Does 2 nm mean the transistors are 2 nanometers wide?

No. Node names stopped matching a physical feature of the transistor years ago. A 2 nm label means a newer generation, not a literal 2 nanometer measurement.

Is a smaller process node always better?

Usually it brings more density or better efficiency, but cost, yield and chip design also decide real performance.

What is Process node? AI glossary

A process node is the name for a generation of chip manufacturing technology. When you hear that an AI chip is built on "3 nanometer" or "2 nanometer," that is the process node. A newer node generally lets a factory fit more transistors on a chip, run them faster, or run them on less power.

The number is a label, not a ruler

It is tempting to read "3 nm" as the size of a transistor. It is not. Decades ago the node name did track a real feature, such as the length of a transistor's gate. Today the name no longer corresponds to any single physical feature of the chip. Wikipedia's article on the 2 nm process states that the label has no relation to an actual physical measurement such as gate length or metal pitch.

So think of a node name the way you think of a phone model number. A higher generation is better, but the number itself does not tell you the exact size of anything. It also means nodes from different companies are not directly comparable by name alone.

How a new node works

Each node is a package of manufacturing steps and transistor designs. Moving from one node to the next usually brings some mix of three gains:

Density: more transistors in the same area.
Speed: the chip can run faster at the same power.
Efficiency: the chip uses less power for the same work.

Sometimes the change is a new transistor shape, not just smaller features. The move from the 3 nm generation to the 2 nm generation, for example, is widely described as a shift from FinFET transistors to gate-all-around nanosheet transistors. Foundries publish their gain targets for each node, and those are targets, not guarantees.

Why it matters if you use AI

AI models run on specialized chips, and the best of them are made on the newest nodes. More transistors and better efficiency mean more computing per chip and more computing per watt. That feeds into how fast a lab can train a model and how cheaply it can serve inference to you.

A chip is more than its node, though. Memory matters as much as raw compute for large models, which is why designers pair processors with high-bandwidth memory. The size of a model's parameter count decides how much memory it needs in the first place.

An example

Say a company announces its next AI accelerator is "built on a 2 nm process." That tells you the chip uses the newest manufacturing generation from its foundry. It does not tell you how fast the chip is, how much it costs, or how it compares to a rival's chip on another company's "3 nm" node. For that you need measured results on real workloads.

Only a handful of factories in the world can make chips on the leading nodes, which is why a node's availability can limit how many AI chips exist at all. When you see a node in the news, read it as "which generation of manufacturing is this chip using," and look for tested performance before drawing conclusions.