AI Computing: The Hardware Behind the Hype

AI has gone from buzzword to daily tool. But “an AI laptop” can mean wildly different things depending on which kind of AI you actually use. Here's a plain-English map of the AI landscape — and, more importantly, what hardware each type really needs.

1. Cloud AI assistants — the AI most people use

Tools like Microsoft Copilot, Claude and ChatGPT do their heavy thinking on remote servers, not on your machine. You're really using a web app, so the hardware demand on your end is light: any modern laptop with a decent screen and a reliable internet connection runs them perfectly. If cloud assistants are your main use of AI, don't overspend chasing AI-specific silicon you won't touch — put the money into RAM, a fast SSD and a good display instead.

2. On-device AI and the NPU

The newest laptops include a Neural Processing Unit (NPU) — a dedicated AI accelerator alongside the CPU and GPU. NPUs handle smaller AI tasks locally and efficiently: live captions, background blur and noise removal on calls, image clean-up, on-device assistants and the “AI PC” features built into modern Windows. They sip power doing it, which helps battery life. If you want your laptop itself doing AI work without leaning on the cloud, an NPU is the spec to look for.

3. Running LLMs locally — the demanding one

This is where hardware really matters. Open-source large language models — Llama, Mistral, Qwen, Gemma and many more — can be downloaded and run entirely on your own machine, with no cloud and no subscription. The appeal is privacy, control and zero ongoing cost; the catch is that these models are hungry. What they need:

• Lots of memory. Model size is everything. Small models (around 7–8 billion parameters) run comfortably in 16GB of RAM; larger, smarter models want 32GB or more. For local AI, RAM is the single most important spec.
• A capable GPU — and its VRAM. A dedicated graphics card dramatically speeds up local models, and its video memory (VRAM) sets how large a model you can load. More VRAM means bigger, more capable models running at usable speed.
• Fast NVMe storage. These models are many gigabytes each. A quick NVMe SSD means they load in seconds instead of minutes, and gives you room to keep several on hand.
• A strong, well-cooled CPU. Even with a GPU, the processor does real work — and sustained AI workloads run hot, so cooling that prevents throttling keeps performance steady.

4. Open source vs closed, local vs cloud

It's worth understanding the trade-off. Closed cloud models (Claude, ChatGPT, Copilot, Gemini) are the most capable and need almost nothing from your hardware, but your data leaves your device and you pay per use or by subscription. Open-source models run locally are private, free to run and work offline, but you provide the horsepower and the very best frontier models still live in the cloud. Many people now use both — cloud for the hardest tasks, a local model for private or everyday work.

What to actually buy

If you mainly use cloud assistants, buy a well-balanced laptop — good RAM, fast NVMe storage, a lovely screen — and don't pay an AI premium. If you want to run models locally, prioritise memory first, then a capable GPU with generous VRAM, fast storage and cooling that holds its speed under load. Not sure which camp you're in? Our team is happy to talk it through and point you to the right machine.