BUILD 2026: What Matters After the Keynote

Yesterday's BUILD keynote ran long. They always do. There's a rhythm to these things — opening segment on vision and numbers, followed by a demo showcase that always works perfectly in the room, followed by product announcements staged for press coverage, followed by the breakout sessions where the actual engineers tell you what's true.

I watched the keynote, skipped most of the morning demos, and spent the afternoon in the technical sessions. Here's the separation.

The Parts That Were Marketing

The Copilot integration announcements were presented as breakthroughs. Most of them were incremental extensions of features already announced at Ignite. The demos were polished. The underlying capability — AI assistance embedded in Microsoft 365 applications — is real, but the pace of the demos suggests production readiness that the enterprise rollout timeline doesn't support.

This is not a criticism of the technology. It's an observation about the gap between keynote time and shipping time. Microsoft has consistently announced features at BUILD that show up in GA six to eighteen months later. That's fine and expected. Price accordingly.

The Parts That Were Real

Azure AI Foundry. The model selection and deployment tooling got meaningfully better. The ability to swap model providers within the same deployment infrastructure — different model for different tasks, same integration layer — is genuinely useful. If you're building production AI systems on Azure, this is worth evaluating. The vendor lock-in risk on model providers is real, and anything that abstracts it is worth the integration cost.

Phi-4 reasoning. Microsoft's small model family got a reasoning-focused variant. The benchmark numbers they showed were good. Small models with strong reasoning are interesting for latency-sensitive and cost-sensitive applications — the use cases where you can't afford a frontier model but need more than a basic language model can do. Worth testing on your actual workload before drawing conclusions.

Semantic Kernel updates. The orchestration framework for multi-step AI workflows got significant updates. Specifically around memory management and multi-agent coordination. The framework is opinionated in ways that will fit some architectures better than others. If you're already bought into the .NET ecosystem and building agentic applications, the new version is worth a close look.

The Infrastructure Disclosure

One thing from the technical sessions: Azure disclosed more about the networking architecture for large-scale AI deployments than I've seen publicly before. Specifically around the InfiniBand fabric design for Blackwell clusters and how they're thinking about workload scheduling at the rack level.

This is the kind of detail that matters if you're designing for very high utilization or very low latency at scale. The public cloud AI infrastructure story has been underspecified relative to what's actually deployed. The fact that they're talking about it is either because customers need it for architecture decisions, or because they're trying to close the gap with the colocation providers who've been more transparent. Probably both.

What to Actually Watch

The real test for everything announced yesterday is production rollout velocity over the next six months. Azure AI Foundry is available. Semantic Kernel 2.0 ships this week. Phi-4 reasoning is in preview.

Preview to GA timelines on Azure have historically varied widely. Some things land in six weeks. Others sit in preview for a year. The signal on which category something falls into is usually in the enterprise customer feedback loop, which you won't see publicly.

Watch the GitHub repo activity on the open-source components. That's a leading indicator. And watch the documentation — when Microsoft starts investing heavily in migration guides and cookbook-style tutorials for a product, that's usually the signal that GA is close and they're building for real adoption, not just press.

One Observation

BUILD is best understood as a market signal, not a shipping schedule. The announcements tell you where Microsoft thinks the market is going and what they're betting on. The actual delivery of that bet is a different story, with a different timeline, that you have to evaluate separately.

The bet being placed is clear: Azure as the default AI infrastructure layer for enterprise, with Microsoft's application layer as the primary demand generator. Reasonable people can disagree about whether that bet wins. But it's coherent, and the capital behind it is real.

Keynotes sell the future. Ship schedules tell the truth.

— Dustin