AMD to pack 50% more HBM into AI chips than Nvidia — a potential windfall for Samsung Electronics

AMD's aggressive push to load more high-bandwidth memory into its AI accelerators is sending a tailwind toward Samsung Electronics' HBM business. Analysts expect Samsung Electronics' share of the HBM market to rise sharply next year as AMD's AI accelerator shipments ramp up in earnest.

AMD said Wednesday that third-quarter sales would reach $13 billion, topping the market consensus of $12.52 billion, as surging demand for its chips from large-scale AI data center buildouts drove a stronger-than-expected outlook.

Second-quarter sales came in at $11.54 billion, up 50 percent from a year earlier and above market expectations. Data center revenue more than doubled year-on-year to $6.72 billion. Adjusted earnings per share of $1.66 also beat the consensus estimate of $1.62.

AMD shares have more than doubled this year as the company has emerged as a credible rival to Nvidia in the AI chip market. The stock surged 7 percent in regular trading before falling 9 percent in after-hours trading after the third-quarter revenue guidance came in slightly below some market expectations.

AMD's ecosystem is expanding rapidly. Last month, Anthropic agreed to purchase AMD's latest GPU, the MI450, at a total capacity of 2 gigawatts starting in the first half of next year. OpenAI, Meta and Oracle have also announced plans to expand their AI infrastructure with AMD.

On a conference call Wednesday, CEO Lisa Su was asked what steps AMD was taking to secure the memory it needs. "We have been working hand-in-hand with our memory partners for years to ensure a strong production and supply ramp for our business," she said. "We have very clear visibility into HBM allocation for the volume we will be shipping in 2027."

Samsung Electronics Chairman Lee Jae-yong raises a glass with AMD CEO Lisa Su ahead of a dinner at Seungjiwon, a Samsung guesthouse in Itaewon-dong, Seoul. [Samsung Electronics]

Samsung Electronics is widely seen as the biggest beneficiary of AMD's expanding ecosystem. When AMD CEO Lisa Su visited South Korea in March, Samsung Electronics' Pyeongtaek campus was her first stop. AMD also named Samsung Electronics as the preferred supplier of HBM4 for its next-generation AI accelerator, the Instinct MI455X. Su later met Samsung Electronics Chairman Lee Jae-yong for dinner at Seungjiwon in Yongsan-gu, Seoul, where the two discussed cooperation in AI chips and memory.

As AMD positions memory capacity and bandwidth as key differentiators in its battle against Nvidia, demand for Samsung Electronics' HBM4 is expected to grow further.

AMD recently unveiled the MI455X AI server GPU and its first rack-scale AI system, Helios, at its AMD Advancing AI 2026 event in San Francisco.

The MI455X carries 432 gigabytes of HBM4 — 50 percent more than the 288 gigabytes in Nvidia's next-generation AI GPU, Vera Rubin.

Su reiterated the memory advantage on the conference call. "Memory bandwidth and capacity are one of the strengths of AMD solutions," she said. "For large-scale AI models, you can get significant benefits from larger memory capacity, and that is reflected in the total cost of ownership."

UBS Securities' projections for HBM customer demand share by vendor. [Created with ChatGPT]

The HBM market was once almost entirely dependent on Nvidia's demand, but the customer base is diversifying quickly as big tech companies and AI chipmakers — including Google and AMD — step up investment in their own accelerators.

UBS projects Nvidia's share of HBM demand will fall from 60 percent this year to 41 percent next year, while AMD's share rises from 7 percent to 13 percent and Google's climbs from 18 percent to 30 percent.

Analysts say those shifts could allow Samsung Electronics to seize leadership of the HBM market next year, when HBM4 supply begins in earnest.

UBS Securities' forecast for HBM bit-share by vendor. [Created with ChatGPT]

UBS forecasts Samsung Electronics will capture a 41 percent share of the global HBM market next year, overtaking SK Hynix at 39 percent to claim the top spot. SK Hynix's HBM market share is expected to decline steadily, from 59 percent in 2025 to 48 percent this year and 39 percent next year.

The HBM market itself is on a steep growth trajectory. UBS projects total HBM capacity will expand 77 percent, from 33.14 billion gigabytes this year to 58.65 billion gigabytes next year.

AMD's HBM consumption is expected to lead all major AI chipmakers in growth over the same period, rising roughly 3.5-fold from 2.17 billion gigabytes to 7.52 billion gigabytes.

Memory pricing, however, remains a key variable. Su said AMD could adjust memory configurations for customers if larger memory did not deliver meaningful total-cost-of-ownership benefits for certain workloads.

"Many customers are very happy with the performance improvements that come with larger memory capacity, but for some workloads — such as mid-sized models — the benefits may not be as significant," Su said. "In those cases, we will adjust the memory configuration accordingly."

The approach is designed to retain cost-sensitive customers by offering more flexible HBM specifications while also addressing concerns that loading high-end HBM could weigh on AMD's own profitability.

go@heraldcorp.com