Build 2026 shows Microsoft understands enterprise IT better than most.
I watched the Microsoft Build keynote last week and came away with one clear thought: even in the age of AI, it’s not just the smartest model that wins, it’s understanding your customer and the systems they already rely on. Enterprise IT is Microsoft’s home turf and Build 2026 was another reminder of how well it understands the day-to-day realities of IT departments.
The real problem: inference is getting expensive
The more employees use AI at work, the faster organisations can burn cash on tokens, often without realising it because the cost is abstracted away from the individual user. Teams discover agentic coding, get excited, and can quickly end up using frontier-model tokens on tasks that a smaller open-source model could have handled for a fraction of the cost. In some corners of the industry, that overspend is even treated as a badge of honour!
The sensible response is to be model-agnostic and fit the model to the task. Reserve expensive frontier models for the genuinely hard tasks. Push the routine work onto cheaper models, and where it makes sense, onto local compute you've already paid for. None of this is novel thinking, but it's good to see Microsoft building the plumbing for it rather than just maximising Azure spend.
A couple of the Build announcements lean directly into that. Microsoft AI made a visible point of cost-efficiency. MAI-Thinking-1 is a 35B active-parameter reasoning model that, on Microsoft’s own benchmarks, looks competitive with Claude Opus 4.6 on SWE-Bench Pro while being much smaller. Independent validation still matters, but the direction of travel is clear: smaller, cheaper models that are good enough for a large share of real work.
Then there's the hardware.
The Surface RTX Spark Dev Box
The Dev Box is a compact Windows machine built around NVIDIA’s RTX Spark platform, with up to 128GB of unified memory and a clear pitch around local AI development. Microsoft and NVIDIA are positioning it for serious local inference, which is exactly where the economics are starting to get interesting.
The pitch is straightforward: work locally where possible and only reach for the cloud when you need frontier-scale models. The payback period will depend on developer spend and workflow, but token costs do not need to be especially high before this starts to look sensible for Windows-based AI development. It will be equally interesting to see how well Microsoft’s tools support local inference on macOS, given how much of that market Apple has already captured.
My main takeaway: the era of “just send it to the cloud” is quietly ending, at least for development and lighter inference. The economics increasingly stop supporting it by default.
Three releases that show Microsoft gets it
A few announcements stood out, less for the technology and more for what they reveal about how well Microsoft reads its own customers.
Microsoft Scout and the Autopilot category
Scout is Microsoft's first "Autopilot," an always-on agent that acts on your behalf rather than waiting to be prompted. It's built on OpenClaw, the open-source agent framework that went viral earlier this year.
Crucially, Scout is being positioned as an always-on agent that operates with its own identity and within existing enterprise controls. That matters more than the “autopilot” label. If Microsoft can make this work safely inside the Microsoft 365 governance boundary, it becomes much easier for large organisations to adopt.
What I'll be watching for is model fungibility, how freely Scout lets you swap in local compute or route through something like OpenRouter to manage cost. Either way it will be fascinating to see what norms are established and how broadly it ends up being used, as it will have real and unpredictable marginal costs.
The GitHub Copilot app
Claude Code, Codex and a handful of other agentic coding tools have converged on a familiar set of UX patterns for running multiple coding sessions at once. The new GitHub Copilot app clearly borrows from that direction, then adds a model picker and the one thing only GitHub can really offer: deep, native integration with GitHub itself.
If Microsoft keeps widening the model options, particularly toward local open-source models and local inference, this is where they could open a real gap on OpenAI and Anthropic; not by having the best model, but by letting you choose the best model for the job.
Rayfin: meeting the vibe coders where they are
Microsoft knows full well that people are vibe coding applications all over the place. The hard part was never building the prototype; it's getting that prototype into production securely and maintaining it.
To tackle this challenge, Microsoft has created Rayfin, an open-source SDK and CLI that lets you (or a coding agent) define an application backend in code and deploy it onto Microsoft Fabric. The data lands in Fabric in your own Microsoft tenant, under your existing governance controls.
Replit is the launch partner, but I expect more app-building tools to adopt Rayfin over time because it is open source. While Fabric-centric data architecture may still feel niche for smaller businesses, it shows Microsoft is willing to meet developers where they already are rather than insisting they come to Azure first.
Why the Microsoft stack still has the advantage
Strip away the announcements and the underlying logic is the same one that's driven Microsoft consolidation for years. Companies standardise on Microsoft not because every tool is best-of-breed (they aren’t!) but because a single stack removes integration complexity, the number of vendors to manage, and ultimately cost.
The frontier AI labs are building better brains and trying to win the application layer to crystallise customer relationships. Microsoft already owns much of that application layer, so it is building the thing enterprises buy; somewhere to run these systems safely, affordably and without adding more vendors to the stack.
Bottom line
The model race gets the headlines, but the boring basics of governance, cost control and meeting customers where they already work are still where enterprise IT is won. On that ground, Microsoft retains a home advantage and Build 2026 suggested it intends to keep it.
If you’re trying to make sure your team are well setup to take advantage of the benefits of AI, we can help. Reach out below.👇