Every GeForce RTX 30 graphics card Nvidia just launched is built on Samsung 8nm, and that is the biggest customer win Samsung’s foundry has landed in years. Ampere is Nvidia’s flagship architecture, the GPUs that power gaming rigs and increasingly the data centers training AI, and Nvidia chose to build them in Korea rather than at TSMC. The interesting part is which node it chose: not Samsung’s newest, but an older, cheaper one.
8nm is not a step forward from the 7nm EUV process Samsung was bragging about last year. It is a refinement of the older 10nm line, a non-EUV node, a half-step that exists precisely because not every chip needs the bleeding edge or wants to pay for it. Nvidia looked at Samsung’s best EUV process and Samsung’s mature 8nm and picked the mature one. For a company shipping enormous GPU dies by the million, a node that is cheap and available and yields well beats a node that is newer and tighter and harder to get.
This is a different kind of win than the Snapdragon flagships, and in some ways a smarter one. TSMC is sold out. Apple buys up its leading-edge capacity years in advance and everyone else fights over what is left, even after TSMC brought EUV online at the front of its 7nm line. Samsung’s pitch to Nvidia was not we are more advanced, it was we have capacity and we will make you a price, and for a gigantic order of big GPU chips that pitch won. Being the available alternative to a sold-out leader is a real business, maybe a more durable one than chasing every node record.
The split inside Nvidia’s own order tells the real story. The data-center Ampere, the A100 that goes into AI training racks, is built at TSMC on 7nm. Only the consumer GeForce dies came to Samsung 8nm. Nvidia kept its highest-margin, most thermally demanding silicon on the better process and handed Samsung the high-volume gaming parts, where cost matters more than the last increment of efficiency. That is the shape of the win, large and real and also carefully bounded. Samsung sits somewhere around the high teens of foundry revenue in 2020 while TSMC holds more than half and keeps pulling away, and a deal like this is how you stop the gap widening, not how you close it.
Landing Nvidia does two things. It fills Samsung’s 8nm lines with high-volume work, and it puts a marquee logo on the customer list that is not a phone company, proof Samsung can serve the GPU and eventually the data-center market that is starting to matter more than mobile. After years defined by Apple and Qualcomm, a giant Nvidia order broadens what the foundry is for. The 8nm node will be remembered less for its technology than for who it convinced.
Ampere runs hot and pulls a lot of power, and gamers are already noticing. Some of that is the architecture and some of it is the node, because 8nm is a generation behind what these big chips could ideally use. Nvidia accepted the tradeoff for cost and supply, and it mostly works, but it plants a question: if Samsung only won this because TSMC was full and expensive, what happens next generation, when Nvidia has more leverage and TSMC has more room. Winning an order on price means living with the day someone else offers a better one.
For right now, in late 2020, this is exactly the customer Samsung needed, proof the foundry can land the biggest names in chips outside of phones. Perso, I would not assume Nvidia stays. The relationship looks like one of convenience more than conviction, the kind of deal that holds as long as TSMC is sold out and unravels the moment it is not. But Samsung will take it, and it should. Filling your most profitable lines with Nvidia silicon while you fight to get EUV yielding is a good place to be standing. The question is whether the marquee customer is a foundation or a tenant who already has one foot toward the door.