Open Task Manager while you're chatting with a local model in Ollama or LM Studio on your brand new "AI PC," and you'll probably see something confusing. The GPU graph spikes. The CPU ticks up. And that NPU your laptop was sold on, the one with the 40 or 50 TOPS sticker on the box, sits flat at zero.
That's not a bug on your end, and it's not a driver issue you can fix by updating something. It's how these tools are built, and understanding why actually matters if you're deciding what to buy or how to set expectations for local AI on a Copilot+ PC in 2026.
What Your NPU Was Actually Designed To Do
The NPU (Neural Processing Unit) in your Snapdragon X, Intel Core Ultra, or AMD Ryzen AI chip wasn't built with "run a 7B chat model" as its primary job. It was built for small, constant, low-power AI tasks that run in the background without draining your battery or spinning up a fan.
Info!
Windows features that genuinely run on the NPU include Windows Studio Effects (background blur, eye contact correction in video calls), Live Captions with real-time translation, Recall's semantic indexing, and small on-device models like Microsoft's Phi Silica. These are lightweight, always-on workloads, not the same job as generating long chat responses from a multi-billion parameter model.
A local LLM chat session is a different kind of workload entirely. It needs to load gigabytes of weights, run repeated matrix multiplications for every single token, and do it fast enough to feel conversational. That's a job GPUs have been optimised for since long before "AI PC" became a marketing term, and it's the job Ollama and LM Studio were originally built around.
Why Ollama and LM Studio Don't Touch the NPU
The short version: Ollama and LM Studio both run on llama.cpp under the hood, and llama.cpp loads models in GGUF format. GGUF was designed to be flexible; model weights get quantised into different bit-widths and dequantised on the fly, and the whole thing is built to run well across CPU and GPU backends like Vulkan, ROCm, CUDA, and Metal.
NPUs don't work that way. They expect models converted into ONNX format and compiled ahead of time for a specific vendor execution provider: QNN for Qualcomm's Hexagon NPU, OpenVINO for Intel, or the Ryzen AI Software stack for AMD's XDNA cores. That conversion has to happen per model, often with a fixed quantisation scheme the NPU's fixed-function hardware can actually execute. There's no dynamic dequantisation step happening on the fly the way GGUF allows.
In other words, GGUF prioritises flexibility and broad compatibility. NPU execution providers prioritise a narrow, pre-compiled, highly efficient path. Those are two different design philosophies, and as of mid-2026, nobody has fully bridged them into a "just works" experience inside the popular local-LLM tools.
Warning!
If you're checking whether your AI PC purchase was "wasted" because Ollama isn't using the NPU, it wasn't. The NPU is still doing its job for Windows AI features running in the background. It just isn't the engine behind your chat window.
Is Anyone Actually Building NPU Support Into These Tools?
There's real movement here, just slower and more fragmented than the "AI PC" marketing implies:
- llama.cpp merged a QNN backend targeting Qualcomm's Hexagon NPU, giving Snapdragon-based laptops a path to NPU acceleration outside the mainstream Ollama/LM Studio builds.
- Windows ML now automatically picks the right execution provider (QNN on Qualcomm, OpenVINO on Intel) for apps built to use it, falling back to GPU or CPU when no NPU provider is available. That's a Microsoft-level plumbing change, not something Ollama or LM Studio opted into directly.
- Community projects like NoLlama (built around Intel's OpenVINO GenAI) exist specifically because Ollama can't touch the NPU on Intel hardware out of the box.
Meanwhile, official feature requests are still open and unresolved. There's a standing request on Ollama's GitHub asking for Rockchip NPU support, and LM Studio has had an open feature request for Intel, AMD, and Qualcomm NPU support sitting unresolved for a while now. Neither team has committed to a timeline.
If you're already running local models on your GPU and want to understand how that whole local-AI landscape fits together, our deep dive on local AI vs cloud AI in 2026 covers why running models on your own hardware became practical in the first place, NPU or not.
NPU Support Across Vendors, As It Actually Stands
| Platform | NPU / TOPS | Vendor NPU Stack | Ollama / LM Studio Uses NPU? |
|---|---|---|---|
| Qualcomm Snapdragon X / X2 Elite | Hexagon NPU, 45 to 80 TOPS | QNN (Qualcomm AI Engine) | No by default. llama.cpp has an experimental QNN backend outside mainstream builds. |
| Intel Core Ultra (Lunar Lake / Panther Lake) | Intel AI Boost NPU, 40 to 50 TOPS | OpenVINO / OpenVINO GenAI | No. Runtimes default to the Arc iGPU via SYCL/Vulkan, or CPU. |
| AMD Ryzen AI 300 / 400 series | XDNA 2 NPU, up to 50 TOPS | Ryzen AI Software | No. Radeon iGPU under ROCm or Vulkan does the heavy lifting instead. |
So Is the NPU On Your AI PC Pointless?
Not pointless, just narrower than the sticker on the box suggests. If you use Windows Studio Effects on video calls, rely on Live Captions translation, or use any app built specifically on Windows ML or a vendor's NPU SDK, the NPU is doing real, useful, battery-friendly work every day. What it isn't doing, at least not through Ollama or LM Studio in 2026, is powering the chatbot window you opened to test out a downloaded model.
- You use video calling apps with background blur or eye contact correction daily.
- You rely on Live Captions or on-device translation features
- You want longer battery life for lightweight, always-on Windows AI features
- You expect NPU-backed local LLM support to mature over the life of the laptop
- Your actual goal is running Ollama or LM Studio for local chat and coding models
- You're choosing between similarly priced laptops based mainly on TOPS numbers
- You need larger models loaded right now, not "eventually, once software catches up"
- You'd rather prioritise RAM and GPU capability, which is what today's tools actually use.
What This Means If You're Buying a Laptop in India in 2026
Copilot+ laptops with a qualifying NPU now start around the ₹70,000 to ₹75,000 mark in India for entry Core Ultra 5 configurations, and go up from there depending on RAM, display, and chassis. If your actual plan is to run local LLMs through Ollama or LM Studio, that NPU line item on the spec sheet shouldn't be your deciding factor. What matters far more today is RAM capacity (since it directly caps which model sizes you can even load), memory bandwidth, and how capable the integrated GPU is, since that's the component doing the actual work right now.
We've covered this from the build side too. If you're weighing a dedicated local AI setup instead of leaning on a laptop's iGPU, our budget AI PC build guide walks through getting usable VRAM without overspending, and our look at the AMD Radeon AI PRO R9700's India pricing is a useful comparison point if you're deciding between a laptop's integrated graphics and a dedicated card for local AI work.
Info!
None of this makes Copilot+ laptops a bad buy. It just means you should buy one for what it's actually good at today (battery life, background Windows AI features, general productivity) rather than assuming the NPU will make Ollama faster the moment you install it.
Frequently Asked Questions
Does my NPU do anything at all when I run Ollama or LM Studio?
No. As of mid-2026, both tools route inference to your GPU (via Vulkan, ROCm, or Metal, depending on platform) or your CPU. The NPU stays idle during that specific workload, even though it may be active elsewhere in the background for other Windows AI features.
Will Ollama or LM Studio ever support NPUs properly?
Possibly, but there's no confirmed timeline from either team. llama.cpp already has an experimental QNN backend for Qualcomm hardware, and Windows ML's automatic execution provider selection shows the underlying plumbing is maturing. Feature requests for Intel, AMD, and broader NPU support remain open and unresolved on both projects' trackers.
Should I avoid buying an NPU-equipped laptop then?
Not necessarily. Most modern laptops come with an NPU anyway as part of current-generation chips, and it's genuinely useful for Windows Studio Effects, Live Captions, and battery-friendly background AI tasks. Just don't pick a laptop for local LLM work based on its TOPS number. Prioritise RAM and GPU capability instead.
