Samsung puts on-prem language models on the fab floor

Samsung will deploy Mistral’s on-premises models across semiconductor manufacturing and engineering, a deal announced during a Franco-Korean summit in Paris. The models stay inside the fence. That is the story. A fab will not send yield data to a public endpoint, and a model that cannot see the data cannot help the yield.

The intended uses are unglamorous and high-value: development-cycle compression, process recipes, yield troubleshooting. Memory, logic design, and the foundry business are all in scope. If it works, it will look less like ChatGPT on a tool and more like a search engine over decades of failure modes. If it does not, it will join the long list of AI pilots that died in IT security review — which, in a fab, is often the correct review.