DEV Community

Cover image for What's Actually Cooking in Azure AI Right Now (Notes from My MVP Prep Rabbit Hole)
Ibne sabid saikat
Ibne sabid saikat

Posted on

What's Actually Cooking in Azure AI Right Now (Notes from My MVP Prep Rabbit Hole)

So here's the thing — I've been buried in Azure documentation for the past few weeks, partly because I'm putting together my Microsoft MVP application, and partly because I genuinely can't keep up with how fast the Azure AI stack is moving anymore. Every time I close one tab, three new "What's New" pages pop up.
I figured instead of hoarding all these notes in a random OneNote page I'll never open again, I'd just dump them here. Consider this a "things I actually found interesting" list rather than a press release. No fluff, no "in today's rapidly evolving digital landscape" nonsense.

  1. Azure Cobalt 200 — the ARM chip everyone's quietly excited about If you're into infra and DevOps like I am, this one's worth a look. Microsoft pushed out early access preview for Azure Cobalt 200, their Arm-based VMs, and the pitch is a 50% performance bump over the previous generation — specifically tuned for agentic AI workloads running on Linux. Alongside it, there's a preview of Lasv5/Laosv5 VMs on AMD's EPYC "Turin" chips. Why does this matter beyond spec-sheet bragging rights? Because "agentic AI" workloads are genuinely different from your typical web app — lots of small, chatty calls, orchestration overhead, tool invocations back and forth. If Cobalt 200 delivers on that promise, it's a real cost story for anyone running agent pipelines at scale, not just a marketing slide. I haven't gotten my hands on one yet (still on the waitlist, story of my life), but if anyone reading this has access, genuinely curious how it holds up under real agent orchestration load.
  2. Foundry is turning into the actual control plane Azure AI Foundry keeps absorbing more of the AI lifecycle into itself, and two additions stood out to me:

Foundry IQ and Fabric IQ — basically making it less painful to wire different data sources and systems together for agent workflows. If you've ever tried to stitch together five different connectors just to get an agent to answer a question grounded in real business data, you'll appreciate this.
Foundry Local — running smaller, efficient models locally instead of always hitting a hosted endpoint. Paired with Azure Local, this is clearly aimed at scenarios where you can't (or don't want to) send everything to the cloud — think regulated industries, edge devices, or just people who are sick of latency and token bills.

Honestly, the local-model angle is the one I'm most interested in, because it's a good counter-narrative to the "just throw a bigger model at it" approach that's been dominating the conversation.

  1. Voice agents got a real upgrade Azure Speech quietly had a big moment. There's now Neural HD V3 voices in public preview with prompt-level instruction control (so you can actually tell the voice how to sound, not just what to say), plus MAI-Voice 2.0 with support for 10+ languages. On the speech-to-speech side, GPT-Realtime 1.5 and a new Azure-Realtime model are in public preview, aimed at natural multilingual voice output. There's also a proper Voice Live Evaluation Harness now, which lets you score voice agents against 13 different Foundry evaluators instead of just vibing your way through QA like most of us have been doing.
  2. Microsoft Discovery is now GA This one flew under the radar a bit, but Microsoft Discovery — their platform for building and governing agentic AI workflows — went generally available. If your org has been experimenting with agents in silos and now needs some actual governance around who's building what, this is the thing to look at.
  3. Certifications are shifting under our feet (again) For anyone else prepping for exams or MVP nomination alongside me — heads up. DP-100 (Azure Data Scientist Associate) officially retired, replaced by AI-300, the Machine Learning Operations Engineer Associate cert. There's also a new AB-100 (Agentic AI Business Solutions Architect) and AB-620 (AI Agent Builder Associate) rolling in to cover the agent-building skill gap. Basically: if your study plan was built around the old ML/data-science cert path, it's time to update the reading list. I had two chapters of DP-100 prep notes that are now, well, decorative.

Why I'm even writing this
Honestly? Writing things down publicly is how I actually retain them. Reading a "What's New" page and forgetting it by dinner isn't learning, it's scrolling with extra steps. If this saves someone an hour of digging through release notes, even better.
If you're also deep in Azure/Cloud/DevOps land right now, drop what you've been poking at in the comments — I'll probably end up reading about it at 1 AM anyway.

Top comments (0)