Microsoft’s Ignite announcements focus on infrastructure, databases and operational tools for running AI agents, while leaving key pricing questions unanswered.
Microsoft used its Ignite conference, running Nov. 17–20 in San Francisco and online, to argue that production AI depends on more than capable models. Its Azure announcements span processors, databases, application backends and agent operations—a broad push relevant to software teams choosing where their workloads and data will live.
The headline infrastructure release is Azure Cobalt 200, Microsoft’s Arm-based virtual machine processor, now generally available. Microsoft says the VMs deliver up to 50% better CPU performance than Cobalt 100 and are planned for 24 Azure regions by year-end. New options include consistent-speed and boost modes, configurations for large in-memory databases, and storage-heavy machines with up to 23 terabytes of built-in storage.
The CPU focus reflects the work surrounding AI agents: coordinating tools and creating temporary workspaces can require substantial computing outside the model itself. Faster processors could improve those workloads, but Microsoft’s performance figure is its own claim, not a guarantee for every application. Buyers should test their software and compare costs against existing machines.
Azure HorizonDB, now generally available, is a fully managed, PostgreSQL-compatible database for mission-critical applications. It includes Entra authentication, private endpoints, encryption options, multi-zone replication and distributed storage. Preview features would let developers manage AI models alongside data, call AI functions from SQL and mirror database data into Microsoft Fabric for analysis. Those additions could reduce integration work for teams already using Microsoft’s cloud, but the AI features remain previews.
Microsoft also made Rayfin and Fabric apps generally available. They combine application building with databases, authentication, backend services and access to Fabric data. Microsoft reports that builders created more than 100,000 apps with Rayfin during its preview. That figure indicates adoption, not proven business value; small firms should consider portability, governance and ongoing operating costs before relying on the tools.
For teams building agents, Microsoft Foundry is adding a feedback loop that uses production traces and evaluations to identify recurring failures and test configuration changes. Multi-turn evaluation, production insights and Agent Optimizer are generally available. Human annotation, Managed Training and Programmed Training are in preview. The tools address a common gap between a convincing demo and reliable performance in everyday use. Their value will depend on the quality of teams’ tests and oversight.
Foundry now offers more than 11,000 models, according to Microsoft. Microsoft-Decision-1.1, now generally available, targets structured tasks such as routing and classification. The company reports a score of 68 on Artificial Analysis Decision Index v2.1 and says it is more than twice as fast as its predecessor. Microsoft positions it as a way to handle routine decisions so larger models can focus on harder problems. The announcement provides no price comparison, so developers will need to check rates and performance for their own workloads.
A preview will let customers deploy Foundry models and agents to their own Azure Kubernetes Service clusters, using existing infrastructure and policies. Microsoft says the option supports open-weight models and lets teams choose GPUs and regions. Preview tools for agent networking, isolation and extending clusters to customer servers add control over deployment, though the arrangement remains within Azure’s platform.
For builders, the practical move is to test the releases that address a current need: Cobalt 200 for CPU-bound services, HorizonDB for PostgreSQL workloads, or Foundry’s evaluation tools for agents already in production. Ignite expands the options, but does not settle the questions that matter to smaller operators: workload-specific prices, migration effort and how easily a system can move between providers. Independent testing remains essential.

