The more advanced Windows AI features actually highlight the awkward situation of AIPC.

Previously, we at SanYi Life have repeatedly criticized the various strange phenomena in the PC industry when adapting AI functions in our relevant content, and we've also conducted some actual "AIPC" tests.

For example, we once tested "whether an old computer without an NPU can run edge-side AI". The result showed that today's CPUs and GPUs can generally support running local large models, and an NPU is not a "must-have option".

Another example is that we've also repeatedly tried to tinker with the predecessor of the modern "NPU", which was then called the VPU, an external computing power unit. Practice has proven that the Windows system can directly recognize this "visual processing unit" and automatically install the drivers, but it cannot use it as an NPU. (We're still continuously experimenting with this topic.)

In addition, we've successively tested and used "AIPC" devices from multiple different brands and architectures for a long time. This includes the common x86 platforms with NPUs, as well as other architecture solutions such as Qualcomm Snapdragon X series and Apple M series.

Based on our own experience, the current AI ecosystem on the PC side actually doesn't lack "software adaptation". Some of these AI software and functions come directly from the driver integrations of relevant hardware manufacturers (a typical example is that NVIDIA and AMD graphics card drivers come with AI frame interpolation, AI upscaling, AI video processing, AI audio noise reduction and other functions), while others come from the "ecosystem partner" circles carefully cultivated by hardware manufacturers (Intel has done particularly well in this regard).

In other words, as long as users really know how to tinker and understand where to find the AI operation solutions suitable for their hardware, even old computers from four or five years ago can actually enjoy the ecological dividends of "AIPC".

However, there's a major drawback to the various "AIPC" ecosystem and software experiences mentioned above. That is, they are either pre - installed by PC manufacturers when leaving the factory (which basically means only new machine buyers can use them), or users need to know some relevant knowledge to find suitable software themselves, and even configure the operating environment and download large models on their own.

Obviously, neither situation is very convenient for ordinary consumers. So, for today's PC users, is there a simpler AI experience method that doesn't require self - configuration and can be more closely integrated with the operating system?

At least in Microsoft's view, there is.

At the end of July 2025, Microsoft published an official blog post, detailing a series of AI functions that the latest version of Windows 11 has introduced.

For example, the search box in Windows 11 now has an intelligent agent feature. Users can let the system automatically perform operations through natural language. For instance, when users say "the mouse cursor is too small" or "lower the screen resolution", the system can automatically change the relevant settings.

Another example is that Windows 11 now also has an "Intelligent text" function based on an edge - side model, which can perform operations such as document summarization, document rewriting, and creating to - do lists based on documents offline.

In addition, Windows 11 also supports using AI to accelerate a series of image - related functions. This includes recognizing the content of local images and describing "what's in the picture" in text, as well as AI - based image beautification, such as "relighting" portraits in the photo app, realizing interactive text - to - image generation in the drawing program, or accurately cropping a certain shape or extracting a certain color in the screenshot app.

Moreover, we must also mention the "Recall" feature of Windows 11. This controversial AI function remembers the documents, pictures, and web page content that users have edited or browsed in the past. According to Microsoft, it's a "user behavior analysis" implemented entirely on the local side based on edge - side computing power, so no data will be uploaded.

So users can ask the system questions in natural language to more easily find the content they've forgotten. For example, when users say "web pages I've read about the XX topic" or "photos from the last meal", Windows 11 will "dig out" the relevant pages and files. Of course, most of the controversies about this function may not stem from distrust of AI, but from the worry that familiar people may "peek" at those "skeletons in the closet" that users don't want to reveal.

After saying so much, do you feel that something's not quite right? That's right. If, like us, you "happen" to love updating the Windows system version and "happen" to have a variety of different PC hardware platforms, from the 11th - generation Core desktop version to the Ryzen 9000X3D, then you'll find that "although the system is the latest version and the PC hardware includes both old platforms and the latest flagships, why haven't I seen any of these Windows 11 built - in AI functions that Microsoft mentioned?"

Yes, this is what we're going to criticize this time. Taking a closer look at the Windows 11 built - in AI functions listed by Microsoft, it's not hard to find that many of them have additional "restriction conditions".

For example, to use the natural - language intelligent agent in Windows 11, at this stage, users must have a Windows 11 AI+ PC device with a Qualcomm Snapdragon chip (that is, it must use a Snapdragon X series processor). This function is not yet compatible with AMD or Intel processors, even though they theoretically have enough computing power.

Another example is that Windows 11 does integrate an edge - side model for text rewriting, but it currently only supports English. And it's not just about working on English (text) content. Users must first change the system language to English to "activate" this function. Moreover, contrary to the situation of interacting with the intelligent agent mentioned earlier, this function does not support the Qualcomm Snapdragon platform and is only available on some Intel and AMD hardware.

This isn't the end. Although Qualcomm launched a Windows PC solution with a built - in NPU early on, and the "AI processors" of Intel and AMD have gone through at least three generations of iterations, when we communicate with relevant hardware manufacturers, they always say that their software and drivers can ensure that even the earliest AIPC platforms are still compatible with the latest code. At the same time, these "AI processors" generally support heterogeneous processing acceleration between the CPU, GPU (integrated graphics), and NPU, and some are even additionally compatible with "computing power integration" with discrete graphics cards.

However, on Microsoft's side, a large number of Windows 11 built - in AI functions can only run on NPUs, and only support a few processor models with an NPU computing power of over 40 TOPs. Considering that a large number of the latest high - performance PC processors' NPUs do not meet this standard, or even don't have an NPU at all, Microsoft's design obviously limits the compatibility of Windows 11 AI functions.

So why does Microsoft do this? From the public information, it seems that they believe it's necessary to promote the "NPU ecosystem" and make it the third major computing unit in future PCs after the CPU and GPU, so that AI programs can "always run" more efficiently without occupying the resources of the CPU and GPU.

It has to be said that the original intention of this idea is indeed good, but it also ignores the objective reality of the current PC industry hardware to some extent. Moreover, this strong binding feature of Windows 11 to "high - computing - power NPUs" doesn't fully meet the needs of all users.

According to Microsoft's technical standards, this "AI desktop workstation" can't even run Windows' AI functions.

After all, as we all know, the advantage of NPUs is high efficiency and low power consumption. But if we only look at objective computing power, the "absolute performance" of NPUs (even the top - level models with 45 TOPs) at this stage is still inferior to that of integrated graphics, and even more so when compared with discrete graphics cards. From the perspective of user needs, not all users use "portable and lightweight laptops", and not everyone cares so much about the device's power consumption or battery life.

For a large number of desktop PC users, or users who use gaming laptops and mobile workstations while being plugged in all year round, isn't it a more reasonable design to call on the GPU or even CPU computing power to achieve faster and more efficient "full - power operation" when using AI applications? Why must we be so fixated on NPUs and emphasize the so - called "low power consumption"?

Of course, Microsoft may have good intentions. They want to promote the architectural innovation of PC processors. But when most people may not be able to really use these "seemingly wonderful" Windows built - in AI functions, it will inevitably make a large number of ordinary consumers question the practicality of AIPC, and even doubt whether Microsoft has insufficient technical capabilities when implementing the "AIPC" concept.

This article is from the WeChat official account "SanYi Life", written by SanYi Jun, and is published by 36Kr with permission.