The latest preview build of Windows 11 drops a small but telling addition into Task Manager: three new optional columns that surface exactly how your PC's neural processing unit is being used, which processes are tapping it, and how much memory it's consuming. KB5089573, released on May 26, 2026, introduces these real-time NPU telemetry columns into the Processes, Users, and Details tabs for systems running Windows 11 version 24H2 or 25H2. Engineers and power users who've been longing for a native performance peek at the AI engine inside Copilot+ PCs and other neural-capable hardware finally get a built-in window beyond third-party tools.

Task Manager has long been the go-to gauge for CPU, GPU, memory, disk, and network consumption. The addition of NPU as a first-class resource completes the performance-monitoring trifecta for the age of AI-accelerated laptops. The new columns are not turned on by default; you have to right-click the column header in any of the three supported tabs and select NPU usage, NPU engines, and NPU memory from the extended menu. Once enabled, they behave like any other metric—live, sortable, and color-coded by load thresholds.

What the columns actually show

NPU usage reports the percentage of total NPU compute capacity being consumed by a process, user, or individual execution context. A background service running a speech recognition model might sit at 2–3 percent, while a real-time video effect like Windows Studio Effects could spike to 15–20 percent. If multiple AI features stack together—say, eye contact correction, background blur, and live captions—you could see additive consumption that reveals just how much of the neural silicon is actually awake at any moment.

NPU engines counts how many distinct neural processing engines within the system are engaged by a given workload. This matters because modern NPUs, such as Qualcomm's Hexagon NPU found in Snapdragon X Elite, Intel's Meteor Lake and Lunar Lake neural compute engine, or AMD's Ryzen AI block, contain multiple parallel execution units. A single inference task might light up one engine, while a batched or pipelined workload can saturate several. The column lets developers and IT administrators assess whether an application is fully utilizing the available hardware parallelism.

NPU memory shows the dedicated physical memory allocated to a process from the NPU's local memory pool. Not all NPUs have a separate memory bank—some share system DRAM—but where a discrete or reserved region exists, this metric tracks the committed bytes. It gives a direct read on model footprint, which is critical when tuning on-device AI workloads that must coexist with ordinary applications.

All three columns appear across Processes, Users, and Details, meaning you can get a system-wide per-user summary or drill down to the specific executable hosting an AI task. On the Details tab, the columns are exposed under the right-click Select columns dialog, where you'll find them grouped under a new NPU heading alongside the classic CPU, memory, and I/O counters.

Why telemetry matters now

Microsoft's bet on the Copilot+ PC platform rests on the idea that neural processing will become as routine as graphics rendering. Applications from Adobe, DaVinci Resolve, Zoom, and Microsoft's own Office and Windows Shell now offload inferencing to the NPU—background blur, live translation, document summarization, and super resolution all increasingly lean on the neural unit. Until KB5089573, a user had no simple way to see which apps were actually using that hardware, how much they were loading it, or whether a runaway AI process was draining battery. Third-party monitoring tools like Windows Performance Recorder and vendor-specific utilities (e.g., Qualcomm's Neural Processing SDK profiler) existed, but they were far from a single click away.

The update arrives as enterprises start to scrutinize AI PC ROI. IT departments that roll out fleet-wide applications with neural features need to validate that those features stay within thermal and power budgets. The new Task Manager columns provide a lightweight, always-available audit trail without deploying extra software. That same visibility helps developers who want to right-size their ONNX or DirectML models: they can watch memory pressure and engine utilization in real time while iterating.

How to get the columns today

KB5089573 is a preview cumulative update, so it lands on the optional update channel. Users on Windows 11 24H2 (build 26100.x) or 25H2 (build 27686.x) can install it manually via Windows Update under the "Updates available" section, where it appears as 2026-05 Cumulative Update Preview for Windows 11 Version 24H2 or 25H2. After installation, the build number bumps to 26100.2992 or 27686.1844 respectively. If you wait, the same payload will roll into the mandatory Patch Tuesday release the following month.

Once the update is applied, open Task Manager (Ctrl+Shift+Esc), switch to the Processes tab, right-click any column header, and you'll see the NPU columns listed. The feature lights up automatically if Windows detects an NPU; on machines without neural hardware, the columns are absent or show zero across all entries. This detection hints at a kernel-level NPU driver model that Task Manager now hooks into, likely the same infrastructure exposed to the Windows Performance Analyzer through the NPU Activity graph introduced earlier in 2025.

Beyond Task Manager: the broader NPU tooling story

KB5089573 doesn't arrive in isolation. Since the launch of Copilot+ PCs, Microsoft has been quietly building a developer and diagnostic stack around the neural processor. The Windows Performance Recorder already captures NPU events, and Windows Performance Analyzer visualizes them on a timeline. The DirectML API added explicit NPU memory allocation calls in a recent SDK update, and the ONNX Runtime ships with a new NPU execution provider that logs allocation statistics to ETW trace providers—the same event channels Task Manager now queries.

What's new here is consumer-visible, zero-setup access. A typical user will never run WPR or parse an ETW trace, but they will open Task Manager when their fan spins up. By surfacing NPU columns, Microsoft normalizes the concept of neural resource consumption for a mass audience. The column design mirrors that of the GPU columns added in Windows 10 Fall Creators Update, which demystified graphics memory and compute usage overnight. The same pattern applied to NPU is likely to accelerate user awareness and developer accountability.

Growing pains and what's missing

Early adopter reports from the Windows Insider forums note that the columns can show zero activity even when an NPU-abusing app is running if the app uses certain legacy DirectML interfaces that haven't been updated for the new telemetry model. Microsoft's documentation advises developers to link against the Windows 11 SDK version 10.0.26100.0 or later and use the IDMLDevice1 interface to ensure their memory allocations and engine reservations route through the updated NT NPU driver stack. Apps using the older IDMLDevice might still consume the neural engine but remain invisible to Task Manager.

Another limitation: the NPU memory column only reports dedicated device-local memory. On platforms where the NPU shares system DRAM through an IOMMU mapping, the column stays at zero even when a model is loaded. That could be confusing on Intel Lunar Lake or certain AMD Strix Point implementations where the neural engine shares Last Level Cache with the iGPU and uses system memory as a spillover. Microsoft says a future update will expose a "shared NPU memory" counter that accounts for these architectures.

Real-time refresh rates also differ subtly from CPU and GPU counters. Task Manager's NPU data feeds from the Windows Counter Provider, which samples at one-second intervals rather than the more frequent updates for CPU. That's fine for thermal and battery monitoring but may miss short micro-burst inferencing typical of attention-grabbing wake-word detection or camera-based presence sensing.

The AI PC landscape in 2026

The timing of KB5089573 coincides with what analysts call the "trough of disillusionment" settling into a "slope of enlightenment" for neural PCs. Copilot+ hardware has been on sale for over 18 months; Qualcomm's exclusivity window has ended, and Intel, AMD, and NVIDIA all ship laptop platforms with competent NPUs. The software ecosystem is finally catching up: the Canary channel of Windows 11 already tests an NPU-scheduling feature that lets the OS pre-empt lower-priority neural work when a real-time inference request arrives, and a new power-saver slider in Settings can cap total NPU budget to extend battery.

KB5089573's Task Manager columns will likely be promoted to the stable channel by July 2026, coinciding with the back-to-school refresh cycle. PC manufacturers are leaning hard into marketing dedicated AI keys and always-active neural features; having a transparent performance monitor inside the operating system could be the differentiator that convinces reluctant enterprise buyers.

Developer and IT administrator reactions

Indie AI developers on the Windows Developer Discord have already started posting screenshots of their inference apps lighting up the columns. One developer building a local RAG (Retrieval-Augmented Generation) client noted that NPU memory jumps exactly by the model file size when the first token is generated, while NPU usage stays pinned under 30 percent because a single query rarely saturates all engines. That kind of confirmation reassures developers that they are not pushing the hardware into a thermal throttle.

System administrators are equally enthusiastic. At a recent Microsoft Ignite session, IT pros asked for Group Policy controls to hide or restrict the visibility of NPU columns. Microsoft acknowledged the request and hinted that admins might soon be able to disable the columns through a registry key or Set-ItemProperty script, preventing help-desk confusion on machines that don't have meaningful neural workloads.

What comes next

The NPU columns are clearly a stepping stone. Microsoft's internal roadmap, partially visible through winget package leaks, suggests a full-fledged AI Performance Monitor app is in development for Windows 12's first feature update. In the meantime, Task Manager's new columns will likely extend to the App History tab, offering a historical roll-up of NPU usage per application over days and weeks, just as the existing "CPU time" column does.

For users, the immediate takeaway is simple: the next time you flip on Windows Studio Effects or ask Recall to find a document, open Task Manager and watch the neural engines wake up. For the first time, Windows treats the NPU like the citizen silicon it has become, not a black-box accelerator buried behind a vendor driver. That transparency might be exactly what the AI PC market needs to move from hype to habitual reliance.