NVIDIA's new Windows laptop chip is built to run AI at home

What happened

On 16 October, NVIDIA announced that its RTX Spark Windows laptops are on sale. The heart of each machine is a single chip that pairs a graphics processor (GPU) with a central processor (CPU). Think of the GPU as the part built for lots of calculations at once, and the CPU as the part that runs the rest of the computer.

The selling point is AI. NVIDIA says the machines are for running "large language models and AI agents locally", meaning on your own laptop rather than on a company's servers. Microsoft is a launch partner and sells its own version, the Surface Laptop Ultra. NVIDIA also offers a 74-billion-parameter model called Nemotron 3.5 Bolt, built for tool use and step-by-step tasks, on supported configurations.

NVIDIA also describes a feature called PAIR, which lets one RTX Spark machine serve AI to other Windows, Linux and macOS computers on the same home network.

These are NVIDIA's own claims. The company also says features are offered on a "when-and-if-available basis", which means they may change. We have not yet seen independent tests.

Why it matters

Running AI on your own machine has three benefits. Your questions stay on your device. The tool keeps working without a connection. And you are not paying for every request. The catch has always been memory. A large model is a very large file of numbers, and the computer has to hold it all at once.

That is why the 128 GB figure matters. NVIDIA calls the memory "unified", which means the GPU and CPU share one pool rather than copying data back and forth. A pool that large is what makes big models practical on a laptop.

The idea behind it

Three ideas from our courses explain what NVIDIA is doing.

Low precision. NVIDIA's headline number, "1 petaflop of FP4 AI performance", uses FP4, a very compact way of storing numbers. Each number takes 4 bits instead of 16 or 32. That is like writing a recipe with fewer decimal places: you lose a little accuracy, but you can fit far more of it on the same counter. Our lesson Why Low Precision explains the trade-off, and Making Low Precision Work shows how models keep their accuracy at these sizes.

The KV cache. When a model writes a long reply, it keeps a running memory of everything it has read so far. That memory grows with the length of the conversation and takes up space. Our lesson The KV Cache: Speed at a Price explains why. Large shared memory gives that running memory room, so long chats and agent tasks do not run out of space.

Inference. Running a trained model to answer a question is called inference, and it is the part you pay for every time you use AI. Making Inference Affordable covers the tricks that make it cheaper, and the LLM inference simulator lets you watch the memory and speed trade-offs change.

One calculation is worth doing yourself. A model's name often gives its size. Nemotron's name says 74 billion parameters, and NVIDIA lists it with a 24 GB memory requirement. That is far less than 74 billion numbers at 16 bits each would need, which is why low precision matters. We have not checked the exact format NVIDIA uses, so treat that as a question to ask rather than a fact.

What to watch

Three questions will decide whether this matters beyond the launch.

• Independent speed tests. NVIDIA's game and AI figures are its own. Reviewers will show what these laptops do in practice.
• Battery and weight. NVIDIA calls the design slim and says it has all-day battery life, but gives no hours and no weight. Those numbers will matter for anyone who wants to run AI on the move.
• Whether PAIR works in ordinary homes. Sharing one machine's AI with other devices is a clever idea. It will need to work reliably with a home router, not just in a demonstration.

The next time you see a new laptop advertised for AI, ask three things: how much memory it has, how many bits its numbers use, and what its AI figure is measured on. Those three answers tell you more than the headline does.