For most of the past year, Microsoft has insisted that the future of AI on Windows belongs to Copilot+ machines. To access the most advanced local AI features, you supposedly needed a PC equipped with a dedicated Neural Processing Unit (NPU). That was the premise. Now, Microsoft seems to be changing the game.
According to the latest documentation, Windows 11’s local Language Model APIs are able to run on non‑Copilot+ PCs as long as they sport an Nvidia GeForce RTX 30‑series GPU (or newer) with a minimum of 6 GB of VRAM. At first glance, this appears to be a developer‑centric tweak, but in practice it could represent one of the most consequential adjustments to Microsoft’s AI‑PC strategy since the Copilot+ line debuted last year. More importantly, it forces us to revisit a lingering question from the start of the AI PC era: were NPUs really a prerequisite for these workloads?
The Copilot+ exclusivity era felt a bit odd
When Copilot+ machines launched in June 2024, Microsoft marketed them as the gateway to native AI experiences on Windows. To qualify, a device needed 16 GB of RAM, an SSD, and an NPU capable of delivering at least 40 TOPS of AI performance. The messaging implied that such specialised chips were essential for running AI locally. While that’s true for efficiency, it never painted the whole picture.
Anyone familiar with AI hardware already knew that GPUs were more than capable of handling these tasks. In fact, modern graphics cards often out‑perform NPUs when it comes to running language models and generative AI apps. That’s why enthusiasts have been relying on GPUs for years, whether they’re experimenting with small language models or image generators. Yet Windows’ built‑in AI features remained locked behind the Copilot+ badge.
This created a strange situation: a gaming rig with an RTX 4070 had more than enough horsepower to run AI models locally, but it couldn’t tap into Microsoft’s native AI framework because it lacked an NPU. Meanwhile, a thin laptop with a qualifying NPU could. The new change doesn’t erase that split entirely, but it certainly narrows it.
Microsoft may be laying the groundwork for AI beyond NPUs
The expanded Language Model APIs let developers harness local AI capabilities on supported Nvidia hardware. Microsoft states that these APIs now operate on non‑Copilot+ systems equipped with RTX 30‑series GPUs or newer, provided they have at least 6 GB of VRAM. The APIs run Microsoft’s compact on‑device language model, Phi Silica, which can be used for summarising text, rewriting content, turning prose into tables, formatting information, and generating responses to prompts.
Think of it as a lightweight, on‑device counterpart to the AI features people usually associate with services like ChatGPT. The key difference is that everything runs locally, not in the cloud. This matters for two reasons: privacy—your sensitive documents, notes, emails and drafts never leave the PC—and performance—local AI can respond instantly without waiting on remote servers, subscriptions, or an internet connection.
What’s intriguing is Microsoft’s distribution model. If an app requires Phi Silica, Windows can fetch the model via Windows Update and execute it locally on compatible hardware. In other words, the OS is starting to treat AI models as just another Windows component rather than a premium add‑on reserved for a specific class of PCs—a notable philosophical shift.
Is the Copilot+ exclusivity ending?
Don’t assume every AI feature is now coming to older Windows machines. Tools such as Recall, Click to Do, and some of Microsoft’s AI‑powered creative utilities still appear tied to systems with NPUs. The current broadened support applies only to the Language Model APIs, which focus on text‑based AI experiences.
History, however, shows that such walls rarely stay up forever. Once Microsoft proves that local AI runs smoothly on mainstream RTX hardware, it becomes harder to justify keeping certain AI experiences exclusive to NPUs. Developers won’t care whether the workload runs on an NPU or a GPU as long as it works well, and consumers certainly won’t. That’s why this update feels more impactful than the documentation change alone suggests.
For now, it’s just one API, but it also marks Microsoft’s first meaningful acknowledgment of a point many PC enthusiasts have been making all along: capable GPUs were never the bottleneck. If local AI can run flawlessly on millions of existing RTX‑powered PCs, the line between a “Copilot+ PC” and an ordinary Windows PC may soon become far less significant than Microsoft originally envisioned.