Nvidia brings atomic simulation to the chip fab

Nvidia is no longer content to sell the chips that train AI models. It now wants a seat inside the fabs that make them. At SEMICON West this week the company unveiled cuPotential, a software library that simulates how atoms and molecules behave during semiconductor manufacturing.

The tool joins CUDA-X, the collection of libraries that turns Nvidia’s graphics processors into engines for scientific computing. It is in early access, and Synopsys, one of the leading chip design software vendors, is already using it to simulate materials. Early tests, Nvidia says, run markedly faster than the methods they replace.

The library matters less than the company it keeps. Applied Materials models its process chambers in Omniverse, Nvidia’s digital twin platform. Lam Research plans to build another Nvidia library into its wafer-etch simulator before the end of the year. KLA and Teradyne, whose business is inspecting and testing chips, are speeding up their systems with Nvidia tools. Belgian research hub imec says it has refined its lithography processes using Nvidia AI.

These are the companies that build the machines every advanced chip depends on. That all of them now run on Nvidia software says a lot about where the company is heading.

From the lab to the clean room

A modern chip goes through hundreds of steps, depositing and etching layers just a few atoms thick. Much of that tuning has traditionally been done by trial and error on real wafers, which is slow and expensive. Simulating how each material will behave lets engineers rule out dead ends without burning silicon. If the simulation holds up, new processes reach production sooner.

That is where the impact could show up. For equipment makers, faster simulation could shorten development cycles and tighten ties with their largest customers. For foundries such as TSMC, Samsung and Intel, locked in a race over each new transistor generation, any shortcut in process tuning carries competitive value.

For Nvidia the stakes are different. Every company that builds its workflow on CUDA-X becomes a little more tied to Nvidia hardware, because switching would mean rewriting software that already works. That lock-in has long been its moat against AMD and against the in-house chips of the big cloud providers. Now it reaches into the supply chain that manufactures semiconductors, including its rivals’ products.

In our view, investors tend to judge Nvidia almost entirely on data center orders, and announcements like this one barely register. Yet the more third-party software depends on its platform, the steadier the demand for its hardware, even if the AI spending cycle cools. For the equipment makers joining in, any effect on their shares will hinge on whether these gains turn into foundry contracts or stay as trade-show demos.

One caveat: the performance gains on show are early test results reported by the companies themselves. The real test comes when these tools run every day on a production line.

Disclaimer: for information only. Trading and investing carry a risk of loss, including the loss of all capital. This is not investment advice.

Versión en español: Nvidia lleva la simulación atómica a las fábricas de chips

trader diary

What if the next piece reached you before it happened?

Monday to Friday, 1h35 before the New York cash open: the pre-market with levels, the calendar and the trading idea of the day. On Sunday, the weekly wrap and the week ahead. Free. No spam. One click to unsubscribe.

You will get an email to confirm (anti-bot). If it does not arrive, check your spam folder.