The Samsung Mistral partnership announced on September 9 in Paris comes with a €3 billion check attached, and most of the coverage led with the money. But the word that actually explains the deal is buried in the third paragraph of Samsung’s own announcement: on-premises.
On-premises means the AI runs on Samsung’s own servers, inside Samsung’s own buildings. Not in a data center owned by someone else. Not through an app that sends your question somewhere and sends an answer back. The model lives where the chips are made, and nothing it reads leaves the property.
That sounds like an IT detail. It’s the whole trade.
Samsung already knows what happens without it. In March 2023, engineers in the semiconductor division were cleared to use ChatGPT at work. Within weeks, three of them had pasted confidential material into it — source code from a chip database, code for a piece of fab equipment, and the notes from an internal meeting. Anything typed into that kind of service goes to a server the company doesn’t control. Samsung banned generative AI tools across the company that May.
So when Samsung went looking for an AI partner three years later, it wasn’t shopping for the smartest model. It was shopping for one that would agree to be locked in a room.
That narrows the list fast. American AI labs largely sell access to models they host themselves — that’s the business. A company willing to hand over its system to run entirely inside a customer’s walls is making a different bet about what it’s selling. Mistral has built around that bet.
And narrowing the list is how a customer turns into an investor. Samsung didn’t just sign a contract; it led Mistral’s funding round, taking an equity stake alongside it, with the announcement made during a Korea–France state summit. The Samsung Mistral arrangement gives each side something the other can’t easily buy elsewhere: Samsung gets a model it can point at process data that has never left the fab, and Mistral gets a chip factory to tune against.
Samsung has made this kind of call before — choosing to license an outside component rather than build one in-house when the clock mattered more than the credit. The pattern is the same. What’s different is that the thing being brought in-house isn’t a part. It’s a system that will sit next to the most sensitive data the company has.
The stated jobs are unglamorous: spot defects earlier, tune process and equipment settings, shorten development cycles. Below two nanometers, small yield differences decide who is profitable, and Samsung intends to extend the Samsung Mistral work across memory, foundry, and logic.
None of which has happened yet. Every claim here is a plan, and plans in semiconductor manufacturing are judged by one number, reported quarters later, which is yield. Until then, what Samsung has bought is a condition.
My take: Samsung didn’t buy a model. It bought a room the model can’t leave.
Not advice. Just how I see it.