Microsoft released KB5096136 this week, a small automatic update for Windows 11 version 26H1 that bumps the AMD Vitis AI Execution Provider from version 2.2604.1.0 to 2.2605.2.0. While the change barely registers in Windows Update history, it signals a significant shift: Microsoft now treats the plumbing that accelerates AI on PC hardware as an integral part of the operating system, not as an app or driver left for users to find and install on their own.

A Routine Update with Outsized Implications

On the surface, KB5096136 is almost boring. It arrives automatically through Windows Update, requires the latest cumulative update for Windows 11 26H1, and appears in update history as “Windows Runtime ML AMD NPU Execution Provider Update.” The package replaces April’s KB5089175, which had delivered version 2.2604.1.0. Microsoft’s support note describes it simply: “includes improvements to the AMD Vitis AI Execution Provider AI component.”

But the brevity hides a deliberate strategy. By servicing this component through Windows Update, Microsoft ensures that any PC with an AMD Ryzen AI NPU running the 26H1 release will have a consistent, up-to-date acceleration layer for AI workloads. This matters because Windows 11’s next competitive frontier isn’t about chatbots or flashy Copilot features—it’s about making local AI inference reliable and invisible.

What’s Inside KB5096136

The update targets the AMD Vitis AI Execution Provider, a piece of software that sits between Windows Machine Learning (Windows ML) or the ONNX Runtime and the physical neural processing unit (NPU) inside AMD’s latest mobile processors. When an app uses Windows ML to run an ONNX model, the execution provider decides whether that model can run on the NPU, GPU, or has to fall back to the CPU. If the NPU is used, the provider translates machine learning operators into instructions that AMD’s hardware can execute efficiently.

Version 2.2605.2.0 is an incremental improvement over the previous 2.2604.1.0, but Microsoft hasn’t detailed exactly what changed. In practice, these updates typically refine operator support, improve performance for certain model architectures, fix bugs, or adjust power behavior. Because the provider is now part of the Windows servicing stack, these tweaks reach devices without requiring users to hunt down AMD’s separate Vitis AI developer package.

You can verify the installation under Settings > Windows Update > Update history; look for the entry labeled “Windows Runtime ML AMD NPU Execution Provider Update.” If you’re running Windows 11 26H1 on an AMD Ryzen AI PC and have the latest cumulative update, the package should already be present.

Who Benefits (and Who Doesn’t)

Everyday Windows users with an eligible AMD-based laptop or desktop may not notice any immediate change. The update doesn’t add a new app or toggle. However, if you use Windows features that rely on local AI—like Windows Studio Effects for video calls, background blur, eye contact correction, or voice focus—the updated execution provider could improve responsiveness or reduce battery drain. Over time, as more apps adopt Windows ML for on-device inference, having an up-to-date provider will ensure broader compatibility.

Power users and developers stand to gain more. If you’re testing ONNX models on an AMD Ryzen AI machine, the new provider might unlock operators that previously forced a CPU fallback, or it could speed up inference for specific vision or language models. It’s not a magic bullet—models still need proper quantization and architecture support—but it removes one more variable. Developers no longer have to bundle a specific version of the Vitis AI EP with their app; they can rely on the system-managed component.

IT administrators face the dual reality of convenience and control. The update is automatic, so there’s nothing to deploy, but that also means changes to AI behavior can arrive silently. In tightly regulated environments where app behavior must be consistent, admins might want to track which execution provider versions are installed across the fleet. For now, because 26H1 is a restricted branch (mostly on new Copilot+ PCs and other latest devices), the immediate impact is narrow, but as this servicing model expands, admins will need to treat these components like graphics drivers—something to validate before rollout.

The Race to Make AI Invisible

Windows has a long history of struggling with hardware acceleration. In the past, making a GPU run a game or a video encoder work meant juggling vendor-specific drivers, SDKs, and API versions. With the rise of AI PCs, the same fragmentation threatened to repeat. NPUs from Qualcomm, Intel, and AMD each had their own toolchains, and developers faced the prospect of supporting every silicon vendor separately.

Microsoft’s answer has been to build a common abstraction through Windows ML and the ONNX Runtime. The idea is straightforward: developers bring their models in ONNX format, and Windows ML figures out which hardware can best execute them. To make this work, hardware vendors supply “execution providers” that translate generic ONNX operations into silicon-specific instructions. These providers are not applications; they are plug-in components that the runtime loads as needed.

AMD’s Vitis AI has long been the development stack for its adaptable SoCs and data center cards, but in the PC space, it became the bridge for Ryzen AI processors. By early 2026, AMD and Microsoft started shipping the execution provider as part of Windows, initially through a separate installer and then, with KB5089175 in April, as an automatic update. KB5096136 continues that cadence, moving the provider to a monthly-like update cycle.

This is exactly how Microsoft mainstreams a technology. By rendering it invisible—just another update that arrives without fanfare—the company turns AI acceleration from a specialty feature into platform plumbing. The goal is that someday, when you open an AI-powered feature, it just works, regardless of whose NPU is inside your laptop.

What to Do Now

If you have a Windows 11 26H1 device with an AMD Ryzen AI processor, there’s nothing you need to do. Open Windows Update, make sure the latest cumulative update is installed, and then check update history to confirm that the “Windows Runtime ML AMD NPU Execution Provider Update” is listed. If it’s missing, you might be missing the prerequisite cumulative update, or your device might not be eligible. Remember, 26H1 is a specific release branch; most Windows 11 users are on 24H2 or 25H2, where this update does not apply.

For developers building AI experiences with Windows ML, test your applications against the new provider version. Pay attention to which models run on the NPU versus falling back to CPU. Use tools like the ONNX Runtime profiling API to verify operator coverage. If you previously maintained a separate download of the Vitis AI EP, consider transitioning to the system-managed version to simplify deployment and ensure users always get the latest bug fixes.

IT teams that supervise Windows 11 26H1 deployments should add execution provider versions to their change management documentation. While the immediate risk of regression is low, subtle changes in AI behavior could affect user-facing features. For example, if a video conferencing app suddenly uses the NPU for background blur after the update, battery life might change. Knowing which components updated and when will help in support scenarios.

Outlook: More AI Updates on the Horizon

KB5096136 is unlikely to be the last such update. Microsoft’s steady march toward treating AI acceleration as a first-class system component means we can expect similar packages for Intel’s NPU (via the OpenVINO execution provider) and for Qualcomm’s Hexagon NPU. The cadence may align with monthly cumulative updates, turning the AI stack into a living part of the operating system.

For users, the promise is less friction and better performance from the growing number of AI-enabled apps. For Microsoft, it’s a way to ensure that Windows remains a competitive platform for local inference—a battle that will define the next generation of PCs. The real test will come when AI features spread beyond early adopter 26H1 devices to the broad install base. If Microsoft can make that transition as boring and reliable as the update we see today, it will have succeeded in turning AI into just another thing Windows does.