Microsoft just moved two headline models into production on Azure Foundry GPT-5.6 is GA, and Claude is GA as a hosted option running on NVIDIA GB300 (Blackwell Ultra) and rolled out a pair of features that change where platform teams put their critical workloads: Fireworks AI (public preview) for high-performance open-model inference, and external key management for Azure Key Vault Managed HSM (public preview). That combination is the most consequential part of this weeks Azure activity: model choice and model hosting are now infrastructure decisions with the same operational and security posture questions as your database or KMS.
Why this matters right now
GAing GPT-5.6 and offering Claude on GB300 isnt just marketing parity. It signals Foundry is being treated as a multi-vendor model-hosting platform with production SLAs a turnkey option enterprises will use to avoid operating GPU fleets. Foundry also added an Asia-Pacific data zone and production-ready hosted agent capabilities in the Foundry Agent Service. Taken together, those moves accelerate two platform trends: centralizing model hosting inside cloud provider control planes, and making agent-first workflows operationally simpler to run at scale.
If you run a platform team, two immediate implications follow:
- Model residency, performance and cost become first-order constraints. Choosing GPT-5.6 vs Claude is now a policy decision with cost and egress implications, not just an API choice. Expect to pin models in your internal catalog, route traffic, and budget for GB300-class capacity where latency or accuracy demands it.
- Agents as first-class runtime: Foundrys hosted agent capabilities push teams toward an agent-centric IDP approach. Agents simplify observability, remediation, and deployments at scale, but they also expand your attack surface and increase audit and governance requirements.
Fireworks AI: the open-model inference play
Fireworks AIs preview is a notable part of Foundry: a managed inference fabric for serving open models at scale with high performance. For teams that want to avoid vendor-model lock-in or run custom weights, Fireworks is a managed alternative to self-hosting on bare metal or building a bespoke GPU fleet.
That said, managed inference fabrics remain operationally complex: model versioning, automatic rollover, cold-start mitigation, and tenant/resource isolation still need platform-level controls. Treat Fireworks as an integrated service to be governed, not just a faster container runtime, or youll hit reliability and cost surprises.
The new trust boundary: Managed HSM external key management
Microsoft previewed external key management for Azure Key Vault Managed HSM allowing key material to be retained outside Managed HSMs while still integrating with Azure services that accept customer-managed keys. This is overdue and the right move for regulated workloads. Dont confuse "external" with "no metadata leakage": control-plane calls, key-usage telemetry, and attestations can still traverse provider services. External KMS removes one risk (custody of key material inside cloud HSMs) but surfaces others around latency, availability, and the operational complexity of cross-cloud or on-prem key managers.
Two practical, non-obvious points
- Treat model endpoints and KMS like databases: enforce RBAC, rotation, disaster recovery, and billing attribution. You need playbooks for model rollback, cold-start mitigation, and KMS failover.
- The announcements didnt include new AKS release notes, Azure DevOps changes, or additional cost-management tooling. If your roadmap counted on AKS or DevOps improvements this week, dont assume they exist yet.
Final take
Microsoft is consolidating the things enterprises worry about model hosting, agent runtimes, and key custody into its platform. Thats useful. Its also a bet that customers prefer a managed, opinionated stack over composing best-of-breed pieces themselves. Good decision for teams that want speed; dangerous for teams that havent updated threat models to include hosted models and external KMS behaviors.
If youre responsible for platform or security, update your runbooks: treat Foundry-hosted models and Managed HSM external keys as first-class infrastructure, and test your incident playbooks now. If you dont, youll find out how expensive the combination of high-cost GPU instances and a misconfigured KMS can be the hard way.
Related reading: see our take on the agent-first shift in platform engineering and the broader Claude availability story in "Platform Engineering Embraces the Agent-First Paradigm" and "Claude Opus 4.7 GA: Availability on Claude API, Amazon Bedrock, and Google Vertex AI."