Samsung Electronics to Triple HBM Production, Over Half Allocated to HBM4 [GTC 2026]

"Applying 4nm Process to HBM4, Wafer Evaluation Underway"

"Developing 2nm Base Die Starting with HBM5"

"Producing 'Grok' 3 in Pyeongtaek, Demand for 4nm Expected to Rise"

Samsung Electronics has decided to more than triple its high bandwidth memory (HBM) production this year compared to last year, allocating over half of that increase to its latest product, HBM4. The company plans to apply a 4-nanometer (nm) process to HBM4 and accelerate the transition to advanced manufacturing, introducing a 2nm process to the base die starting with its 8th-generation high bandwidth memory (HBM5).

Sangjun Hwang, Vice President and Head of Memory Development at Samsung Electronics' Device Solutions (DS) division, spoke to reporters at the company’s booth during NVIDIA’s annual developer conference, 'GTC 2026,' which opened in San Jose, California, on March 16 (local time). In response to questions about HBM production, he said, "We are ramping up production rapidly and there are no significant issues," confirming the company’s aggressive expansion plan.

Hwang stated, "Our goal is for HBM4 to account for more than half of our total HBM output," adding, "In situations where supply is slightly constrained, prioritizing premium products is beneficial for the entire industry." He explained that, given the global memory supply shortage, Samsung has no choice but to coordinate supply volumes according to the needs of strategic partners and those requiring mass-produced products.

NVIDIA CEO Jensen Huang is taking a commemorative photo at the Samsung Electronics booth at the opening of 'GTC 2026' in San Jose, USA, on the 16th (local time). From left, Sangjun Hwang, Vice President of Memory Development at Samsung Electronics; Jensen Huang, CEO of NVIDIA; Jinman Han, President of the Foundry Division at Samsung Electronics. Photo by Joint Press Corps.

He specifically mentioned plans to expand total HBM supply by about three times compared to the previous year. He added that HBM4 is currently undergoing evaluation at the wafer stage. He also emphasized, "The competition in HBM is no longer just about capacity; it’s now a competition of performance and power efficiency," noting that "securing power efficiency is crucial, especially given the significant power constraints in data centers."

Samsung Electronics also announced that for HBM5, which follows the newly unveiled HBM4E, it will apply a more advanced 2nm base die process compared to the existing 4nm process. Additionally, the core dies used for stacking in HBM5 and HBM5E will utilize 10nm-class 1c (6th generation) and 1d (7th generation) chips. "While there are cost concerns, leveraging advanced processes is inevitable to align with the product goals and concepts that HBM aims for," he explained.

Jinman Han, President of the Foundry Division at Samsung Electronics, also attended GTC for the first time in two years since 2024, presenting the company’s foundry technologies at its booth. He particularly focused on the next-generation inference AI accelerator, 'Grok 3 LPU,' which drew attention after NVIDIA CEO Jensen Huang mentioned during the GTC 2026 keynote that "Samsung Electronics is manufacturing it."

Han explained, "We are currently producing the Grok 3 LPU using a 4nm foundry process at our Pyeongtaek facility," and noted, "We have received more orders this year than we initially expected." He added, "Since we are also manufacturing the base die for Samsung HBM4 with a 4nm process, we expect demand for the 4nm process to increase significantly going forward."

The Grok 3 chip is an ultra-large chip with a die area of over 700 square millimeters, and the package structure incorporates a single chip per package. Due to its large size, wafer demand is expected to increase substantially if AI inference demand grows in the future.

Regarding why NVIDIA chose Samsung Foundry to produce the Grok 3 LPU, Han explained, "Since 2023, before NVIDIA acquired Groq, we and Groq were already engaged with each other, and our engineers directly assisted with the design." He added, "When NVIDIA and Groq began collaborating, we wondered whether they might use another foundry, but after evaluating our chip’s performance, they must have concluded there was enough potential. Our 4nm process is by no means inferior."

Asked when revenue from the Grok 3 LPU would begin, he replied, "Mass production will start at the end of the third quarter or early in the fourth quarter," and added, "We’ll have to see the market response, but I believe demand for the Grok 3 LPU will increase significantly next year."

This content was produced with the assistance of AI translation services.