Microsoft has released a behind-the-scenes update that quietly boosts the AI capabilities of Windows 11 PCs equipped with NVIDIA RTX graphics cards. KB5089168, now rolling out automatically through Windows Update, upgrades the NVIDIA TensorRT-RTX execution provider to version 2.2604.1.0 for systems running Windows 11 24H2 or the newer 25H2 releases. The update doesn’t add any flashy new features or interface changes—instead, it fine-tunes a critical piece of plumbing that helps apps run machine learning tasks faster on your local GPU.

The update replaces an earlier package, KB5083460, and requires the latest cumulative update for Windows 11 to be installed first. Once applied, it shows up in your Windows Update history under the label “Windows ML Runtime Nvidia TensorRT-RTX Execution Provider Update (KB5089168).” There’s no need to download drivers separately or flip any switches; eligible devices will receive it silently in the background.

A closer look at the update

KB5089168 is not a typical Windows patch. It doesn’t fix security vulnerabilities or squash user-facing bugs. Instead, it updates a component that sits at the intersection of Windows Machine Learning (Windows ML), ONNX Runtime, and NVIDIA’s RTX ecosystem. In plain terms, this execution provider acts like a translator: it takes AI models built in the standard ONNX format and converts them into optimized instructions that can run directly on the Tensor Cores inside NVIDIA RTX GPUs.

Microsoft describes the update as containing “improvements to the execution provider component,” though it hasn’t published a detailed changelog. The new version number—2.2604.1.0—is the clearest signal that something has changed under the hood. The package applies exclusively to Windows 11 24H2 and 25H2, which share a modern platform foundation that supports this kind of modular AI servicing. If your device is on an older Windows release, you won’t see this update at all.

One key detail: the update is gated behind the latest cumulative update. That means if your PC has pending Windows updates, KB5089168 won’t appear until those are installed. Microsoft has increasingly tied AI infrastructure components to the cumulative update baseline, ensuring that the runtime environment is consistent before dropping in a new acceleration layer.

How this affects you

For most people, the impact will be invisible—and that’s by design. KB5089168 isn’t about delivering a new app feature; it’s about making existing AI-powered tasks potentially faster, more responsive, or more power-efficient. If you regularly use applications that leverage Windows ML or ONNX Runtime for local inference, you might notice snappier performance after the update installs. But it’s not a gaming or graphics driver, so don’t expect higher frame rates in your favorite titles.

Everyday users are most likely to benefit indirectly. Photo editors that use AI for background removal, video conferencing tools with real-time background blur, or creative apps that offer AI-assisted denoising or upscaling could see improvements in processing time. Even some voice assistants and transcription tools lean on local AI models—these might respond more quickly. The catch: the app itself must be written to take advantage of the TensorRT-RTX execution provider. Many popular applications still rely on generic GPU paths or cloud inference, but the number of RTX-optimized apps is growing.

Content creators and power users with RTX 30-series (or newer) GPUs stand to gain the most. Workflows like AI-powered image generation, video effect rendering, or 3D asset processing often run complex models that can clog a CPU or even a standard GPU path. The TensorRT-RTX provider is built specifically for client PCs, emphasizing fast startup, low latency, and efficient use of GPU memory. If you’re using an app like Adobe Photoshop’s Neural Filters, Topaz Video AI, or DaVinci Resolve’s AI features, KB5089168 could shave precious seconds off each operation.

Developers get a different kind of benefit: the update reinforces Microsoft’s model of keeping execution providers as system-managed components. Instead of bundling large vendor-specific libraries into their app packages, developers can rely on Windows Update to deliver and update the acceleration backends. This reduces app bloat and simplifies deployment, but it also means developers must design their software to gracefully handle cases where the provider is missing or outdated. Apps should check for available execution providers at runtime and fall back to CPU or DirectML paths if necessary.

IT administrators and enterprise managers might view the update with cautious interest. On one hand, it simplifies fleet management—AI acceleration components are now serviced through familiar update channels. On the other hand, it introduces a new software dependency that can be blocked by update policies, deferrals, or missing cumulative updates. If a line-of-business application relies on a specific inference path, IT teams need to ensure the execution provider is present and functional across managed devices. Checking update history becomes a new step in troubleshooting AI-related performance issues.

The bigger picture: how Windows became an AI platform

KB5089168 didn’t come from nowhere. It’s the latest milestone in a quiet transformation that has turned Windows 11 into a modular AI runtime platform. To understand why an execution provider matters, it helps to know how Windows ML and ONNX Runtime work.

Windows ML is a set of APIs that let applications evaluate machine learning models locally. ONNX Runtime is the cross-platform inference engine that often sits underneath, handling the actual number crunching. The clever part is that ONNX Runtime uses execution providers to split model computation across different hardware backends. An app loads an ONNX model, and the runtime decides which provider—CPU, DirectML, CUDA, TensorRT, or an NPU plugin—should run which parts of the graph.

Historically, Windows leaned heavily on DirectML, a hardware-agnostic acceleration layer that works across GPUs from NVIDIA, AMD, and Intel. DirectML remains important for broad compatibility, but it can’t always extract the same level of performance as a vendor-tuned provider like TensorRT. TensorRT itself began life in NVIDIA’s datacenter world, where graph optimization, kernel selection, and quantization have been refined over years. TensorRT-RTX adapts that expertise for consumer PCs, focusing on smaller footprint, faster model compilation, and portability.

Microsoft’s move to service execution providers through Windows Update mirrors how graphics drivers and media codecs are handled. With the PC hardware landscape now including NPUs, discrete GPUs, integrated graphics, and even cloud chips, a one-size-fits-all acceleration approach no longer works. By making execution providers swappable and updatable independently, Microsoft can push improvements to NVIDIA RTX users without waiting for a major OS release or requiring app developers to ship new versions.

NVIDIA’s RTX advantage is clear: Tensor Cores are designed for the kind of matrix math that AI models crave, and the CUDA ecosystem is already mature. But Microsoft can’t afford to play favorites. The same Windows ML framework must also work well with AMD’s Ryzen AI, Intel’s Core Ultra NPUs, and Qualcomm’s Snapdragon X chips. KB5089168 demonstrates how vendor-specific updates can coexist within a unified platform—a balancing act that will define Windows’ AI future.

What you need to do

For the vast majority of users, the answer is: nothing. KB5089168 installs automatically through Windows Update, provided your PC meets the requirements. The key prerequisites are:

  • Windows 11 version 24H2 or 25H2
  • The latest cumulative update already installed
  • An NVIDIA RTX GPU (30-series or newer recommended, though older RTX cards may also be supported)
  • A current NVIDIA display driver (no specific version required beyond a fairly recent one)
  • Windows Update not paused or blocked by policy

To verify that the update is in place, open Settings → Windows Update → Update history and look for the entry. If it’s missing, first ensure all pending updates are installed and a restart isn’t waiting. If the entry still doesn’t appear, the device might not be eligible yet—Microsoft often phases these component updates gradually.

On managed enterprise devices, IT policies may control whether dynamic updates like KB5089168 are downloaded. Administrators should check their Windows Update for Business rules and ensure that “Receive updates for other Microsoft products” is enabled if necessary. In some cases, execution provider updates may be classified as optional or driver updates, so verify that those aren’t blocked.

Troubleshooting tip: If an AI-powered app seems slower than expected or falls back to CPU even after the update, the app might not be using the TensorRT-RTX provider. Developers must explicitly register and select the provider at runtime. A properly built app should report which hardware backend is active—look for settings or logs that mention “execution provider” or “inference device.”

What’s next

KB5089168 won’t be the last of its kind. Microsoft’s servicing model for AI components is still taking shape, and we can expect similar updates for AMD, Intel, and Qualcomm providers as those ecosystems mature. The real test will be whether Microsoft starts publishing detailed changelogs for these updates. Developers and IT pros need to know which operator patterns are newly optimized, which bugs are fixed, and whether any regressions are known. Without that transparency, each update becomes a gamble.

For users, the trajectory is encouraging. Local AI is becoming a first-class citizen on Windows, not something that requires a Ph.D. to wrangle. As more apps adopt Windows ML and ONNX Runtime, updates like KB5089168 will make a tangible difference in everyday tasks—from photo editing to document summarization. The ultimate goal is a Windows platform where any AI model can run efficiently on whatever hardware is available, without users having to think about execution providers, driver versions, or update history. That day isn’t here yet, but KB5089168 moves the needle a bit closer.