Samsung Electronics Co. has won long-awaited validation from Nvidia Corp. for its latest high-bandwidth memory, clearing a critical hurdle in the race to supply chips powering the next wave of artificial intelligence hardware.
The South Korean tech giant recently passed Nvidia’s qualification tests for its fifth-generation 12-layer HBM3E product, according to people familiar with the matter on Friday.
The approval comes about 18 months after Samsung completed development of the chip and follows a string of failed attempts to meet Nvidia’s demanding performance standards.
The milestone marks a symbolic recovery of technological credibility for Samsung in a field that has become one of the most strategically contested corners of the semiconductor industry.
Nvidia’s flagship B300 AI accelerator, as well as MI350 from Advanced Micro Devices Inc. (AMD), are among the systems set to deploy the high-capacity memory.
ALREADY SUPPLYING 12-STACK HBM3E TO AMD
Samsung had already shipped HBM3E 12-stack chips to AMD, but Nvidia, the dominant buyer of advanced memory for AI workloads, had remained out of reach.
Industry officials said the breakthrough owed much to a decision by Jun Young-hyun, Samsung’s chip business head and vice chairman, to redesign the DRAM core for HBM3E earlier this year, addressing thermal performance issues that had dogged earlier versions.
Volumes of 12-layer HBM3E chips to be supplied to Nvidia are expected to be relatively small, as Samsung is the third supplier to secure approval, following SK Hynix Inc. and Micron Technology Inc., sources said.
“For Samsung, the supply is less about revenue and more about pride,” said an industry executive. “Recognition from Nvidia means its technology is back on track.”
REAL BATTLEGROUND: HBM4
The real battleground, however, is already shifting to the next generation.
HBM4, the sixth iteration of high-bandwidth memory, is scheduled to debut in Nvidia’s next-generation graphics architecture Vera Rubin, the successor of the Blackwell AI chip, next year.
Samsung is aiming to close the gap with rivals by deploying its most advanced 10-nanometre-class DRAM (1c) and a 4 nm logic die produced through its foundry, compared with the 1B DRAM and 12 nm logic processes adopted by competitors.
Early performance indicators are encouraging for Samsung.
Nvidia has asked suppliers to push HBM4 data transfer speeds beyond 10 gigabits per second, well above the current industry standard of 8 Gbps.
Samsung has demonstrated 11Gbps, outpacing SK Hynix’s 10 Gbps and leaving Micron struggling to meet the requirements, according to people briefed on the situation.
Samsung plans to ship large volumes of HBM4 samples to Nvidia this month, with the goal of securing qualification at an early stage.
IN HBM4 TALKS WITH NVIDIA, BROADCOM, GOOGLE
In April, the Korean chipmaker said it is in talks to supply customized sixth-generation HBM4 chips to major AI chipmakers, including Nvidia, Broadcom Inc. and Google LLC.
Samsung said it could begin supplying HBM4 chips in large volumes to its clients as early as the first half of 2026.
Samsung is currently lagging behind its crosstown rival SK Hynix in the HBM race, having failed to secure deals for its advanced memory chips with Nvidia.
SK Hynix is the primary provider of the latest HBM chips to the US AI chip designer.
To catch up to SK Hynix, Samsung last year teamed up with its foundry archrival Taiwan Semiconductor Manufacturing Company Ltd. (TSMC) to jointly develop HBM4.
Analysts said success in the next round of HBM4 chip testing would allow Samsung to claw back market share in a memory segment central to AI computing.
Asked about its 12-layer HBM3E chip supply to Nvidia, Samsung just said it is “proceeding with schedules as planned with major customers.”
By Jeong-Soo Hwang, Chae-Yeon Kim and Eui-Myung Park
hjs@hankyung.com
In-Soo Nam edited this article.
