Why a CPU and GPU are no longer enough – And why you now need an NPU.
To understand what is changing, it helps to take a brief look under the bonnet of modern clients. Something significant has happened there. Alongside the familiar CPU and GPU, a third component has joined the line‑up—the NPU (Neural Processing Unit). Its primary role is to run AI functions locally on the device.
Modern Copilot+ devices come equipped with an NPU as standard. They are not designed to provide a “better Copilot”, but to meet the technical minimum required—one element of which is an NPU capable of at least 40 TOPS. The distribution of tasks across the three processing units is, in principle, easy to understand:
Microsoft is not the only one driving this forward. Through DirectML—the API that allows developers to connect their applications to the NPU—more and more software partners are adding new AI capabilities to their products. Applications such as McAfee, Adobe Photoshop, Adobe Premiere Pro, CapCut and OBS Studio already make use of the NPU today. The infrastructure is in place and development is moving steadily. Much of this functionality already exists, some still in pilot phases, some widely available, but not yet fully activated or in its final, expanded form.
Why is that? Microsoft is waiting until enough devices with an NPU are available on the market. The software is effectively waiting for the hardware. Organisations that purchase devices without an NPU are relying on the CPU and GPU to absorb these workloads. That is possible, although neither efficient nor energy‑saving. Most importantly, it comes with noticeable performance losses.
Why traditional procurement thinking needs to evolve.
For decades, procurement revolved around a single question: “What level of performance do we need today?” Buyers could also rely on the fact that even entry‑level models offered more power than the devices purchased years earlier and still in use. Those who wanted additional headroom simply opted for a more powerful variant, which provided a reliable performance buffer for the years ahead.
Experts now agree that this mindset will not hold for the upcoming device cycles. The real question has become whether the devices we purchase today will be capable of supporting the software demands of the coming years. Choosing Copilot+ devices is not a luxury in this context, but a form of proactive risk management.
Organisations purchasing devices today are not investing for 2026 alone, but for 2027, 2028 and potentially the years beyond. Those planning without an NPU risk a scenario the client market has not seen in more than a decade—devices that may fail to reach their intended lifecycle because they can no longer keep up with the software demands.