三星电子扩大为英伟达Vera Rubin平台量产Grok3与供应HBM4

Samsung Electronics (005930.KS) is establishing a commanding presence in Nvidia's next-generation AI computing platform, Vera Rubin. Its foundry division has begun mass production of Grok3, the language processing unit (LPX) that serves as Vera Rubin's AI inference accelerator, while its memory division is expanding supply of sixth-generation high-bandwidth memory (HBM4) for the next-generation Rubin graphics processing unit (GPU). The two pillars of Samsung's semiconductor business are broadening their alliance with Nvidia, the world's largest AI chip company, across the board.
Nvidia announced on the 24th, at the Hot Chips 2026 semiconductor conference held in California, that the Grok3 LPX had entered mass production. Grok3 is an inference accelerator that handles the computation of generating answers one word at a time after an AI reads long context. It generates AI responses up to four times faster than competing products, measured against Google's Gemma4.
The chip is actually produced at Samsung's foundry. Grok3 is being manufactured on the 4-nanometer process at the S5 line of Samsung's Pyeongtaek campus. Nvidia CEO Jensen Huang said at GTC 2026 in March this year that "Samsung is making Grok3."
Analysts say the sharp improvement in Samsung Foundry's yield played a decisive role in Nvidia's decision to formalize Grok3 mass production this time. According to industry sources, the production yield for Grok3 has improved roughly two- to threefold recently compared with April. It has now risen above 80%, the level considered to allow stable mass production. As yields, which had been low in the first half of this year, improved significantly, Nvidia was able to officially declare Grok3 mass production.
This reflects the growing maturity of Samsung Foundry's 4-nanometer process. Samsung Electronics has raised process stability by mass-producing the "base die," which handles computation in the bottom layer of HBM4, on the same family of processes. As a result, the Pyeongtaek 4-nanometer line has recently approached 100% capacity utilization. This trend is becoming a turning point for Samsung's foundry business, which has posted losses for an extended period.
The first customer for the Grok3 LPX is Nebius. It is a leading "neo-cloud" provider offering GPU-based cloud computing services specialized in AI model training and inference. Starting at year-end, Nebius plans to install the Grok3 LPX in its inference platform, the Nebius Token Factory, and become the first in the industry to launch a commercial cloud service using it.
Cooperation between Samsung Electronics and Nvidia does not stop at Grok3. Samsung's advanced memory also goes into the Vera Rubin platform that Nvidia begins shipping this month. Samsung Electronics is supplying HBM4 for the Rubin GPU. HBM is a key memory that enables a GPU to rapidly exchange the data it needs for computation, and it has been developed through six generations to date.
As Rubin GPU production ramps up faster than expected, Samsung Electronics, which has the world's largest memory production capacity, is expected to play an even greater role. Global investment bank KeyBanc recently raised its Rubin GPU shipment forecast by 9%, from 1.75 million units a year to about 1.9 million units.
Behind the potential increase in Rubin GPU supply is the expansion of HBM4 supply by memory companies. The industry expects Samsung Electronics' HBM supply this year to rise 20% from last year in gigabyte (GB) terms.
Samsung Electronics is also widening its cooperation with Nvidia by supplying a low-power DRAM module for the Vera CPU, called SOCAMM2, as well as NAND flash-based solid-state drives (SSDs).
An industry official said, "The cooperation between Samsung and Nvidia is moving from individual components toward a total AI hardware solution," adding, "Attention is on whether the long-term agreement (LTA) relationship formed in memory will extend to the foundry segment."