Beyond the Node

The short version

Smaller transistors are no longer the only way forward

For fifty years, faster chips meant smaller transistors. That is why a handful of factories now make nearly every advanced AI processor. This paper asks what happens if the transistor stops being the best place to look for progress.

1. The leading edge has become a members-only club

A 2 nm wafer is reported at about $30,000 and the factory behind it at about $28 billion. Designing a first chip costs $30–100 million or more. Very few companies can pay that.

2. Most of a chip’s energy isn’t spent computing

In an ordinary processor, only about 5–10% of the energy goes into arithmetic. The rest goes into deciding what to do next and moving data around.

3. A different kind of chip avoids both costs

Build it as a grid of small tiles, each with a little arithmetic and a little memory, and plan the flow of data in advance. Companies already ship chips like this on 12–14 nm.

4. On the right jobs, that can make up for older transistors

The paper shows the maths, the evidence and the places it could fail.

What we are not claiming. That a 65 nm chip beats a 3 nm GPU at everything. The claim is narrower: for specific jobs, mainly AI inference, a well-organised older-node chip can win on cost and energy.

Inside the paper

What you will find

1. The short version
2. Why the leading edge became a members-only club
3. Where the energy really goes
4. The idea: a chip built from tiles
5. Does the maths work?
6. Risks and a roadmap
7. About the research: method, evidence grades and references