The guide's advice is broadly reasonable, but its specs are presented more firmly than they deserve. One line in its laptop section could cost buyers real money. Below is what Mashable recommends, what Microsoft and the hardware vendors document, and where the two differ.
What Mashable recommends
The guide, written by Joe Hindy, covers five areas:
- Operating system: Windows, macOS and Linux can all run AI. It names Windows and Ubuntu as the two most popular choices. Windows wins for everyday use because Copilot is built in. Ubuntu suits people who want to build their own AI setup, but it is harder to learn.
- RAM and storage: It suggests a baseline of about 32GB of RAM and 1TB of storage, and says more is better.
- GPU: Despite the author's own preference for AMD, the guide recommends Nvidia because of tool compatibility. The RTX 5070 with 12GB of VRAM is the suggested starting point, with the RTX 5080 and RTX 5090 as better options. It says most guides treat 8GB of VRAM as the minimum.
- CPU with an NPU: It names Intel Core Ultra, AMD Ryzen AI and Snapdragon X. It notes that Snapdragon isn't a realistic choice for a self-built desktop.
- Laptops: It says the same advice applies to laptops, and claims most qualifying laptops let you upgrade RAM and storage yourself.
It's a sensible outline. Most of the problems are in the details.
Section summary: The guide's priorities are right for heavy local AI work: GPU and VRAM first, then RAM and storage, then the CPU. Its numbers are suggestions, not requirements.
Windows vs. Ubuntu: Copilot doesn't mean local AI
The guide favors Windows partly because Copilot is built in. That's true, but Copilot is mostly a cloud service. Having Copilot on a PC says nothing about whether that PC can run a local language model, an image generator or a coding assistant.
Windows does have its own tools for local AI. Microsoft's documentation describes Windows ML as its local inference engine, built on ONNX Runtime. It can run models on CPUs, integrated or discrete GPUs, and NPUs. Windows can install and update the hardware-specific acceleration components through Windows Update.
There's a version requirement. According to Microsoft, CPU and GPU support through DirectML works on all supported Windows versions. The hardware-optimized components for NPUs and specific GPUs need Windows 11 version 24H2 (build 26100) or later. If you're building an AI PC on an older Windows 11 release, update before you start comparing performance.
The guide's points about Ubuntu (more flexibility, a steeper learning curve) are the author's opinion, not tested results. Many popular open-source AI tools run on both systems. Choose the one you're comfortable maintaining.
Section summary: Windows has a documented local-AI platform, but Copilot doesn't prove anything about local performance. Run Windows 11 24H2 or later if you want the hardware-optimized acceleration paths.
GPU and VRAM: the desktop numbers are correct
The guide's VRAM figures for desktop cards are accurate. GamersNexus's coverage of Nvidia's announcement listed the 5090 at 32 GB, the 5080 at 16 GB, the 5070 Ti at 16 GB, and the 5070 at 12 GB.
Capacity isn't the only thing that matters. Bandwidth, meaning how quickly the GPU can move data in and out of its memory, matters too. Nvidia says the RTX 5070 has 12GB GDDR7 memory, and has 672 GB/sec of total memory bandwidth. Puget Systems pointed out that the RTX 5090's jump to 1.79 TB/s is an exciting development for AI workloads, which are often constrained by memory bandwidth. Puget also called the RTX 5090 theoretically the best consumer GPU for AI workloads that money can buy.
The RTX 5070 Ti, which the guide doesn't mention, is worth a look. It has the same 16GB as the RTX 5080. Puget Systems listed launch prices of $550 for the RTX 5070, $1,000 for the RTX 5080 and $2,000 for the RTX 5090. Tom's Guide listed the RTX 5070 Ti at $749 · 16GB GDDR7. Those were launch prices. Today's street prices may be quite different, so check current listings before you set a budget.
Two caveats on the guide's GPU claims:
- "Nvidia runs better than AMD" is too broad as stated. Nvidia has a real advantage in software support, but the gap depends on the tool. Microsoft's own documentation shows at least one built-in Windows AI feature supporting both brands.
- "8GB is the floor" depends on the task. Whether a model fits depends on its size, its precision and how compressed (quantized) it is. A heavily compressed small model needs much less memory than an uncompressed one.
Section summary: The desktop VRAM figures are accurate. Bandwidth matters as well as capacity, and the RTX 5070 Ti is a reasonable 16GB option that sits between the cards the guide names.
Laptops: same GPU name, less VRAM
The guide says its advice applies to laptops too. For GPUs, that's misleading, because Nvidia's laptop chips don't carry the same VRAM as desktop cards with the same name.
Tom's Guide's breakdown of the RTX 50-series laptop GPUs lists RTX 5090 · $2,899 · 24GB GDDR7 RTX 5080 · $2,199 · 16GB GDDR7 RTX 5070 Ti · $1,599 · 12GB GDDR7 RTX 5070 · $1,299 · 8GB GDDR7. Nvidia confirms that the RTX 5090 Laptop GPU has 24GB of GDDR7 memory, which is 8GB less than the desktop card.
| GPU name | Desktop VRAM | Laptop VRAM |
|---|---|---|
| RTX 5090 | 32GB | 24GB |
| RTX 5080 | 16GB | 16GB |
| RTX 5070 Ti | 16GB | 12GB |
| RTX 5070 | 12GB | 8GB |
So the guide's "12GB starter card" would only have 8GB in a laptop, which is the minimum the guide itself mentions. If VRAM is your main concern, compare the actual memory figure in the laptop's spec sheet, not the GPU's name.
The guide's claim that most qualifying laptops are user-upgradeable is also unsupported. In many thin laptops the RAM is soldered to the board, and storage slots vary from model to model. Before you buy, check the manufacturer's service manual or spec sheet to see whether the memory is socketed and how many M.2 slots there are. If you can't upgrade later, buy the 32GB configuration now.
Section summary: Laptop GPUs often have less VRAM than desktop cards with the same name, and many laptops can't be upgraded. Read the spec sheet before you buy.
NPUs: having one doesn't mean a PC meets Copilot+ requirements
The guide says a CPU with an NPU "sets you up for success." Microsoft's documentation is more specific. Its Copilot+ PC developer guide says many of the new Windows AI features need an NPU that can run at 40+ trillion operations per second (TOPS). It names chips that qualify, including the AMD Ryzen AI 300 series and the laptop-focused Intel Core Ultra 200V series.
An NPU below that line won't unlock those features. Intel's product brief for its Core Ultra desktop chips (Series 2) describes them as having more than 35 platform TOPS. That figure combines the CPU, GPU and NPU, so it isn't comparable to Microsoft's NPU-only requirement. When a spec sheet lists TOPS, check which number it is.
NPUs also don't help unless software uses them. Microsoft says applications must be specifically programmed to use the NPU. Windows ML picks the best available processor and falls back to the GPU or CPU if needed. You can check whether the NPU is doing anything: on supported devices, Task Manager shows NPU usage.
For a desktop with a strong Nvidia card, the NPU is a minor factor. Microsoft describes NPUs as built for low-power, continuous AI tasks, and says discrete GPUs generally give the best performance for image, video and generative AI. That's the work the guide is aimed at.
Section summary: NPUs matter most for low-power and Copilot+ features. On a GPU-heavy desktop, put your money into the graphics card first.
Microsoft's local language model also runs on gaming GPUs
Microsoft's AI documentation shows the GPU and NPU support overlapping. Its table of Windows AI features lists Phi Silica, its small on-device language model, as available on Copilot+ PC NPUs. It also lists it for non-Copilot+ PCs with Nvidia GeForce RTX 30-series or newer GPUs, or AMD Radeon RX 9060-series or newer, as long as the card has at least 6GB of VRAM.
GPU support comes with conditions: Developer Mode must be turned on, you need the latest driver from the GPU maker, and the current GPU version is marked experimental. Microsoft also says Phi Silica is being replaced by a new on-device model called Aion Instruct. It starts reaching Windows Insider devices in November 2026 and regular devices in January 2027, when Phi Silica will be removed. Downloaded on-device models can be managed under Settings > System > AI Components.
A practical checklist for building or buying
- Start with the workload. Chatting with a small model, generating images, editing video and developing models need very different hardware.
- Check the model's requirements for memory and supported software before choosing a GPU.
- Prioritize VRAM and bandwidth. On desktops, the guide's RTX 5070, 5080 and 5090 tiers are a sensible ladder, and the 5070 Ti is a 16GB option between them.
- On laptops, look up the actual VRAM figure. Don't assume it matches the desktop card with the same name.
- Treat 32GB of RAM and 1TB of storage as a sensible starting point, not a guarantee. Model files can be several gigabytes each.
- If you want Copilot+ features, check the NPU's own TOPS rating. Copilot+ features need 40+ TOPS from the NPU alone.
- Run Windows 11 24H2 or later for Windows ML's hardware-optimized acceleration.
The bottom line
Mashable's guide gets the main points right: browser AI and local AI are different, the GPU and its memory matter most, and Nvidia is the safe choice for software compatibility. Where it falls short is in treating "AI PC" as one standard configuration. Copilot doesn't make a PC good at local AI, and a CPU with an NPU doesn't automatically meet Copilot+ requirements. An "RTX 5070" laptop has 8GB of VRAM, not 12GB. Work out what you want to run, check the actual spec sheets, and buy as much VRAM as your budget allows.