top of page

The AI PC Consolidation: NPUs Bring More Intelligence Onto the Laptop

Nov 1
6 min read

How Intel, AMD, and Qualcomm are making local AI a standard part of mobile computing

Ā 

For years, laptop AI mostly meant sending a prompt, photo, or recording to a remote service and waiting for a response. That pattern is changing. Intel, AMD, and Qualcomm now build dedicated neural processing units (NPUs) into laptop processors, giving applications another place to run AI workloads directly on the computer. The shift is turning ā€œAI PCā€ from a marketing label into a more concrete hardware and software category.

Ā 

An NPU is a processor block designed for the repetitive mathematical operations used by many AI models. It works alongside the central processing unit (CPU), which handles general tasks, and the graphics processing unit (GPU), which excels at parallel work and graphics. A suitable AI app can send sustained, relatively low-power tasks—such as noise removal, background effects, speech recognition, live captions, or model inference—to the NPU. The aim is to keep the laptop responsive and reduce battery use while the task runs.

Ā 

This does not mean every AI task moves off the cloud, or that every application automatically uses an NPU. The model, operating system, drivers, application, memory, and workload all affect where the work happens and how quickly it finishes. An NPU is one part of a more distributed computing system, not a miniature data center.

Ā 

Three approaches to the same shift

Intel: Core Ultra brings a hybrid design to more PCs

Intel’s Core Ultra Series 3, introduced at CES 2026, extends the company’s integrated AI PC strategy. Intel says the top mobile configurations provide up to 50 NPU TOPS and that the series is powering more than 200 PC designs. TOPS means trillions of operations per second; the figure is a peak hardware rate, not a promise that every app or model will run at that speed. 1

Ā 

The wider idea is to distribute work across the CPU, GPU, and NPU. Intel describes the NPU as a low-power target for ongoing tasks, while its integrated graphics can help with heavier visual workloads. Its partner software examples include AI-assisted calls, image and video editing, security detection, and local model tools. Whether a feature runs locally depends on the particular app and its supported hardware path. 2

Ā 

Image: Intel Core Ultra Series 3 laptop, pictured by WIRED.
Image: Intel Core Ultra Series 3 laptop, pictured by WIRED. 7

Ā Ā 

AMD: Ryzen AI pairs an XDNA NPU with flexible software

AMD’s Ryzen AI 400 family brings the company’s XDNA-based NPU into a newer generation of laptops. The exact ceiling varies by processor: AMD’s Ryzen AI 5 435 specification lists up to 50 NPU TOPS and up to 59 TOPS across its AI engines overall. Keeping those figures separate is important: total platform compute is not the same as NPU throughput. 3

Ā 

AMD’s developer tools give software teams the option to target the NPU alone or combine it with the integrated GPU. That flexibility suits a range of local workloads, from lightweight assistants and speech tasks to image tools and document question-answering. AMD also supports local model workflows through software such as Lemonade, which can direct inference to the available NPU or GPU. 4

Ā 

Image: AMD Ryzen AI 400 chips, photographed by ServeTheHome.
Image: AMD Ryzen AI 400 chips, photographed by ServeTheHome. 8

Ā Ā 

Qualcomm: Snapdragon X2 pushes NPU throughput higher

Qualcomm’s Snapdragon X2 Elite and X2 Plus continue its Arm-based Windows PC push. Qualcomm’s product specifications list up to 85 NPU TOPS for select X2 Elite configurations; other listed versions are rated at up to 80 TOPS. The earlier Snapdragon X Elite generation is rated at up to 45 TOPS, illustrating how quickly the advertised capacity is changing. These are model-specific peak figures, not a guarantee of faster results in every application. 5

Ā 

Qualcomm emphasizes concurrent, low-latency on-device AI and has built out tools for developers to deploy models through Windows ML, ONNX Runtime, and its own AI software stack. That support matters because hardware only becomes useful when applications can reach it. Snapdragon’s Arm architecture also makes compatibility and application availability part of the buying decision, alongside the NPU figure itself. 5Ā 6

Ā 

Image: Qualcomm Snapdragon X2 Elite product image.
Image: Qualcomm Snapdragon X2 Elite product image. 5

The useful comparison is not just the TOPS number

The current headline figures—up to 50 NPU TOPS for Intel Core Ultra Series 3, 50 for the cited AMD Ryzen AI 5 435, and 80–85 for selected Snapdragon X2 Elite parts—show that dedicated AI hardware is now a regular point of competition. They should not be treated as a simple ranking. Vendors can use different precision formats and measurement methods, and a laptop’s power limits, cooling, memory, drivers, and model support influence real results. 1Ā 3Ā 5

Ā 

Microsoft’s Copilot+ PC program has helped standardize the category by requiring an NPU capable of more than 40 TOPS for its qualifying systems. Microsoft’s developer guidance also stresses that the NPU must be deliberately targeted by software. Windows ML can identify available hardware and route supported models through an appropriate execution provider, but that does not make every application NPU-aware. 6

Ā 

For local generative AI, memory capacity can be as important as the NPU. Model size, quantization, context length, and the amount of memory available to the GPU or NPU all shape what a laptop can run comfortably. Smaller speech, image, and language models are a natural fit for efficient on-device use. Larger models may need the GPU, a hybrid CPU/GPU/NPU path, or a remote service. AMD’s developer material similarly distinguishes low-power NPU workloads from larger models better suited to integrated graphics. 4

Ā 

Local processing can improve privacy—but does not guarantee it

Keeping a prompt, recording, or document on the laptop can reduce the amount of sensitive information sent to a cloud provider. It can also lower network latency and allow certain tasks to continue without an internet connection. That is particularly useful for private notes, internal documents, or live captions during travel.

Ā 

But an NPU does not make an application private by itself. A program can still upload data for account syncing, analytics, backups, or cloud features. Local files also remain exposed to risks such as malware, weak device security, and unauthorized access. Users and organizations still need to check each app’s data settings, permissions, and retention practices. Local execution changes the available privacy choices; it does not replace them.

Ā 

A hybrid future, not a cloud-free one

Cloud AI remains valuable when a task needs a very large model, current web information, shared collaboration, or computing power beyond a laptop’s thermal and memory limits. The local-first approach is strongest for frequent, bounded tasks where speed, offline access, predictable latency, or data control matters. Many products will use both: the laptop handles routine work, while a cloud service handles requests that need more capacity or fresh information.

Ā 

The larger shift is that the laptop now has a dedicated engine for AI that can work quietly in the background without asking the main processor to do everything. As apps, APIs, and optimized models mature, users should see more features that work directly on their machines and fewer reasons to upload every small task. The deciding factor will be practical support in the software people use, not simply the largest number on a processor box.


Closing Thoughts

My view is that NPUs matter most when they quietly make everyday AI useful: a call sounds clearer, captions appear without a network connection, or a private document can be searched without being uploaded. The silicon creates the option; good software and transparent controls determine whether people trust and use it. AI PCs are becoming more capable, but their strongest case is practical convenience and choice—not the claim that every task should run locally.

Ā 

What does this look like in practice

A student could transcribe a lecture offline, then ask a small local model to find a definition in personal notes. A consultant could search a confidential project folder without sending every document to an external service. During a video call, the NPU could handle background blur and noise suppression while the CPU keeps the meeting app responsive. If the user asks for current news or a complex analysis beyond the laptop’s capacity, the software can offer a cloud option instead.

Ā 

Why does this make a difference?

It gives the computer a way to perform routine AI work with lower delay, less dependence on an internet connection, and fewer automatic transfers of private input. It can also preserve CPU and GPU capacity for other tasks and improve efficiency for sustained workloads. The benefit is not absolute: results depend on compatible apps, model size, system memory, and sensible privacy settings. The difference is that local processing becomes a real option built into mainstream laptops, rather than a specialty setup.


References

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page