Nvidia’s Jensen Huang Thanks Samsung at GTC, Highlighting Close Cooperation from Memory to Foundry

[The Financial News] Jensen Huang, CEO of Nvidia, singled out Samsung for special mention in his keynote at the annual developer conference NVIDIA GTC 2026 (NVIDIA GPU Technology Conference 2026), underscoring the close collaboration between the two companies.
During his keynote on the 16th (local time) at the SAP Center at San Jose in California, Huang introduced an inference-only chip and said, “Samsung is manufacturing the ‘Groq 3’ language processing unit (LPU) chip for us,” adding, “We are ramping up production as fast as we can right now. We are truly grateful to Samsung.”
Huang explained that this chip will be integrated into Nvidia’s next-generation artificial intelligence (AI) chip system, the NVIDIA Vera Rubin platform, and noted, “Shipments will likely begin in the second half of this year, probably around the third or fourth quarter.”
The Groq 3 LPU (Language Processing Unit) is designed to work alongside Nvidia’s Rubin GPU, splitting roles to boost inference performance and efficiency. Huang’s remarks confirmed that Samsung Electronics’ foundry business is producing the chip. As a result, Samsung Electronics can now showcase its collaboration with Nvidia not only in high-bandwidth memory (HBM) for Nvidia GPUs, but also in the foundry segment.
At the GTC venue, Samsung Electronics also used its exhibition booth to actively promote its memory partnership with Nvidia. The company publicly unveiled for the first time a physical chip of its next-generation HBM4E (High Bandwidth Memory 4E), known as seventh-generation HBM, along with a wafer of core dies used for stacking.
HBM4E, which targets sample shipments in the second half of this year, is expected to support a data rate of 16 Gbps per pin and a bandwidth of 4.0 TB/s. These figures surpass the 13 Gbps data rate and 3.3 TB/s bandwidth of the latest sixth-generation HBM4 (High Bandwidth Memory 4), which entered mass production and began shipping last month.
Samsung Electronics plans to accelerate development of HBM4E by leveraging the technological competitiveness it has built up through mass production of HBM4 based on its 1c DRAM process, a sixth-generation 10-nanometer-class node, as well as Samsung Foundry’s design capabilities for the 4 nm base die, the key component mounted at the bottom of the HBM stack.
The company says HBM4E’s performance is being realized through optimized collaboration that brings together all of its internal strengths, including memory, its own foundry, logic design, and advanced packaging technologies. Observers see Samsung’s decision to unveil HBM4E immediately after starting mass shipments of HBM4 as an effort to highlight its lead over other memory rivals such as SK hynix.
In particular, Samsung Electronics used an “HBM4 Hero Wall” at the exhibition to emphasize that this kind of competitiveness is a unique strength of an integrated device manufacturer (IDM).
Through video presentations, Samsung Electronics also showcased its hybrid copper bonding (HCB) packaging technology, which connects stacked chips and improves thermal resistance by 20% compared with thermo-compression bonding (TCB), while supporting ultra-high stacking of 16 layers or more.
Products highlighting the company’s partnership with Nvidia were prominently displayed as well. These included HBM4 used in Nvidia’s next-generation Rubin GPU, SOCAMM, a low-power memory module for servers designed for the NVIDIA Vera CPU, and the PM1763 enterprise SSD storage. With these, Samsung Electronics underscored that it can supply all the memory for the NVIDIA Vera Rubin platform.
Top executives will also take the stage at GTC. Song Yong-ho, head of the Samsung Electronics AI Center, is scheduled to deliver a special invited talk at NVIDIA GTC 2026 on the 17th, where he will present Samsung Electronics’ “total solution” vision for memory that underpins Nvidia’s AI infrastructure innovation.
soup@fnnews.com Im Soo-bin Reporter