The era of the AI PC has arrived: laptop makers, silicon designers and Microsoft have reworked the basic architecture of Windows notebooks so that a dedicated Neural Processing Unit (NPU) sits alongside the CPU and GPU, and Microsoft’s Copilot+ stack is the first operating-system-level effort to expose those on-device AI capabilities to everyday users.
Background / Overview
The last desktop-era architectural shift for mainstream Windows laptops introduced powerful integrated GPUs and then discrete GPUs; now there’s a third co‑processor purpose-built for AI. That NPU performs the low‑precision matrix and tensor math that underpins modern models for image recognition, speech transcription, on‑device translation, noise suppression, live video effects and lightweight generative tasks. Microsoft’s Copilot+ program explicitly targets systems with NPUs rated at
40+ TOPS (trillions of operations per second), a practical threshold the company uses to ensure a good on‑device experience. OEMs and silicon vendors are responding: AMD’s Ryzen AI family, Intel’s Core Ultra (Lunar Lake / 200V series) and Qualcomm’s Snapdragon X line all include NPUs capable of bringing many Copilot experiences on‑device. The result is a new product category often marketed as
AI PCs or
Copilot+ PCs — machines designed to run local inference for responsiveness, privacy and energy efficiency while preserving the traditional strengths of laptops.
The foundation of the AI PC: Ryzen AI and the NPU
What the silicon brings
AMD’s Ryzen AI 9 HX 370 — the chip AMD supplies in certain premium notebooks — is an exemplar of the new generation. It pairs a high‑performance Zen5 CPU cluster, integrated RDNA graphics, and an XDNA2 NPU block. Public technical listings and vendor spec pages show the chip’s NPU is rated in the 40–50 TOPS class (the combined chip TOPS number that counts CPU+GPU+NPU operations is quoted higher), which is precisely the performance band Microsoft targets for Copilot+ features. Why this matters: NPUs are optimized for the tiny, repetitive linear algebra operations that run most neural nets. Compared to moving every task to the cloud, performing that math locally reduces latency, keeps private data on the device by default, and—when properly engineered—consumes far less energy for continuous or background AI tasks (like live captions or eye‑contact correction) than running a CPU‑oriented solution. That energy/time tradeoff is fundamental to the “always‑on but battery‑friendly” promise of AI PCs.
Real silicon numbers and practical limits
- AMD’s published part listings and independent hardware databases put the Ryzen AI HX 370 in the ~50 TOPS (NPU) / ~80 TOPS (chip combined) range depending on measurement mode and precision. These are rated figures; real‑world usable throughput depends on thermal headroom, memory bandwidth and driver maturity.
- TOPS is a useful comparative metric but not an absolute predictor of every workload. Software stack optimizations, kernel drivers (ONNX/DirectML), and model quantization strategies change how many usable inferences a given TOPS budget delivers.
Windows 11 and Copilot+: the OS built for on‑device AI
Microsoft has reworked Windows 11 to make AI a first‑class system capability. Rather than bundling Copilot into a single app, Copilot+ integrates local AI primitives into the OS experience and selectively offloads heavier work to the cloud when needed.
Key Copilot+ features delivered today
- Windows Studio Effects — real‑time webcam and microphone improvements like background blur, voice isolation, eye‑contact correction and automatic framing, accelerated by the NPU so effects run in foreground/video calls without stalling user workflows.
- Live Captions and Real‑time Translation — device‑generated captions for audio and the option to translate speech into other languages locally when the device meets Copilot+ hardware criteria.
- Cocreator / Paint and AI creative tools — on‑device support for generating assets, removing objects, or suggesting edits in Photos, Paint and Clipchamp, where NPUs speed up inference and make the tools feel responsive.
- Click to Do, Highlights, and Guided Help — context‑aware UI affordances where Copilot can highlight parts of a shared window, give step‑by‑step instructions, or execute small tasks with user permission. These interactions are central to the “multimodal” vision Microsoft is shipping.
The Windows team also provides developer guidance (ONNX runtime and NPU device APIs) so ISVs and independent devs can take advantage of on‑device acceleration. That investment in tooling is a signal Microsoft wants the ecosystem to follow.
Copilot Vision: giving Windows sight
Copilot Vision is the clearest demonstration of how multimodal, on‑device AI changes workflows. When the user explicitly shares a window or app, Copilot analyzes visual content, recognizes text and UI elements, and combines that input with voice or typed prompts to deliver contextually accurate assistance. It’s not passive; Copilot can highlight where to click, translate text on the page, summarize content, or explain charts while you continue working. Because Vision requires user consent for screen sharing and explicit session control, Microsoft has positioned it as a privacy‑first capability that does not continuously monitor the desktop (unlike the controversial Recall preview). The Windows Insider updates show rolling improvements — two‑app sharing, “highlights” to point users to exactly the UI elements they need, and a staged rollout across markets.
Why this is more than a gimmick
Multimodal AI blends:
- Voice (natural language inputs and commands),
- Vision (what’s on your screen), and
- Pointer/touch context (where you click or highlight).
That union makes Copilot feel like a companion that understands
tasks, not just text or images. For knowledge work — assembling reports, summarizing dense web pages, debugging snippets inside an IDE — this changes friction points into single‑step instructions. Business users, trainers and creative pros stand to gain the most, especially when the visual analysis is fast because the NPU handles inference locally.
On‑device image generation and creative workflows
AI-assisted image generation on laptops is no longer purely cloud‑centric. Copilot and Paint’s Cocreator can run lighter model components on the NPU to speed prompt parsing, layout suggestions and initial drafts; the cloud still supplies the heavy lifting for ultra‑high‑fidelity renders when needed. This hybrid approach shortens iteration cycles and keeps more pre‑publication work private and offline. It transforms AI illustration from a separate workflow into part of the creative app’s core editing loop. For creators this means:
- Faster prompt feedback and local previews.
- A smoother editing loop with NPU‑accelerated filters and auto‑corrections.
- Reduced latency and increased privacy for early concept work.
However, full generative fidelity and very high‑resolution final renders still often rely on server‑side models for now; expect that balance to shift as NPUs gain more TOPS and software quantization improves.
Real‑world performance: what the Zenbook S16 test shows — and what is unverified
The PCQuest hands‑on with the ASUS Zenbook S16 (powered by an AMD Ryzen AI 9 HX 370) illustrates the practical benefits of the NPU‑enabled approach: real‑time Studio Effects during multitasking, snappier Copilot replies, and faster creative tools thanks to on‑device inference. The article reports a UL Procyon Video Editing Benchmark 1.5 score of
7,741 for Premiere Pro export tasks and notes smooth 1080p/4K handling with Radeon 890M GPU assistance while the NPU supports AI effects. Those observations align with the architectural expectation — a powerful CPU/GPU/NPU trio enables balanced creative workloads on thin‑and‑light systems.
Caveat on benchmarks: the exact UL Procyon score cited in the review does not appear in the public benchmark databases indexed during verification, so that single numeric result should be treated as the reviewer’s reported figure rather than an independently validated database entry. Benchmarks are highly sensitive to thermals, power limit tuning (PL1/PL2/TDP), driver versions and project settings, so different SKUs and firmware revisions will vary. In short: the
direction of the result (strong export performance and responsive AI features) is consistent with the hardware; the specific score is a reviewer data point.
Why AI PCs matter for everyday users — real benefits
The on‑device AI shift is substantive for daily laptop use. The most immediate, practical advantages include:
- Better video call quality via local, low‑latency noise suppression and eye‑contact correction that run without cloud round trips.
- Seamless multimodal interactions that combine voice, screen context and pointer actions so tasks (summaries, guided edits) take fewer steps.
- Faster creative workflows because iterative steps are accelerated by NPU inference and local previews.
- Offline AI features that improve privacy — many Copilot+ experiences can be surfaced even without continuous cloud connectivity when hardware and software allow.
- Lower apparent latency for assistant responses because local models or model fragments run on the NPU instead of a remote server.
These advantages make Copilot+ devices appealing not just to power users but to anyone who values privacy, responsiveness and fewer cloud dependencies.
Risks, limitations and minefields
The technology is promising, but the rollout exposes meaningful challenges and user risks that IT buyers and end users must weigh.
Privacy and the Recall debate
Microsoft’s
Recall (a timeline of periodic screen snapshots indexed for later search) exemplifies the tension between convenience and privacy. Even when data is stored locally, features that capture activity histories create attack surfaces that must be hard‑partitioned, encrypted and opt‑in. The initial criticism of Recall forced Microsoft to strengthen encryption, require Windows Hello authentication and move toward not enabling Recall by default — but user trust takes time to rebuild once privacy concerns surface.
Software maturity and ecosystem readiness
- Many AI features require updated drivers, runtime libraries (ONNX/DirectML) and application support to be fast and stable. Early devices may ship before the full software stack is mature, leading to inconsistent user experiences.
- Some early ARM‑based Copilot+ laptops ran into compatibility or emulation friction with legacy apps; with AMD and Intel x86 solutions, that friction is reduced—but app developers still need to adopt NPU‑aware patterns to reap full benefits.
Security and manageability
On‑device models and cached local data call for robust platform security: Pluton, secured‑core PCs and firmware integrity are important but not complete solutions. Enterprises will demand assurances around model update provenance, secure sandboxes (Agent Workspace / Copilot Actions), and policy controls before deploying mass‑market Copilot+ features.
Cost, battery tradeoffs and expectations
Copilot+ hardware currently sits in mid‑to‑premium price points. Buyers must balance perceived long‑term utility against higher up‑front cost, and remember that
not every AI function is instantaneous offline — some generative tasks still use cloud resources for best quality. Battery life improvements for AI workloads are real, but heavy mixed workloads still tax thermals and runtime.
How to evaluate a Copilot+ AI laptop (practical checklist)
- Check the NPU spec: 40+ TOPS is Microsoft’s Copilot+ threshold; higher TOPS generally mean more local capabilities.
- Confirm memory and storage: ≥16 GB RAM and ≥256 GB SSD are Microsoft’s recommended baseline for Copilot+ experiences.
- Review the software stack: ensure Windows 11 updates, vendor drivers, and the Copilot app are current; preview features often roll out via the Microsoft Store or Insider channels.
- Test the experience: check live captions, Studio Effects, Copilot Vision highlights and sample creative tasks during a demo to judge latency and accuracy.
- Verify privacy defaults: make sure Recall‑style features are opt‑in and check whether captured data is encrypted and protected by Windows Hello or equivalent safeguards.
The road ahead: what to expect next
- NPUs will keep getting faster and more power efficient; the balance with GPU/CPU capabilities will determine whether fully offline generative models become commonplace on laptops.
- Microsoft will continue iterating Copilot app features — more robust Copilot Actions, deeper app integrations and richer multimodal reasoning — but regulatory scrutiny (especially in the EU) and enterprise security needs will shape rollout timelines.
- Developers who adopt ONNX, quantized model formats and NPU runtimes will be rewarded: apps that ship NPU-accelerated features will feel snappier and use less battery than those that rely on CPUs alone.
Conclusion
The combination of AMD’s Ryzen AI silicon, Microsoft’s Copilot+ software layer and OEM hardware design is more than a marketing label — it signals a platform shift. On‑device NPUs make a practical difference for the kinds of interactive, multimodal assistance that are now part of Windows 11: faster Copilot replies, Copilot Vision that can
see and coach you through tasks, and creative tools that feel immediate instead of cloud‑bound. The ASUS Zenbook S16 review captures these advantages in daily use, showing how a modern CPU/GPU/NPU trio changes expectations for notebooks.
That said, the category is young. Expect unevenness: privacy discussions around Recall, software maturation, driver and app support variability, pricing and the need to verify reviewer benchmark claims are all part of the adoption path. For buyers, the right approach is cautious optimism: evaluate a Copilot+ laptop by its real, day‑to‑day benefits (video calls, captioning, on‑device editing) rather than by a single spec or demo, and demand details about how your data is secured and when features will be available offline.
AI PCs are not a passing fad. They are an architectural upgrade that transforms how Windows laptops behave and how users work. With Ryzen AI, Copilot+ and a growing ecosystem of NPU‑aware software, on‑device intelligence is already changing the practical experience of personal computing — and the next two years will decide how widely that transformation touches mainstream computing.
Source: PCQuest
Ryzen AI and the rise of the AI PC: How Copilot+ and on-device intelligence are transforming Windows laptops