The node Samsung is putting in front of customers right now is not its most advanced one

The node Samsung Foundry is now putting in front of customers is neither 2nm nor 4nm. It is 5nm. With the 4nm line booked out through next year, the sales pitch has shifted to offering fabless customers who cannot absorb the cost of a finer node the generation below it, where yields are already settled.

AI demand drives 4Q25 global top 10 foundry revenue up 2.6% QoQ, Samsung gains share (TrendForce, 2026-03-12)

출처: trendforce.com

Who filled the 4nm line

The largest customer filling that line is not an outside fabless firm. It is the memory division of the same company. HBM is a stack of DRAM layers, and the board at the very bottom that organises the signals and sends them out is the base die. Samsung builds that board not on a memory process but on the foundry's 4nm logic process. So the better memory sells, the fuller the foundry line gets. On top of that, industry sources say, came volume from Nvidia's dedicated inference chip. Nvidia licensed Groq's technology in December 2025 and unveiled the Groq 3 LPU at GTC in March 2026, a chip that uses no HBM at all and instead keeps 500 megabytes of SRAM on the die itself. The board that holds up HBM and the chip that refuses to use HBM are coming off the same 4nm line.

GTC 2026 keynote: long live the inference king (EE Times, 2026-03-20)

출처: eetimes.com

That article describes the Groq 3 LPU as a chip paired with Rubin GPUs and handed only the token-generation work. It fills in the part the original post names and moves past.

The road the pushed-out customers take

The claim that Chinese and Indian fabless firms are coming to Samsung's 5nm to work around sanctions comes from a Samsung supply-chain partner, quoted anonymously. English-language reporting confirms only part of it. China's NIO already uses Samsung 5nm for its autonomous driving chip, and Samsung is reported to be in talks with BYD on autonomous driving SoCs. No Indian customer is named anywhere.

SF5A stands between cars and servers

The 5nm line is classified as SF5A, a process dedicated to automotive. Hyundai's ADAS chip targeting 2030 production, a system on chip from BOS Semiconductor, and an automotive application processor from Telechips are all booked on it. Automotive chips take years to qualify, and once they are in they stay for the life of the model. Putting server and AI orders in those slots lifts revenue now, but it shrinks the room that car volumes will need when they reach production a few years out. That is also why a full line does not immediately mean profit. Samsung's non-memory business is reported to have posted a loss of about 600 billion won in the second quarter of 2026. The split is whether this volume converts into share. If the outside customers on 5nm stay through the next generation and push Samsung's quarterly foundry share into the 8 percent range, filling the line will have meant moving position; if share sits near 7 percent, this run stays a season the company's own memory division paid for.