There's a particular vertigo to stepping away from this field for a few months and coming back to find the roadmap rearranged. While I was off being deliberately unreachable, Microsoft held the two events that set the tone for its whole data year — the Fabric Community Conference in Atlanta in March, and Build in late May — and between them they didn't just ship features, they changed the story. I've now done the thing I'd been putting off since I got back: read six months of announcements in a weekend, with coffee and a healthy suspicion. This is the belated recap, and being late is rather the point. Half the noise from a launch has evaporated by the time you read it cold, months later, which makes it much easier to see what actually mattered. So here's my assessment of what I missed — not a news dump, which every blog already did in real time, but the view from someone reading it after the hype had time to settle.
The word of the year is "agentic," and I have feelings
You could not read the Build coverage without tripping over the word agentic every third sentence. The headline framing was "the agentic era," with Fabric positioned as the data backbone underneath it, and the flagship of that story is agentic analytics — AI that builds and updates your data environment from plain-language instructions. The most concrete expression is Agent Skills for Power BI: an assistant that can generate semantic models, build report pages and layouts, produce the underlying report files, and even publish, all from a description of what you want. Point at your data, say what you need, and the machine assembles the analytics.
Let me be fair before I'm sceptical, because there's something genuinely real here. The tedium of standing up a semantic model and a first set of reports is exactly the kind of mechanical work that an assistant should absorb, and if it works even half as smoothly as the demos, it saves real time. But here is the belated, dust-settled version of my reaction: an agent that builds your data environment is only as trustworthy as the data and the definitions underneath it. An AI that confidently auto-generates a semantic model on top of an ungoverned mess produces a fast, fluent, authoritative wrong answer — and now it does it without a human even pausing to notice. The agentic era doesn't retire the boring governance work. It raises the stakes on it, because the human who used to catch the dodgy join is being designed out of the loop. I'll come back to this, because it's the through-line of everything I missed.
The one that's cleverer than it first looks: Microsoft IQ
Buried under the agentic headlines was the announcement I think is structurally the most important, and it took reading it late to see why. Microsoft IQ is an umbrella over several "intelligence" layers — Work IQ over Microsoft 365 activity, Foundry IQ for enterprise knowledge retrieval, a new Web IQ for grounding agents in the live web, and, the one I care about, Fabric IQ: a shared semantic layer over your structured business data. Strip the branding and here's what it's really saying: for agents to be useful over enterprise data, they can't reason over raw tables — they need a governed layer that encodes what the business actually means. A "customer," a "region," a certified "revenue," with the definitions and relationships made explicit.
That is the opposite of hype. That is Microsoft quietly conceding that the semantic and governance groundwork is the thing that makes the agents work at all — that you cannot bolt intelligence onto chaos. Fabric IQ is a bet that the winners in the agentic era will be the organisations that did the unfashionable work of defining and governing their business semantics. I've been saying some version of that for years; it's oddly satisfying to watch it become a product tier.
The performance headline: GPU in the warehouse
The Build number that got people talking was GPU acceleration built into the Fabric Data Warehouse, via NVIDIA — Microsoft's claim being the first fully managed warehouse with GPU acceleration, and benchmarks of roughly three to seven times faster depending on concurrency. My belated take: impressive, and also, read the fine print. Those are Microsoft's own internal benchmarks against unnamed "comparable providers," which is exactly the genre of number that deserves a raised eyebrow until independent workloads confirm it. More usefully — GPU acceleration matters enormously for some workloads and not at all for others. If your warehouse pain is concurrency and heavy analytical queries, this could be genuinely transformative. If your pain is that your data is a mess and your refreshes are wasteful, a faster engine just lets you be wasteful more quickly. Speed is wonderful; it is not a substitute for the fundamentals, and no benchmark ever fixed a governance problem.
FabCon's quieter, and arguably better, story
Build got the agentic fireworks, but FabCon in March — merged this year with SQLCon under the banner of unifying databases and Fabric on one platform — carried the announcements I'd actually reach for first on a Monday. Direct Lake on OneLake went generally available, which means import-class query speed on lakehouse data with native security and no refresh, now a supported thing you can build production on rather than a preview to experiment with. The Govern tab gained admin insights at GA — real visibility into domains, capacity utilisation, workspace activity, and protection coverage, which is the unglamorous observability that keeps a Fabric estate from turning into a swamp. OneLake got workspace tags for discoverability, and shortcuts to SharePoint and OneDrive reached GA, so more of the estate connects without copying. There's even a Runtime 2.0 in preview on Spark 4.x and Delta 4.x for the heavy compute crowd.
None of that trended. All of it is the operational maturation that decides whether you actually enjoy running this platform. The pattern of the whole year, seen late, is clear: the loud announcements were about agents, and the quiet ones were about giving those agents a governed, observable, well-defined foundation to stand on.
Meanwhile, in the community threads
The official blog is the brochure. The community threads are the review, and I read those far more carefully — because the people posting in the practitioner corners of the internet are the ones who'll actually have to run this on a real, messy tenant on a Tuesday. Reading back through the reaction while I was away, the honest mood was a long way from the keynote's.
The first question under almost every agentic announcement wasn't "how clever" — it was "what's this going to cost?" And fairly: persistent, multi-step agent reasoning is compute-hungry, Microsoft was cagey on pricing at Build, and the community's lived experience is that Fabric capacity consumption already surprises people and throttling already bites. Bolting agents that think in loops on top of a CU model people struggle to predict is, understandably, making the folks who pay the bills nervous. Right behind cost came the platform-readiness point, and it's the one I'd underline: the whole "semantic glue" story assumes you've genuinely adopted Fabric and built a clean semantic layer — and an awful lot of the community, honestly, hasn't. "Great demo, but my tenant is a swamp" was the energy, and it's the correct energy. Then the governance worry, echoing the surveys showing most CIOs still put governance, security, and explainability at the top of their agentic-AI fears: an agent that auto-builds a model on ungoverned data is the practitioner's nightmare, not the marketing's dream. And underneath it all, the weary veteran's question — can Microsoft actually execute coherently across the enormous stack it already owns, or is this another wave of announcements that outruns the delivery?
I find that skepticism reassuring, not cynical. It's the sound of the people who run this stuff pointing at exactly the right things. The gap between the keynote and the forum is the gap between the demo and the Tuesday, and the forum is almost always closer to the truth.
What being late taught me
Here's what the belated vantage clarified, and it's the reason I'm glad I read all this cold instead of in the launch-week rush. Every headline of the year — the agents that build your models, the GPU warehouse, the IQ layers — points at the same conclusion, and it's the one I sketched the day I got back when I laid out what I was watching for the second half, only sharper now that I've seen the detail. The agentic era makes the boring fundamentals more decisive, not less. When an AI builds your semantic model, the quality of your governance becomes the quality of your AI. When agents reason over your data, your semantic definitions become the ceiling on how right they can be. Microsoft's own answer — Fabric IQ, the governance tooling, the semantic layer — is a tacit admission of exactly that.
So I missed two big events, and the thing I feared while I was away — that I'd come back to a landscape I no longer recognised — turned out to be unfounded. The vocabulary changed; the lesson didn't. The organisations that will get the most out of this genuinely remarkable wave of tooling are the ones who did the unfashionable groundwork, and the ones chasing the agentic demo on a rotten foundation are about to learn the same lesson everyone learns, only faster and more publicly this time. Caught up, and oddly reassured. Now let's see what Ignite does to all of this.