closest thing Azure gives you to a capacity plan, and it is worth reading in full rather than remembering “Premium does 100 Gbps”.

  • Premium, TLS and IDPS both disabled: 100 Gbps
  • Premium, TLS on, IDPS in alert-only mode: 100 Gbps
  • Premium, TLS on, IDPS in deny mode: up to 22 Gbps
  • Single TCP connection, IDPS in alert or deny mode: up to 600 Mbps
  • Premium out of the box, before autoscale: up to 18 Gbps

Read the third and fourth lines together. The moment you move IDPS from alert to deny — which is the entire point of running Premium in a regulated environment — your ceiling drops to 22 Gbps aggregate and 600 Mbps for any single flow. A nightly backup job or a large replication stream is a single flow.

The published figures assume a maximum autoscale instance count of 20. With prescaling you can reach 50 instances, and the numbers change. Autoscale kicks in at 60 percent average throughput or CPU, or 80 percent of connection capacity, and takes five to seven minutes to complete. Plan for that gap.

John’s update frames these as more than doubling the previous limits. The current numbers I can confirm; the size of the jump against your own baseline I cannot, so measure before you assume 2×. If you are designing egress inspection for a Swiss or EU regulated workload, the deny-mode row is your design constraint, and it belongs in the landing zone baseline rather than in a spreadsheet somebody finds later.

Azure Front Door: mutual TLS (preview) and batch rule updates (GA)

Mutual TLS is in preview on Azure Front Door Premium. Front Door authenticates the client certificate before the request reaches your origin, with revocation checking via OCSP and validation against allowed domain lists.

Client certificate authentication at the edge is what a lot of partner and B2B integrations actually need, and doing it at the origin means every origin implements it slightly differently. Moving it to the edge is the right layer.

Batch rule updates went generally available for Standard and Premium. Multiple rule changes apply as one atomic operation: all of them or none.

That matters more than it sounds. Replacing a rule usually means removing the old one and adding the new one. If the remove lands and the add does not, you are sitting in a gap with no rule at all — and on a WAF-adjacent rule set, that gap is a security incident. Atomic is the only sane default here.

Compute

Markdown for Agents in Azure App Service (preview)

App Service can now convert your app’s HTML response to markdown automatically. The client asks for it with an Accept: text/markdown header; nothing in your application changes.

The numbers from Microsoft’s internal testing across more than 637,000 pages: converted responses were 97 percent smaller at the median than the source HTML, with a median conversion time of 2 milliseconds. Scripts and styles are stripped; headings, paragraphs, links, lists, images, emphasis and code survive.

The constraints: Windows App Service only for now, Basic tier or higher, all public regions. Linux support is planned before general availability.

Ninety-seven percent smaller is a direct cut to what you pay per agent request, and it needs no application change. Your existing authentication, authorisation and network controls still apply, which is the part that makes this deployable rather than merely clever. If you run any HTML surface an agent reads, turn this on and measure your token spend before and after.

Observability

AKS control plane metrics — check the label before you plan on it

Worth being precise here, because the status depends on which metrics you mean.

Free platform metrics for the API server and etcd are collected automatically for every AKS cluster at no cost, queryable in metrics explorer, and usable for alerts. That part needs nothing from you.

The fuller control plane set through managed Prometheus — scheduler, controller manager, cluster autoscaler, node auto-provisioning — still carries the preview label in the documentation and still needs the AzureMonitorMetricsControlPlanePreview feature flag and the aks-preview CLI extension. Only apiserver and etcd are scraped by default; the rest are off until you edit the configmap. Azure Private Link is not supported, and self-hosted Prometheus cannot scrape these at all.

John’s update called this generally available. The docs I can read still gate it behind a preview feature flag. Check the status on your own subscription before you put it in a production monitoring design — a preview feature carries no SLA, and a control plane dashboard nobody supports is worse than none.

Database

PostgreSQL upgrade validation checks

Azure Database for PostgreSQL flexible server can now validate a major version upgrade without performing it. In the portal: Upgrade → target version → set Action to Validate only → Start. From the CLI, with Azure CLI 2.89.0 or later:

Terminal window
az postgres flexible-server upgrade \
  --resource-group <resource_group> \
  --name <server> \
  --version <target_version> \
  --validate-only

It does not restart the server and does not begin the upgrade. You get a pass/fail list with remediation guidance, a downloadable CSV, and pg_upgrade logs under Server logs — the logs need server logs enabled on the server first, so switch that on before you run the check rather than after.

Major version upgrades fail on the boring things: an unsupported extension, a leftover logical replication slot, a prepared transaction nobody closed. Finding those during a maintenance window is how a two-hour change becomes a rollback. Run the validation, fix the list, then schedule the window.

AI

MAI-Thinking-1 in Microsoft Foundry (preview)

Microsoft AI’s first reasoning model is in public preview in Microsoft Foundry. It is a sparse mixture-of-experts model with 35 billion active parameters out of roughly one trillion total, so only the experts a given request needs are activated.

The claim Microsoft leads with is provenance: trained end to end on Microsoft’s own infrastructure, on clean data, with no distillation from third-party frontier models. On benchmarks it lands toe to toe with Claude Opus 4.6 on SWE-Bench Pro, and reports 97.0 percent on AIME 2025 and 94.5 percent on AIME 2026.

For a Swiss or EU shop the provenance claim is the interesting one, not the benchmark. “No distilled third-party weights, trained on infrastructure we control” is a sentence you can actually put in front of a compliance function, and it is a different conversation from “we call a frontier API and hope”. Benchmarks move every quarter; a defensible training story does not.

It is preview, so it belongs in an evaluation, not in a production path. If you are wiring it into anything real, the content filters and the tool gateway matter more than the model choice — a preview model behind a governed gateway is a manageable risk, the same model with direct tool access is not.

What I left out

Three items from this week’s update are not above, and it is worth saying why rather than quietly dropping them.

  • Fabric item-level recovery on by default from August 23rd. The documentation I can reach still says item recovery is off by default with a 7-to-90 day retention range, which contradicts the three-day default described in the video. Unverified, so omitted.
  • Shared Premium SSD v2 and Ultra disk dynamic resize. Live resize for these disk types over NVMe controllers has been generally available since Q3 2025, and the current documentation still excludes shared disks. I could not confirm the shared-disk extension against a Microsoft source.
  • ExpressRoute resiliency guard. Already covered in CW32 — same preview, no change.

Final thoughts

Go and read the Firewall performance table against your actual egress design. Twenty-two Gbps aggregate and 600 Mbps per flow with IDPS in deny mode is the number that governs the topology, and “Premium does 100 Gbps” is the number most people are carrying around. If those two are far apart in your head, fix that this week.

Then turn on Markdown for Agents if you run a Windows App Service that an agent reads. It is a header, it needs no code change, and a 97 percent smaller median response goes straight to your token bill.

Everything else here is evaluation work, not action. MAI-Thinking-1 is preview, the AKS control plane metrics are still labelled preview whatever the video says, and the PostgreSQL validation check pays off the day you actually schedule an upgrade.

Sources

  1. John Savill, “Azure Update 14th August 2026,” YouTube, https://www.youtube.com/watch?v=KPNPsO-72Kw
  2. “Azure Firewall performance,” Microsoft Learn, https://learn.microsoft.com/azure/firewall/firewall-performance
  3. “Announcing public preview: Markdown for Agents in Azure App Service,” Microsoft Community Hub, https://techcommunity.microsoft.com/blog/appsonazureblog/announcing-public-preview-markdown-for-agents-in-azure-app-service/4537023
  4. “Mutual TLS authentication (Preview) in Azure Front Door,” Microsoft Learn, https://learn.microsoft.com/azure/frontdoor/mutual-tls
  5. “What is a rule set? Azure Front Door,” Microsoft Learn, https://learn.microsoft.com/azure/frontdoor/front-door-rules-engine
  6. “Monitor AKS control plane metrics,” Microsoft Learn, https://learn.microsoft.com/azure/aks/control-plane-metrics-monitor
  7. “Upgrade validation checks in Azure Database for PostgreSQL flexible server,” Microsoft Learn, https://learn.microsoft.com/azure/postgresql/configure-maintain/how-to-run-upgrade-validation-checks
  8. “Introducing MAI-Thinking-1,” Microsoft AI, https://microsoft.ai/news/introducing-mai-thinking-1/
  9. “MAI-Thinking-1,” Microsoft Foundry model catalog, https://ai.azure.com/catalog/models/MAI-Thinking-1