The TOPS line in a Core Ultra table looks like a speed. Microsoft uses it as a door. A Copilot+ PC, in Microsoft's definition, is a Windows 11 machine whose neural processor can do more than 40 trillion operations a second. Intel's laptop parts clear that line. Lunar Lake, the Core Ultra 200V, is rated at up to 48 TOPS, and Microsoft's own Copilot+ pages name that chip as one of the processors that qualify. Series 3, the laptop line launched in January 2026, was given an NPU of up to 50 TOPS in Intel's launch figures, which The Register reported. The desktop Series 2 part is rated at 13 TOPS on Intel's support page. That desktop chip does not open the door.

What sits behind the door is Windows, not a chat window you set up yourself. On a Copilot+ PC, Microsoft says its supported AI APIs always run on the NPU. The graphics and processor columns in that table are for machines that missed the bar. They are not a switch you can flip once the NPU qualifies. The developer list, on Microsoft's page dated October 8, 2026, is Phi Silica, text recognition, image sharpening, image description, pulling an object out of a photo, and erasing one. An image generator is on the list too, but it is optional and downloads later because of its size. Speech recognition is marked experimental. A developer API for live translation is marked not yet supported on that same page.

Phi Silica is the language model in the list, and Microsoft calls it a small one. It ships already installed on the NPU of a Copilot+ PC. The jobs are short: rewrite a paragraph, summarize a thread, read text out of an image. Microsoft's support note describes it as a Transformer-based small language model, tuned for the NPU, using speculative decoding so the text comes out faster at low power. It is not a model you pick by name in a chat application. Microsoft also says this model is on the way out. A replacement called Aion Instruct starts reaching Windows Insider machines in November 2026 and retail machines in January 2027, and Phi Silica comes off the machine when that happens.

Microsoft's product page names a wider set, and it hedges the dates. Recall, Click to Do, improved search, Live Captions with translation, and Windows Studio Effects are listed as Copilot+ experiences that use the NPU, delivered through 2026, varying by device and region, with updates still rolling out. The camera tools are the piece a person actually notices. Blur, eye contact, and framing can stay on for a whole call, because the NPU is the chip Microsoft expects to leave running while the graphics stay asleep. That low-power path is what the 40 TOPS bar is buying.

A model you choose is a second door, and Intel documents that one separately. OpenVINO's text pipeline for the NPU wants 4-bit weights, packed symmetrically, either INT4 or NF4. NF4 is supported on Series 2 NPUs, the Lunar Lake generation, and on later NPUs. The default prompt cap is 1024 tokens, with room reserved for at least 128 tokens of reply, and the context those two numbers add up to is the whole window. Intel says a Series 2 machine may need more than 16 GB of memory once a prompt passes 1024 tokens on a model past 7 billion parameters. The examples named in that note are Llama-2-7B, Mistral-0.2-7B, and Qwen2-7B. The guide prints no tokens-per-second figure, so this piece does not invent one.

The chat applications people already use are not that pipeline. Opening one and loading a model does not, by itself, hand the work to the Intel NPU. The NPU also has no private memory bus. On Lunar Lake the memory is soldered on the package, 16 or 32 GB, shared with the processor and the graphics. By the half-byte count this desk uses, a 7 billion parameter model is about 3.5 GB of 4-bit weights before any context. A 32 billion parameter model, Qwen3 32B in the size this desk has been using, is about 16 GB of weights on that count, which fills a 16 GB Lunar Lake chip before Windows takes a share. Series 3 lifts the top laptop to 96 GB. It lifts the NPU from 48 TOPS to 50. The memory is what can hold a larger file. The two extra TOPS do not change the Windows features, and they do not change which application can see the chip.

The desktop is the awkward machine. Its NPU misses 40 TOPS, so Phi Silica is not preinstalled there and the chip is not a Copilot+ PC. Microsoft does allow that same small model on a graphics card when the PC missed the NPU bar: an Nvidia GeForce RTX 30-series or newer with at least 6 GB, or an AMD Radeon RX 9060 or newer with at least 6 GB, and only with Developer Mode turned on. Intel's own Arc graphics are absent from that list. A desktop with a qualifying Nvidia card can host Microsoft's small model, and the card is doing the work, not the 13 TOPS NPU. A large chat model still wants that card's own memory. The NPU does not bring a bus of its own to substitute for it.

The Series 4 desktop rumor is an NPU 6 block at 74 TOPS, from a November 24, 2025 post that VideoCardz quoted. Partner sheets reported on April 12, 2026 name that block on the listed desktop dies. At 74 TOPS the chip clears Microsoft's line of 40, so Phi Silica and the image and text tools land on the NPU the way they do on the 48 and 50 TOPS laptops. The jump from the desktop part sold today is 13 TOPS to 74. A chat model you choose is unchanged by that jump, because this NPU shares ordinary memory and the rumor prints no token rate. Laptop Series 4 has no TOPS figure of its own in these notes, so the column uses the desktop number. Cores, the new socket, and the early 2027 window sit in the Series 4 piece next to this one.

Microsoft pins its own small language model to the NPU on a Copilot+ PC, and that PC does not get to run the same API on the graphics chip instead. The desktop NPU misses the bar. A qualifying Nvidia or AMD card can host that small model, with Developer Mode on. Neither fact is a rating for a 32 billion parameter chat model.