I had the opportunity to attend Microsoft’s recent launch of its new Surface products, which featured the first Copilot+ PCs, a new category of Windows PC specifically designed for AI. At that launch, Microsoft showed its new vision for the PC, incorporating AI at the OS level while also encouraging developers to build applications for this new platform.
While on the outside the platform doesn’t look like a significant departure from Windows versions of the past, in fact it implements many changes under the hood to both hardware and software. Digging into the details, I see many things from Microsoft that give us a glimpse into the future of the PC—and some of these tangible experiences and tools are available today.
The Copilot+ PC
The Copilot+ PC fundamentally differs from the “AI PC” of the past six months, mainly because nearly every company in the PC space has its own definition of what constitutes an AI PC; but Microsoft’s definition is very specific. It’s also important to note that Copilot+ doesn’t merely imply a PC that’s been optimized to run the base-level cloud-based Microsoft Copilot. Copilot+ PCs are specially designated for running device-side AI at the operating system level and above. This has set a very specific set of requirements for hardware and architecture as well as optimizations for specific models.
For starters, Microsoft has set minimum hardware specs including 40+ TOPS on the NPU and minimum RAM configurations of 16GB. For Microsoft to launch Copilot+, it had to deepen its already close partnership with Qualcomm and resolve some existing issues of using Windows on Arm. For now, the Copilot+ PC is exclusively powered by Qualcomm’s Snapdragon X Plus and Snapdragon X Elite processors, both of which feature NPUs with 45-plus TOPS. The NPU TOPS minimum requirement is essential because it sets a standard for the performance and power consumption that developers can expect from a Copilot+ PC, thereby letting them know how much AI compute they have to play with in their applications. If you’d like to know more about NPUs and their origins, check out this more detailed analysis I wrote a few weeks ago.
These Copilot+ PCs come with a new version of Windows 11 that is inherently AI-accelerated and optimized for Arm processors. In fact, this is the first time ever that Microsoft has led with Arm-based processors as opposed to Intel’s x86 architecture, which is also used by AMD. During the launch, Microsoft made some very bold performance claims against Apple’s 15-inch M3 MacBook Air, including 58% better sustained multithread performance, all while also claiming 20% longer video playback with 22 hours of battery life. Microsoft has worked to alleviate concerns about Arm64 app compatibility, claiming that 87% of the total app minutes people spend on their PCs will be in apps that already have Arm64-native versions. Microsoft also has a new x86 emulator called Prism, which it claims works at least as well as Apple’s Rosetta 2.
As part of the launch, Microsoft introduced a pair of new Surface PCs, the Surface Pro (Gen 11) and Surface Laptop (Gen 7), which both meet all of the minimum specs for a Copilot+ PC. Microsoft also hosted a plethora of PC partners on site at the event to show off their own Copilot+ PCs, which I will discuss in more depth in a forthcoming article comparing prices, specs, and capabilities. The important thing to know is that Copilot+ PCs start at $999 and will begin shipping on June 18—the first day the Snapdragon X Elite is available to the public.
Microsoft’s First-Party AI Apps And OS-level AI
At the Copilot+ launch, Microsoft demonstrated multiple first-party AI apps that it believes will change the Windows user experience. One of the most controversial—and potentially helpful—of these apps is Recall, which periodically takes screenshots of your device usage and then allows you to go back to find things you have seen while working or browsing. Microsoft has already pivoted away from an opt-out implementation towards an opt-in to satisfy critics.
Another app is Cocreator, which allows a user to create something in Paint and use generative AI to help iterate that artwork into a more polished product. Cocreator can also help a user transform an existing photo or image into a particular style—one of the earliest applications of GenAI. Microsoft also enhanced its already AI-accelerated Studio Effects app with new Enhanced Windows Studio Effects, further improving lighting and adding new features such as an eye-contact teleprompter.
Another new OS-wide capability is the implementation of live captions from more than 40 languages into English; this can be done in any app and is a good example of the power of having OS-level AI acceleration. For gaming, Microsoft also implemented a feature called Automatic Super Resolution (Auto SR for short), which improves gaming performance by rendering games at a lower resolution and then using the NPU to boost the resolution back to its native level.
Third-Party Apps And Frameworks
While Microsoft has assembled a respectable list of developers already shipping Copilot+-capable applications, the company needed to implement a new Windows Copilot Runtime to enable both third-party and first-party apps. This new Copilot Runtime is the AI core that enables Copilot+ PCs by extending the Microsoft Copilot stack to Windows. Part of the Copilot Runtime includes the Copilot Library, which is a set of APIs that are powered by the more than 40 on-device AI models that ship with this new version of Windows.
Microsoft has also implemented a new Semantic Index, which helps to redefine search on Windows and is the backbone of features including Recall. Microsoft says that it will expose this capability to developers later in the year with its Vector Embeddings API so developers can implement their own vector store and RAG within their own apps. Microsoft also created Phi Silica, a small local language model based on the Phi family of models, which are specifically designed for running on-device and optimized for the NPU. Microsoft also announced support for PyTorch with its DirectML framework designed to abstract hardware and make it easier to support diverse hardware. Microsoft is also anticipating using AI for the Web with the introduction of WebNN, which also leverages DirectML. WebNN is an API for Windows that is expected to help enable the open Internet to access AI hardware including CPUs, GPUs or NPUs on the device to accelerate Web-based AI experiences.
Adobe came hot out of the gate with support for Photoshop, Lightroom and Express in Windows Copilot+ and said it will add support for Illustrator, Premiere Pro and other apps later this summer. Adobe’s suite of apps has heavily embraced AI creation and editing; in the past, most of those capabilities have relied on the GPU, which can be performant but also power-hungry. Adobe’s biggest competitor in the video-editing space, DaVinci Resolve, also announced support for features such as Magic Mask in Resolve Studio. DaVinci Resolve has shown itself time and again to be an ambitious adopter of new technologies and features, and I’m not surprised that it was involved in the Copilot+ launch.
Microsoft showcased other apps including CapCut, LiquidText, djay Pro and Cephable to demonstrate the power of the NPU. I was especially impressed by how easily the NeuralMix capability in the djay Pro app isolated different tracks within a song to make it easier to queue them up for a DJ. Cephable, an accessibility app, was the one app demoed that I could see myself using, and it demonstrated the benefits of using NPU acceleration with its accessibility tool. I was able to control a slide deck with the mere tilting of my head. It was cool to see how much of the NPU it used in real time while tracking my face and head movements. Cephable, even compared against the M3 MacBook and an x86 plus GPU configuration, showed improvements to performance, battery life and latency by running on the NPU.
These apps merely scratch the surface of what’s going to be possible with Windows on Copilot+ PCs, and we can expect that it will take some time for developers to get a handle on the full potential of the platform and optimize their software for it. Thankfully, developers now have more tools than ever, including the newly announced Snapdragon Dev Kit for Windows, which is supported by Qualcomm’s new AI Hub that the company recently expanded to support on-device PC models.
Wrapping Up
Microsoft has undoubtedly changed the trajectory of the PC industry with the announcement of the Copilot+ PC and the suite of first-party apps that support it, plus the new frameworks and third-party apps that have come along for the ride. Microsoft has been working with Qualcomm on this for years, and this launch has ushered in a new era of Arm-based computing. It will be interesting to see how Apple responds in a few weeks at WWDC with its own AI push, but I don’t believe that Apple’s approach will be as mature. Yes, Apple does benefit from complete vertical integration, but the cooperation between Microsoft and Qualcomm seems so tightly synced that things aren’t far off from that for these Surface Copilot+ PCs.
I believe that in the wake of Copilot+, Apple is now the one playing catch-up; after all, it did launch the M4 iPad Pro somewhat prematurely with heavy AI PC counter-messaging. I believe that the AI race has only just begun for the PC and Mac, and that we’re very much in the early innings of this AI game, so it’s still far too early to declare any winners. We will likely continue to see companies try to compete with one another on AI TOPS performance, whether on the NPU, CPU or GPU. Copilot+ will hopefully help to resolve some of the anticipated fragmentation issues with x86 and Arm platforms and will hopefully smooth out the typical friction that we see with new Windows platforms. By setting hard expectations on NPU performance and minimum memory specs, Microsoft is helping developers understand what kind of hardware they can expect from Copilot+ PCs. I am excited to use these new Copilot+ PCs in the real world and compare them against the M4 iPad Pro as well as older PC platforms. Look for more of my coverage on this in the weeks and months to come.
Read the full article here
