Do you need an AI PC?

The short version

Every laptop worth buying in 2026 has a sticker on it. Copilot+ PC is not a marketing flourish — it is a specific hardware bar, and a specific set of Windows features sits behind it.

On 2 June 2026, at Build, Microsoft announced that “Windows AI APIs are expanding beyond NPUs to CPUs and GPUs, bringing local AI experiences to a much broader set of Windows 11 devices”. That is a real change, and it was widely read as the end of AI-PC exclusivity.

Six weeks later, Microsoft’s developer documentation published the table showing what that expansion actually consists of. Of the nine Windows AI APIs listed, one runs on a GPU. Two run on a CPU. The remaining six — all of the image work — are NPU-only. And the one that made it to GPU requires you to turn on Developer Mode, join the Windows Insider Experimental Channel, and install a beta driver straight from NVIDIA or AMD.

So the honest summary is this: the expansion is real, it is aimed at developers rather than buyers, and if you want the consumer feature list Microsoft advertises, you still need the NPU. Whether that is worth paying for is a different question, and the answer for most people is not on its own.

1. What Copilot+ actually means as a spec

Microsoft’s published Windows 11 specifications set the bar precisely. A Copilot+ PC needs:

• A processor or SoC with a neural processing unit capable of 40+ TOPS (trillion operations per second)
• 16 GB of DDR5 or LPDDR5 memory
• A 256 GB SSD or UFS drive
• Windows 11, meeting all the standard minimum requirements

The named silicon is Qualcomm Snapdragon X series, Intel Core Ultra 200V and 300V series, and AMD Ryzen AI 300 and 400 series. If your laptop is older than that generation, no amount of RAM or GPU will get you the badge.

Two of those requirements are worth noticing for a reason that has nothing to do with AI. A 16 GB memory floor and a 256 GB storage floor are simply a decent laptop in 2026. Part of what the Copilot+ label has quietly done is drag the bottom of the market up, and that is a genuine consumer win regardless of whether you ever touch an AI feature.

2. What the badge gates

Microsoft lists more than twenty Copilot+ exclusive features on its specifications page. The consumer-visible ones include Recall (still labelled preview), Click to Do, improved Windows search, Live Captions, Windows Studio Effects, Agent in Settings, automatic super resolution, flexible Voice Access commands, the Snipping Tool’s perfect screenshot, and a cluster of generative tools in Paint (Cocreator, generative fill, object select, sticker generator) and Photos (image creator, relight, restyle, super resolution).

Some of those are processor-specific on top of being Copilot+ specific — certain Paint and Photos capabilities are noted as Snapdragon X only. Which is to say the badge is not even a single tier.

Recall deserves its own sentence, because it is the feature that sold the category and the one people are most wary of. It stores periodic snapshots of your screen locally so you can search your own past activity in plain language. Microsoft requires Secure Boot, BitLocker or device encryption, and Windows Hello sign-in before it will run, and it is opt-in. If that trade does not appeal, you have just removed one of the strongest arguments for the NPU.

3. What Build 2026 actually announced

The Windows Developer Blog post of 2 June 2026 is the primary document, and it is more measured than the coverage was.

The specific expansions named: the existing in-box small language model on capable GPUs, and Video Super Resolution and Speech Recognition on CPUs — all in public preview. Alongside that, Microsoft positioned Microsoft Foundry on Windows as the unified local AI platform, with Windows ML as the inference layer that “unlocks unmetered intelligence on Windows” across all silicon, and introduced Microsoft Execution Containers (MXC), an early-preview, policy-driven sandbox for AI agents on Windows and WSL.

There was hardware, too, but not the sort you buy at Officeworks: a Surface RTX Spark Dev Box with a petaflop of AI compute and 128 GB of unified memory, and an NVIDIA-powered DGX Station for Windows coming in Q4 2026, pitched at running trillion-parameter models locally. These are developer workstations. They tell you where Microsoft thinks local AI is going; they tell you nothing about which laptop to buy.

4. The table Microsoft published six weeks later

This is the part that reframes the announcement, and it is worth quoting the caveat verbatim. On the Windows AI APIs page, updated 15 July 2026, Microsoft notes: on a Copilot+ PC, supported APIs always run on the NPU, and the GPU and CPU columns “describe expansion to non-Copilot+ devices — they are not alternative backends you can opt into on a Copilot+ PC”.

The nine APIs and where they run:

• Phi Silica (the local language model) — NPU and GPU
• Speech Recognition — NPU and CPU (optional, removable)
• Video Super Resolution — NPU and CPU
• Text Recognition (OCR), Image Super Resolution, Image Description, Image Segmentation, Object Erase, Image Generation — NPU only

Every piece of image intelligence is still behind the NPU. That is not an accident of scheduling. Image models are where the NPU’s efficiency advantage is largest and where a CPU fallback would be least usable.

5. The GPU path has a lot of fine print

The one API that crossed over is the interesting case, because reading its requirements tells you how early this is.

To run Phi Silica on a GPU you need a GeForce RTX 30 series or newer, or a Radeon RX 9060 series or newer, with at least 6 GB of VRAM. Then:

• Developer Mode must be enabled in Settings.
• You need a Windows Insider Experimental Channel build (26300.8553 or later) and Windows App SDK 2.2.2-experimental9 or later.
• You need the latest driver installed directly from NVIDIA or AMD. Microsoft explicitly warns that drivers from Windows Update or your PC maker “may not be sufficient and can cause failures or degraded performance” — and that an OEM update can overwrite the driver you installed and break it again.
• The model is not preinstalled. It downloads on first use, several gigabytes, in the background through Windows Update.

And it is not the same experience once running. Microsoft’s own comparison says GPU inference loses prompt compression (which limits usable context length) and speculative decoding (which means lower tokens per second), with higher latency and higher power draw — the last of which matters a great deal on a laptop. Output quality is stated to be identical, because the weights are the same. Speed, battery and responsiveness are not.

None of that is a criticism of the engineering. It is a public preview doing what public previews do. But “AI PCs are no longer exclusive” and “a developer on the Experimental Channel with a beta driver can run one of the nine APIs on an RTX card” are very different sentences, and only one of them is on the record.

6. The model behind the badge is being replaced in November

Here is the fact most buying guides have not caught up with, and it is sitting at the top of Microsoft’s own documentation page.

Phi Silica is being replaced by Aion Instruct. The published timeline: a standalone sideloadable package for testing in early October 2026; rollout to Windows Insider Preview devices in October 2026, with both models present and the active one chosen by a controlled feature rollout; and in November 2026, Aion Instruct reaches retail devices and Phi Silica is removed. Microsoft also announced an Aion 1.0 Plan reasoning model at 14 billion parameters, shipping in-box on capable devices.

For a developer that is a migration plan with dates on it. For a buyer it is a useful piece of context: the on-device language model that ships with Windows is roughly eighteen months old and is being swapped out inside four months. If you are weighing a purchase on the strength of what local AI does today, weigh it lightly. The hardware bar has been stable since 2024. What runs on top of it has not.

7. What the NPU is genuinely good at

All of the above is a case against buying for the badge. It is not a case against the hardware, which does something a GPU cannot.

An NPU is built for the specific integer maths that quantised models need, at very low power. That matters for a category of feature that is not a party trick: things that run continuously. Live captioning a two-hour meeting. Background blur and gaze correction on every video call. Windows deciding, in the background, that a screenshot contains a phone number. A discrete GPU can do that work and will empty your battery doing it; Microsoft’s own transparency note makes exactly this point about the GPU path.

So the useful framing is not “does the NPU make AI possible” — a good GPU already did, and the cloud does more. It is “does the NPU make always-on AI cheap enough in watts to leave switched on”. That is a real capability, and it is the one thing on this list you cannot retrofit.

8. If you are buying a laptop in Australia right now

Stated plainly as our advice rather than anyone’s finding:

• Do not pay a premium for the badge alone. On a like-for-like machine, Copilot+ is a reasonable tiebreaker. As a reason to step up a price bracket, it is not.
• Buy the 16 GB and the SSD, not the TOPS. Those two requirements will do more for the next five years of your machine than any current AI feature will.
• Check which Copilot+ tier you are getting. Some Paint and Photos features are Snapdragon-only. Two laptops can carry the same sticker and not run the same features.
• Decide about Recall before you buy, not after. It is opt-in, it requires Windows Hello and disk encryption, and if you would never enable it you have removed the flagship reason for the category.
• If you already own an RTX 30-series or newer machine, you are not stranded. Foundry Local and Windows ML run on Windows 10 and later, and there are over twenty open models available through Foundry Local today. You will not get the in-box Copilot+ features, but local AI as a capability is not gated on a purchase.
• Do not let Windows 10 be the reason you panic-buy. Support for Windows 10 ended on 14 October 2025 and paid Extended Security Updates exist to bridge the gap. That is an upgrade deadline, not an AI-PC deadline — any supported Windows 11 machine clears it.

Where things stand, August 2026

The bottom line

The AI PC is a real category with a real hardware advantage, and the advantage is narrower and more specific than the marketing suggests. It is not that the NPU can run AI and other chips cannot. It is that the NPU can run AI continuously, on battery, which is why the features Microsoft gates behind it are the always-on ones.

Everything else is in motion. The expansion to GPU and CPU is genuine but early, and today it reaches one API of nine behind a developer setting and a beta driver. The in-box model is being replaced in November. The feature list is not even consistent across the silicon that qualifies.

Which leads to an unglamorous conclusion. Buy the laptop, not the badge. In 2026 nearly every good Windows machine happens to be a Copilot+ PC anyway — you will almost certainly end up with an NPU without paying extra for one, and that is the right way to acquire it. If a retailer is asking you to step up a price bracket for the AI features specifically, ask which ones, on which chip, and whether you would turn them on. That question is harder to answer than it should be, and that is the most useful thing to know before you spend the money.

Sources

Facts, dates and requirements above are drawn from these primary sources, captured 11 August 2026:

• Microsoft — Windows 11 specifications (Australian site), Copilot+ PC minimum requirements (40+ TOPS, 16 GB, 256 GB), named silicon, and the list of Copilot+ exclusive features including processor-specific limitations
• Windows Developer Blog — Build 2026: furthering Windows as the trusted platform for development (2 June 2026), the “expanding beyond NPUs to CPUs and GPUs” announcement, public preview scope, Windows ML, MXC, Surface RTX Spark Dev Box and DGX Station for Windows
• Microsoft Learn — What are Windows AI APIs? (updated 15 July 2026), the supported-hardware table, the note that APIs always run on the NPU on a Copilot+ PC, GPU VRAM and Developer Mode requirements, and model download behaviour
• Microsoft Learn — Get started with Phi Silica (updated 24 July 2026), the Aion Instruct replacement timeline for October and November 2026, GPU prerequisites and driver warnings, and the NPU-versus-GPU comparison covering prompt compression, speculative decoding, latency and power
• Microsoft Learn — Use local AI with Microsoft Foundry on Windows (updated 6 August 2026), the Foundry on Windows components, per-API device support, and Windows 10-and-later support for Foundry Local and Windows ML
• Microsoft Learn — Copilot+ PCs developer guide, what the NPU does, why it is more power-efficient than a CPU or GPU for AI work, and how Windows ML selects an execution provider
• Microsoft Learn — Extended Security Updates (ESU) program for Windows 10, the 14 October 2025 end-of-support date and the paid ESU bridge
• Microsoft — Windows 10 Consumer Extended Security Updates, the consumer enrolment path, which has separate terms to the commercial program