LA Today WireZDNET’s key takeaways Microsoft and Nvidia launched a new PC platform this week. Those AI-focused machines are powerful but expensive. They also highlight the weaknesses in the Copilot+ PC brand.
Microsoft just rebooted its entire AI PC effort. If you watched this week’s announcement of the next generation of Windows laptops powered by the Nvidia RTX Spark , it’s hard to come to any other conclusion. Also: Microsoft’s new Surface Laptop Ultra finally has a starting price More from ZDNET These are powerful PCs.
My inner gadget freak looks at the new Surface Laptop Ultra and says, “I want one!” But my inner accountant is pointing to the extreme price tag and asking the obvious question: “What exactly are you going to do with this thing?” I’m afraid I don’t have a good answer. These new workstations are designed for AI-focused developers who want to work with large language models on local hardware instead of sending those workloads off to the cloud. Those developers can probably make a decent economic case for that purchase.
Running those jobs locally means you’re not paying hundreds of dollars a month to buy tokens from OpenAI or Anthropic or Microsoft. For those of us who aren’t spending a fortune every month on AI tokens, all that processing power is an unnecessary luxury. Welcome to AI PC 2.0 This isn’t Microsoft’s first push for an AI-centric PC platform.
The company unveiled the Copilot+ PC brand with a flourish in May 2024, highlighting a list of AI-powered features that would be exclusive to those devices. At the top of the list was a feature called Recall, which the company touted as a “photographic memory” for your PC. Recall turned out to be a privacy disaster for Microsoft .
At the time, security expert Kevin Beaumont issued this dire prediction : “I think they are probably going to set fire to the entire Copilot brand due to how poorly this has been implemented and rolled out. It’s an act of self harm at Microsoft in the name of AI...” He was, unfortunately, right. Also: Got a Microsoft 365 subscription?
Your storage is about to shrink Recall was supposed to ship in October 2024. It was supposed to be the signature feature of the new platform. Two years later, it’s still in preview, and at this week’s event Microsoft didn’t mention the feature at all.
In fact, outside of the official documentation for the Recall (Preview) feature, I can’t find any mention of it in any of Microsoft’s marketing communications from 2026. The Copilot+ PC specification sounded like a big deal at the time. To earn that label, a Windows 11 PC had to have at least 16 GB of memory (double the minimum for a regular Windows 11 PC) and it needed a neural processing unit (NPU) capable of processing a minimum of 40 trillion operations per second — 40+ TOPS.
Here’s how Microsoft touted those capabilities at its launch event: Copilot+ PCs leverage powerful processors and multiple state-of-the-art AI models, including several of Microsoft’s world-class [small language models], to unlock a new set of experiences you can run locally, directly on the device. This removes previous limitations on things like latency, cost and even privacy to help you be more productive, creative and communicate more effectively. The Nvidia Spark RTX devices are capable of up to 1 petaflop of AI performance.
That’s 25 times the computing power of the current generation of Copilot+ PCs. They don’t just meet the spec to be called a Copilot+ PC, they blow it away. If those Copilot+ PCs were AI PC 1.0, these new RTX Spark machines are absolutely AI PC 2.0.
The Copilot brand isn’t working As I noted last month, Microsoft’s big reboot of the Copilot client was engineered for the company’s enterprise customers, not for consumers. In fact, consumers have decidedly mixed feelings about AI in general. Pew Research, in a recent survey , found that 52% of respondents in the United States are more concerned than excited about AI.
Globally, only about 13% of people are more excited than concerned. So, maybe tying the brand of your flagship PCs to your AI chatbot isn’t such a good idea. At this week’s launch event, there was only one mention of the Copilot+ PC brand.
It went by so fast that if you blinked, you might have missed it. This was the only mention of the Copilot+ PC event at the joint Microsoft-Nvidia launch event this week Screenshot by Ed Bott/ZDNET Copilot+ PCs are thin and light. They do “everyday tasks” using local AI.
That’s not a sales pitch for power users. Contrast that modest pitch with this slide, listing just a few of the things Microsoft and Nvidia want you to know that an AI PC 2.0 can do. Screenshot by Ed Bott/ZDNET Microsoft’s partners are also running away from the Copilot branding.
Dell’s press release for its three new Nvidia RTX Spark-equipped machines doesn’t include the word Copilot at all. HP likewise avoided any mention of the Copilot brand in its announcement . At the start of 2026, it was obvious that those AI PCs weren’t selling to consumers, and global shortages of memory and storage driving PC prices higher are only making things worse.
What is an NPU good for, anyway? Part of the problem with Copilot+ PCs is that the “exclusive” AI-powered features on those machines aren’t all that compelling. Windows Studio Effects and live captions feel more like product demos than must-have capabilities.
The AI-powered image editing features in the Photos and Paint app are slightly more substantial, but does anyone really use them? And, of course, there’s Recall. I have an office full of PCs that meet the Copilot+ spec, equipped with the latest CPUs from AMD, Intel, and Qualcomm.
The point of this class of hardware is that the NPU is supposed to work AI magic on the computer, delighting you with its results and protecting your privacy by not sending your data to OpenAI or Anthropic via Microsoft’s client software. But I’ve regularly used Task Manager’s Performance tab to monitor what the NPU is doing, and most of the time the answer is “absolutely nothing.” The one exception is when Recall is turned on and I see a spike every few minutes as the system captures a snapshot, analyzes its content, and stores the result in the secure Recall database. Every other AI-related task is being processed in the cloud, and the local NPU is being ignored.
In the launch event for those fancy new RTX Spark-based AI PCs, I watched as Microsoft’s Kayla Cinnamon offered an excellent demo of what an agent can do for a professional developer. If you watch the video, you can see her Task Manager window, in which the high-powered Nvidia RTX SparkGPU is being pushed to the limit while the onboard NPU is doing ... absolutely nothing. Screenshot by Ed Bott/ZDNET It all makes one wonder why Microsoft made such a push to position the NPU as the most important component, when the GPU is the real workhorse for AI-related jobs.
Looking for AI PC 3.0 AI PC 1.0 wasn’t powerful enough. AI PC 2.0 is powerful but too expensive for the mainstream. So, as every Microsoft watcher knows, we wait for version 3.0.
That’s the one that will actually be a) affordable and b) capable of running the sort of local models that currently require the $4000+ PCs Microsoft showed off this week. One reason that isn’t possible on the current generation of Windows 11 hardware is that discrete GPUs are constrained by dedicated video memory. By contrast, machines built around Nvidia’s RTX Spark use unified memory, which gives the CPU and the GPU shared access to memory.
If that unified-memory design sounds familiar, it’s because you’ve seen it for years on Apple’s MacBook Pro and other high-end Macs. But even a midrange Mac Mini can do AI tasks that a Copilot+ PC can’t handle, thanks to the combination of Apple’s M-series CPUs and unified memory. With the Surface Laptop Ultra and equivalent PCs from its OEM partners, Microsoft is trying to win over developers who would otherwise choose a MacBook Pro.
But that’s a niche market. If Windows PCs are going to have a future, those technologies need to move down from workstation-class PCs into the mainstream. The question now is just how long we need to wait for AI PC 3.0.
Source: ZDNET
