At a TechMentor session next month, Microsoft deployment veteran Michael Niehaus will tackle a question that has quietly frustrated IT administrators since Copilot+ PCs hit the market: Which Windows 11 features actually use the neural processing unit inside your laptop, and which are just as happy running on the CPU, GPU, or in the cloud? The session, "Windows 11: Using Local AI Silicon," takes place August 5, 2026, at Microsoft headquarters in Redmond and promises to cut through the marketing fog surrounding local AI acceleration.

Niehaus, now programme director at 2Pint Software and a former 16-year Microsoft engineer, is well-known in enterprise circles for his no-nonsense approach to Windows deployment and management. His 75-minute talk, part of TechMentor & CyberSecurity Live!, aims to give IT pros a practical map of what the NPU actually contributes to Windows 11—and, by extension, to the bottom line of any organization buying new hardware.

What’s really happening inside your PC

The neural processing unit is a dedicated accelerator for machine-learning inference. Unlike a CPU’s general-purpose cores or a GPU’s many parallel shaders, an NPU is built specifically for the matrix math that drives today’s AI models. In theory, that means lower power draw and higher throughput for tasks like image recognition, speech transcription, or text generation.

In practice, Windows 11’s AI architecture is a hybrid beast. Some features run entirely on-device; others lean on the cloud; many bounce between CPU, GPU, NPU, and remote servers depending on what’s available. The NPU is not a universal replacement for other silicon. A workload that is “local” doesn’t automatically mean it’s NPU-accelerated.

Take Windows Studio Effects, introduced in 2022. Its background blur, eye contact, and automatic framing effects can run on the NPU of a Copilot+ PC, but they also work fine on a conventional Intel or AMD GPU. The difference is usually efficiency: NPU execution sips battery, while GPU or CPU execution runs hotter and drains a charge faster. For a laptop in a conference-call marathon, that efficiency can matter a lot. For a plugged-in desktop with a discrete GPU, it’s barely noticeable.

Since the Copilot+ PC launch in 2024, the most visible NPU-native feature is Phi Silica. Microsoft describes it as a small language model optimized for on-device use—think summarization, rewriting, and text-to-table conversion inside apps like Word or OneNote. On a Copilot+ PC, Phi Silica runs directly on the NPU. But here’s the twist: Microsoft has also expanded Phi Silica support to some non-Copilot+ systems, where it will happily run on a supported GPU instead. The model adapts to whatever hardware is available.

That flexibility is deliberate. According to Microsoft’s Windows AI documentation, the platform spans CPU, GPU, NPU, and cloud execution. A developer writing an app that uses the Windows AI APIs—collectively branded as Microsoft Foundry on Windows—can ask for “local inference” without specifying which accelerator to use. Windows decides at runtime what to target. That means an app that hums along on an NPU today might fall back to a GPU tomorrow if the NPU driver misbehaves, or vice versa.

The API surface itself covers a growing list of capabilities: text recognition, image feature extraction, speech-to-text, and local language-model tasks. For developers, the promise is enormous: add AI-powered features without having to package and tune a model for every device class. For IT buyers, the promise is murkier. “Runs on NPU” is not a feature checkbox; it’s a performance characteristic that matters only when the workload and the hardware align.

What it means for you

If you’re a home user, the NPU in your shiny new laptop is likely making a handful of features feel snappier or last longer on battery. The camera effects you toggle in Settings → Bluetooth & devices → Cameras might now be listed as “NPU-optimized.” Voice typing in a quiet room may have lower latency. The Photos app might search your library for “beach” without an internet hiccup. But don’t expect every AI mention in Windows to suddenly become offline and instantaneous. Microsoft Copilot, the web-based assistant, still talks to the cloud. Many of the “AI” touches in Paint and Snipping Tool rely on remote servers. The NPU is a helper, not a replacement for datacenter horsepower.

If you’re an IT pro, the picture demands sharper scrutiny. The decision to standardize on Copilot+ PCs—or any PC with an NPU—should be tied to specific workloads, not vague future-proofing. Start by inventorying your fleet. If you’re buying new laptops, look beyond the NPU TOPS number and ask: which Windows features and third-party apps will actually use this chip? For example, if your field workers need offline speech transcription for compliance reports, an NPU-equipped PC running a local speech model could be transformative. If your call-center agents spend all day in a browser-based CRM, the NPU’s efficiency gains in a video call may be marginal.

Task Manager, as of Windows 11 version 24H2, now shows NPU utilization as a separate graph, much like CPU and GPU. That gives support teams a simple way to verify that a workload is lighting up the accelerator. If your vendor promises that a line-of-business app “runs on NPU,” open Task Manager, run a typical task, and see if the bar moves. A flat line means the app is either falling back to another processor or simply not using AI at all.

Remember too that NPU support varies by chip generation. Qualcomm’s Snapdragon X Elite and X Plus were first to market with Copilot+ compliance (40 TOPS or more). Intel’s Core Ultra 200 series (Lunar Lake) and AMD’s Ryzen AI 300 series have since joined the club. Older NPUs—like those in Intel Meteor Lake or AMD Ryzen 7040—offer fewer TOPS and may not accelerate all Windows AI features. Consult Microsoft’s published Copilot+ PC list to confirm your model’s capabilities.

How we got here

Windows’ flirtation with local AI didn’t start with the NPU. Windows ML, introduced in 2018, let developers run pre-trained models on any DX12-capable GPU. DirectML, a low-level API, gave game developers and content apps access to GPU-accelerated inference. But adoption was limited. Most apps either ignored on-device AI or relied on the CPU, which was too slow for real-time tasks.

The chip industry changed the calculus. Apple’s M-series chips, with their unified memory and dedicated neural engine, showed that always-available AI could feel like magic—instant photo object removal, voice isolation that didn’t require a cloud round-trip. Microsoft responded with the Copilot+ PC initiative in 2024, setting a high hardware bar: 40 TOPS of NPU performance, 16 GB of RAM, and 256 GB of storage. The message to Intel, AMD, and Qualcomm was clear: if you want the Windows logo optimized for the AI era, include a serious NPU.

Qualcomm was first to cross the line with the Snapdragon X Elite, launching in mid-2024 alongside the first wave of Copilot+ Surface devices. Intel and AMD followed in late 2024 and early 2025 with Lunar Lake and Strix Point (Ryzen AI 300), respectively. The result is a fragmented landscape: millions of PCs now have an NPU, but the software that uses it is only beginning to emerge.

Microsoft’s strategy is to make local inference a platform capability, not a one-off bullet point. That’s why the company is pushing Phi Silica as a built-in model, why the Windows AI APIs expose a “local LLMs” endpoint, and why docs explicitly advise developers not to assume a specific hardware backend. The goal is to make NPU acceleration as invisible to users as GPU acceleration already is—something that just works when the right app is running.

What to do now

For IT departments, the near-term work is methodical:

  1. Audit your device inventory. Identify which machines have an NPU and which meet the Copilot+ threshold. Use a tool like Microsoft’s own PC Health Check app or a commercial endpoint management tool to pull NPU capability details.
  2. Pinpoint actual use cases. Don’t buy AI PCs for the sake of “AI.” Identify tasks that would benefit from offline, low-latency inference: real-time speech translation for traveling salespeople, on-device document classification for legal teams, camera-driven inventory scanning in a warehouse. Then verify that the necessary software stack supports NPU execution.
  3. Test, test, test. For each candidate workload, run a pilot on the target hardware. Open Task Manager to confirm NPU activity. Measure battery life impact. Compare performance against the same task running on a GPU or CPU. A workload that consumes 3 W on the NPU versus 15 W on the GPU could significantly extend a laptop’s workday.
  4. Engage with your app vendors. Ask them explicitly: does your Windows client use the Windows AI APIs, and does it target the NPU? If the answer is vague, push for a roadmap. Many enterprise ISVs are still adapting to the idea that local AI is a real, deployable option.
  5. Prepare your Windows image. Make sure you’re deploying Windows 11 24H2 or later, which includes the full NPU driver stack and Task Manager integration. Older builds may not enumerate the NPU correctly.

For developers, the path forward is already documented. Start with the Windows AI platform overview on Microsoft Learn. Explore the Microsoft Foundry API for tasks like text extraction (OCR), image captioning, and local language model invocation. The key benefit: you don’t need to ship a model. Windows provides the model, and the runtime picks the best hardware. If you’re maintaining a .NET desktop app, the Windows AI APIs are callable from C# and C++. If you’re on WinUI 3 or WPF, you can integrate summarization or text analysis with a few lines of code.

And for everyone, consider attending Niehaus’s session, either in person or via on-demand recording. The practical value of his talk isn’t in theoretical AI hand-waving; it’s in mapping the specific Windows 11 features that light up an NPU—camera effects, audio cleanup, local language models, and more—to the hardware sitting on desktops right now.

The outlook

The next 18 months will see NPUs become a table-stakes feature, not a premium differentiator. Intel’s upcoming Panther Lake and AMD’s future chips are expected to push NPU performance well beyond 40 TOPS. Microsoft will almost certainly expand the Windows AI API surface, likely adding vision and speech models that rival cloud quality. And enterprises that build a muscle around testing and deploying NPU-friendly workloads today will be in a far better position when those workloads become mission-critical.

Don’t mistake the NPU for a silver bullet. It’s a tool—an efficient, specialized tool for a particular class of computation. Knowing when to use it, and how to know it’s being used, is the real skill. That’s what Niehaus’s session aims to teach, and it’s what any organization buying PCs in 2026 needs to learn.