Samsung Foundry Deepens Nvidia Ties, Unveils 2nm HBM5 Development Plans

SAN JOSE — Nvidia CEO Jensen Huang visited Samsung Electronics' exhibition booth immediately after delivering his keynote speech at Nvidia GTC 2026 here on March 16.

Huang signed Samsung wafers used for Grok language processing unit (LPU) production and sixth-generation High Bandwidth Memory (HBM4) manufacturing, writing "Really Fast Grok" and "Amazing HBM4" respectively. The CEO also posed for photographs standing between Han Jin-man, President of Samsung's Foundry Business Division, and Hwang Sang-joon, Vice President of Memory Development.

Attendees described the scene as symbolic of the deepening partnership between the two companies, now extending from HBM to foundry operations.

Industry observers said Samsung has established itself as a core partner spanning all areas of next-generation AI accelerator platforms through its foundry division's entry into Nvidia's supply chain.

During the keynote, Huang unveiled plans to integrate GPUs, CPUs, storage, and LPUs into a unified design for the next-generation Vera Rubin AI accelerator platform. Samsung has been supplying HBM4, memory module SoCam2, and storage device PM1763 to Nvidia, and has now expanded the partnership by adding Grok3 LPU chip production.

The foundry business, long considered a weak point for Samsung, has secured another major contract with a big tech company, strengthening its turnaround prospects. Samsung signed deals with Tesla and Apple last year, demonstrating its competitiveness. The company is producing Grok3 at its Pyeongtaek facility using 4-nanometer process technology, with shipments expected as early as the second half of this year.

Samsung also unveiled the physical HBM4E product for the first time, emphasizing its commitment to securing technology leadership in the AI chip market. The product operates at 16Gbps with memory bandwidth of 4.0TB/s, surpassing HBM4's 11.7Gbps speed and 3TB/s bandwidth.

Vice President Hwang announced plans to maintain technological superiority in successor products including HBM5.

"The base die for HBM4E uses 4-nanometer process technology, and we plan to develop HBM5 and HBM5E using 2-nanometer process," Hwang said. "The core dies for HBM5 and HBM5E will use sixth and seventh-generation 10-nanometer class technology."

Samsung previously lagged behind rival SK Hynix through HBM3E but has positioned itself for a reversal by becoming the first to mass-produce HBM4.

"We will increase HBM production more than threefold compared to last year," Hwang said. "Our goal is to have HBM4 account for more than half of our total HBM output."

SK Hynix unveiled its HBM4 combined with Vera Rubin for the first time. The company also showcased collaboration results with Nvidia, including the GB300 GPU with HBM3E, the DGX Spark supercomputer featuring low-power LPDDR5X DRAM, and liquid-cooled eSSD storage.

SK Hynix is focusing on current-generation flagship products to solidify market share, as HBM4 and HBM3E are expected to account for a significant portion of AI chip demand. According to market research firm TrendForce, SK Hynix captured 59% of the HBM market last year, far ahead of Samsung's 20%, driven by HBM3E dominance.

SK Hynix also demonstrated "custom HBM" technology, considered key for HBM4E, signaling its readiness to compete with Samsung on next-generation technology. Custom HBM optimizes performance through tailored designs for specific AI chips.

"Starting with HBM4E launching next year, customized products optimized for various AI chips will emerge in earnest," an industry official said. "SK Hynix is also preparing for this."