What Is an AI PC? On-Device AI Explained

An AI PC is a personal computer built to run AI models on the machine itself instead of sending every request to a data center. In practice that means a processor with a neural processing unit (NPU), a part of the chip designed for AI calculations. The NPU works alongside the usual central processor (CPU) and graphics processor (GPU). Companies draw the line differently. Intel describes the category by that three-part design. Microsoft sets numeric minimums for its Copilot+ PC label.

What does an NPU do, and what is TOPS?

Microsoft's developer guide describes an NPU as a chip built for demanding AI jobs. It gives real-time translation and image generation as examples. Microsoft's support page says an NPU works through large amounts of data at once. It says an NPU needs less power for AI tasks than a CPU or GPU. The developer guide ties that efficiency to longer battery life.

Intel describes a division of labor:

• the CPU for quick responses
• the GPU for high-throughput work
• the NPU for sustained AI tasks at low power

NPU capacity is advertised in TOPS, short for trillions of operations per second. Hardware alone is not enough. Microsoft notes that software has to be written to use the NPU. It also notes that models usually have to be converted, or quantized, into a lower-precision format before an NPU can run them.

What counts as a Copilot+ PC?

Microsoft's Windows 11 specifications page reserves the Copilot+ PC label for machines with:

• a processor with an NPU capable of 40 or more TOPS
• 16 GB of DDR5 or LPDDR5 memory
• 256 GB of SSD or UFS storage

As of October 6, 2026, the page lists these chips as qualifying:

• AMD Ryzen AI 300 and 400 series
• Intel Core Ultra 200V and 300V series
• Qualcomm Snapdragon X series

Standard Windows 11, by comparison, requires 4 GB of memory and 64 GB of storage. Microsoft's May 2024 launch announcement said the first Copilot+ PCs would be available from June 18, 2024, starting at $999.

The label controls access to software. Microsoft ties features such as object selection in Paint and super resolution in Photos to Copilot+ PCs. It limits some others to Snapdragon X models. Recall lets users search what they have previously seen on screen. It is still listed as a preview. After feedback about privacy and security, Microsoft said in June 2024 that Recall would be off by default. It said turning Recall on would require Windows Hello enrollment.

Are Macs AI PCs?

Apple does not use Microsoft's label, but its chips contain a comparable component called the Neural Engine. When it introduced the M4 chip in May 2024, Apple said the 16-core Neural Engine could perform 38 trillion operations per second. It claimed the Neural Engine was more powerful than the NPU in any AI PC at the time.

Apple's support page for Apple Intelligence, dated September 14, 2026, lists Macs with an M1 chip or later as compatible. It says the features can take up to 14 GB of storage on some Macs. Those are Macs with an M3 chip or later and at least 12 GB of memory. On other compatible models, it says, the features can take up to 8 GB. The same page says certain features rely on server-side models, so Apple's system mixes local and cloud processing.

Can I run AI locally on my laptop?

Often yes, and an NPU is not required. Microsoft's developer toolkit Foundry Local runs chat and speech-transcription models on Windows, Apple silicon Macs and Linux. It uses a GPU or NPU when one is available and falls back to the CPU. Microsoft says prompts and outputs are processed on the device, with the network used to download models.

LM Studio is a desktop app for running models locally. It recommends at least 16 GB of RAM. On Windows, it also recommends at least 4 GB of dedicated graphics memory.

The limit is model size. Microsoft says the Foundry Local catalog is restricted to models that have been compressed, tested on consumer hardware and kept small enough to distribute to end users.

What to watch

Supply and price are near-term pressures. IDC's preliminary figures show worldwide PC shipments fell 4.9% from a year earlier to 68.2 million units in the second quarter of 2026. By IDC's count, that was the first decline after nine straight quarters of growth. IDC blamed a memory shortage that it does not expect to ease until early 2028. It said vendors were raising prices. IDC added that interest in on-device AI keeps growing as cloud computing costs rise.

New designs are also arriving. TechCrunch reported that Nvidia announced a PC chip called RTX Spark in June 2026. Machines from Dell, HP, Lenovo, Microsoft and others were expected in the fall, TechCrunch reported. PCWorld reported in September that Nvidia had confirmed the first of them, a Lenovo laptop, for October. TechCrunch reported in May that Google had unveiled Googlebook laptops built around its Gemini assistant. Google opened pre-orders on September 21. It said the laptops start at $899, carry NPUs rated above 45 TOPS and would reach US stores on October 4.