Microsoft dropped a small but significant update into Windows Update last week, and unless you check your update history obsessively, you probably missed it. KB5089175 delivers version 2.2604.1.0 of the AMD Vitis AI Execution Provider to Windows 11 version 26H1, and while it won’t add a shiny new button to your desktop, it does something far more consequential: it quietly strengthens the backbone of AI acceleration on AMD-powered PCs.

What Actually Changed Under the Hood

KB5089175 updates a single component: the AMD Vitis AI Execution Provider. This isn’t a driver or an app. It’s a module that plugs into ONNX Runtime and Windows Machine Learning (Windows ML), the common AI runtime layers Microsoft ships inside Windows. When an application hands an ONNX-format AI model to Windows ML, the execution provider steps in and figures out how to run the model’s operations on AMD hardware—typically the Neural Processing Unit (NPU) found in Ryzen AI-equipped laptops and desktops.

The new version, 2.2604.1.0, replaces the previous KB5079260. Microsoft’s support note is characteristically terse, but the update history label tells you what’s going on: “Windows Runtime ML AMD NPU Execution Provider Update.” That phrasing is deliberate. “Windows Runtime ML” signals a shared platform, and “AMD NPU Execution Provider” points to the hardware-specific acceleration path.

Only devices running Windows 11 version 26H1 receive this particular package. That’s not the current mainstream build for most users—Microsoft positions 26H1 as a specialized release for next-generation hardware platforms, not a broad feature update. Your PC must also have the latest cumulative update for 26H1 before KB5089175 will install.

What It Means for You—Broken Down by Use Case

For Everyday Windows Users with AMD Ryzen AI Hardware

You won’t see a new icon, a splash screen, or a performance counter. The update installs silently through Windows Update. After it lands, applications that rely on Windows ML or ONNX Runtime for AI workloads may run more efficiently on your AMD NPU. That can translate into snappier background blur in video calls, faster on-device image recognition in Photos, or lower power draw when your laptop is doing sustained AI work without a network connection.

But—and this is crucial—the update doesn’t instantly turbocharge every app. Software must explicitly use Windows ML or ONNX Runtime and hand off ONNX models that AMD’s execution provider can accelerate. If an app bundles its own runtime or uses a different AI framework, KB5089175 won’t change its behavior.

How do you know you even have it? Open Settings > Windows Update > Update history and look for “Windows Runtime ML AMD NPU Execution Provider Update (KB5089175).”

For Developers Shipping AI Apps on Windows

This update is a quiet vote of confidence in Microsoft’s framework-dependent deployment model. Instead of shipping every conceivable NPU, GPU, and CPU runtime with your app, you can target Windows ML and let Microsoft’s servicing stack keep the execution providers current. That shrinks installer sizes and reduces the nightmare of tracking vendor-specific library versions.

The trade-off? You’re now dependent on Microsoft’s update cadence. If KB5089175 introduces a regression, your users will encounter it before you do, and you’ll have fewer levers to fix it yourself. Test your workloads against this provider, especially if you rely on ONNX operators that an NPU parses differently from a CPU or GPU.

A practical checklist:

  • Validate ONNX operator coverage with the AMD Vitis AI Execution Provider
  • Quantize models appropriately for NPU execution
  • Verify that unsupported operations gracefully fall back to CPU
  • Measure first-run compilation time (it can be substantial for complex models)
  • Track which KB versions you’ve tested against

For IT Administrators and Enterprise Deployments

Most organizations won’t touch 26H1 for months. Microsoft recommends 24H2 and 25H2 for broad deployment. Still, the update is a harbinger: AI runtimes are becoming a standard servicing item, just like graphics drivers or .NET Framework patches. If your organization plans to pilot 26H1 devices, build an inventory process now. Track the Windows build, cumulative update level, AI component KB numbers, execution provider versions, and NPU driver revisions.

AI acceleration can reduce cloud dependency for sensitive workloads, but it also creates new governance questions. If an app processes local data through an ONNX model accelerated by a vendor NPU, who is responsible for output accuracy, telemetry, or model provenance? Start documenting the AI software supply chain early, because auditors will ask.

How We Got Here: A Brief History of AI Abstraction on Windows

Microsoft’s AI-on-Windows story didn’t start with Copilot+. It began with WinML APIs back in the Windows 10 era, then gained momentum with ONNX Runtime, an open-source, cross-platform inference engine. The missing piece was a clean way to let Windows automatically select the best hardware backend without requiring developers to code separate paths for Intel, AMD, Qualcomm, and NVIDIA.

Enter execution providers. In ONNX Runtime, an execution provider is a pluggable module that knows how to map ONNX operators onto a specific piece of silicon. Some are built by Microsoft (like the CPU provider), others by silicon vendors. AMD’s Vitis AI stack targets not just Ryzen AI NPUs but also AMD Adaptable SoCs and Alveo data-center cards. The Windows client execution provider is the piece that slots into Windows ML.

Earlier this year, Microsoft released KB5079260, the predecessor to KB5089175, also for 26H1. The rapid replacement suggests an aggressive refinement cycle—exactly what you’d expect when an OS starts treating AI inference the way it treats Wi-Fi or Bluetooth: a platform capability that must stay current and stable.

This shift matters because NPUs are finicky. They excel at sustained, low-power inferencing for a specific set of operators, but they choke on models that aren’t quantized correctly or that contain ops outside their coverage. A polished execution provider can hide much of that complexity, turning a “technically works” experience into a “just works” one.

What to Do Now—Actionable Steps

If you suspect your device should have KB5089175 but don’t see it:

  1. Confirm you’re actually running Windows 11 version 26H1. Type winver in the Start menu and check the version line. If it says 24H2 or 25H2, this update doesn’t apply to you—wait for the equivalent package on your branch.
  2. Install the latest cumulative update for 26H1. KB5089175 requires it as a prerequisite.
  3. Manually trigger a check for updates: Settings > Windows Update > Check for updates.
  4. Look in Update history under “Other updates” or “Driver updates.” The entry labeled “Windows Runtime ML AMD NPU Execution Provider Update (KB5089175)” confirms installation.

If you’re a developer:

  • Pull the ONNX Runtime release notes for version 1.18 (or later) to see which AMD Vitis AI provider version maps to which ORT build.
  • Test your model’s performance with the NPU as the primary device and force a CPU fallback to compare.
  • Use the ONNX Runtime logging API to diagnose which operators are being assigned to which provider.
  • Monitor AMD’s Ryzen AI developer hub for updated Vitis AI quantizer tools and operator support tables.

If you’re an IT admin:

  • Add the following fields to your device inventory: AI component KBs, execution provider versions, and NPU driver versions.
  • When evaluating 26H1 hardware, include AI workload testing in your standard application compatibility checks.
  • Document which line-of-business apps depend on Windows ML or ONNX Runtime, and plan for provider version drift across OS builds.

Outlook: What to Watch Next

KB5089175 won’t be the last execution provider update we see this year. Microsoft is likely building a pipeline where Intel OpenVINO, Qualcomm QNN, and NVIDIA TensorRT-RTX providers receive the same servicing treatment. The question is whether the support documentation matures along with the bits. Right now, the KB article tells you almost nothing about what actually changed in version 2.2604.1.0—no list of fixed operators, no performance delta, no known issues. Developers and enterprises need that detail, and the current opacity will become a friction point as AI workloads move into production.

Also keep an eye on the divergence between Windows servicing branches. If 24H2, 25H2, and 26H1 each carry different execution provider versions, developers will face a fragmented platform that undercuts the abstraction promise. Microsoft’s challenge is to keep the AI layer coherent without forcing every AI-capable PC onto a single annual release track.

The AI PC era will be won or lost not by TOPS numbers or marketing slides, but by the boring, behind-the-scenes plumbing that turns hardware promises into real-world reliability. KB5089175 is a small piece of that plumbing, and its silent arrival is a sign that Microsoft intends to treat it seriously.