Samsung Electronics is firmly demonstrating its technological capabilities in the high bandwidth memory (HBM) market and is expanding supply to major global customers such as NVIDIA. Samsung, which at one point lagged in this area, has achieved a technological turnaround by receiving NVIDIA’s highest evaluation for its next-generation HBM4. With HBM sales projected to more than triple in 2026, attention is focusing on whether Samsung can replicate its strategy of overwhelming technological leadership in semiconductors in the HBM and customized AI chip markets.

The emergence of HBM and its inevitability in the AI era

With the rapid spread of generative AI, the structure of data center servers is changing dramatically. The architecture is shifting from conventional general-purpose processor-based systems to a combination of GPUs (graphics processing units) and high bandwidth memory (HBM). HBM is an advanced memory technology manufactured by vertically stacking multiple layers of conventional DRAM like a building.

The core technology of HBM is known as TSV (Through Silicon Via), which drills vertical conduits inside the semiconductor to transmit data rapidly. While conventional DRAM exchanges data only horizontally, HBM is interconnected three-dimensionally, including vertically, which dramatically increases data transfer speed (bandwidth). This is a key element in meeting the requirement to process massive volumes of data quickly during AI model training and inference.

HBM products are categorized by generation: 4th-generation HBM3 (2021), 5th-generation HBM3E (2023), and the latest 6th generation, HBM4. NVIDIA’s GPUs mount multiple HBM devices together to perform ultra-high-speed AI computation. For example, NVIDIA’s next-generation AI accelerator Rubin (to be launched in the second half of 2026) will mount multiple high-performance HBM4 chips to deliver ultra-high-speed AI computation.

Samsung’s HBM development failure and lessons learned

Interestingly, Samsung Electronics was among the first companies to develop HBM technology. However, in 2019, the management of Samsung’s semiconductor division decided to downsize the dedicated HBM team, judging that the HBM market would not be large enough. This contrasted with latecomer SK hynix, which focused on HBM development for survival, even in what was then a small market.

Management’s rationale at the time was that due to the specificity of being “high-performance but very expensive memory,” it would be difficult to expect mass-market demand. This reflected Samsung’s organizational culture of concentrating on businesses with “guaranteed profitability.” The structure of allocating people and resources to business units with certain current profits, rather than to uncertain future technology development, had become entrenched.

The outcome was devastating. As generative AI exploded in 2023, HBM demand far exceeded expectations. SK hynix’s HBM production was sold out through its 2025 output, and competitors also fell into a state of supply shortage. It was problematic that Samsung failed to read the market’s potential, but industry assessments were even harsher on the fact that it took more than a year to improve product quality after the market opportunity had become clear.

At the time, there was a view that Samsung’s HBM failed to meet NVIDIA’s quality requirements due to packaging technology issues. However, industry experts pointed to the possibility of fundamental problems in the DRAM design itself. Performance degradation caused by heat generation and power consumption is estimated to have been the reason it did not pass NVIDIA’s rigorous tests.

2025, signs of a technological turnaround

Samsung, however, turned crisis into a turning point. Through intensive technology development and capacity expansion, its market share surged from 17% in the second quarter of 2025 to a significantly higher level in the third quarter, overtaking US-based Micron to reclaim second place. While SK hynix maintained the top spot with a 57–64% share, Samsung’s catch-up began.

Of particular note is that Samsung’s next-generation product HBM4 received the highest scores in NVIDIA’s strict quality evaluations. In December 2025, NVIDIA evaluated Samsung’s 12-layer HBM4 as achieving 11 Gbps (gigabits per second) per pin. This far exceeds the JEDEC international standard of 8 Gbps and meets the high specifications NVIDIA required for Rubin.

Once the performance superiority of Samsung’s HBM4 was validated, NVIDIA requested far larger volumes from Samsung than initially expected. This appears to align with NVIDIA’s push for a multi-sourcing strategy and diversification of its supplier base. Samsung plans to begin supplying HBM4 modules to NVIDIA in the first half of 2026, and the semiconductor industry expects full-scale supply to begin from the second quarter.

HBM revenue in 2026, the reality behind a threefold surge

According to securities industry projections, Samsung Electronics’ HBM revenue in 2026 is expected to surge about threefold from the 2025 level to reach KRW 26 trillion. HBM shipment volume is also projected to triple year-on-year to 11.2 billion Gb, with next-generation HBM4 accounting for about half of total shipments.

What is even more striking is that this outlook is not merely a simple recovery from the previous downturn. According to KB Securities, Samsung Electronics’ operating profit in 2026 is expected to approach KRW 100 trillion, up 129% from 2025, with the key growth drivers identified as rising DRAM prices and expanding HBM shipments. The fact that Samsung recorded KRW 20 trillion in operating profit in just the fourth quarter of 2025 indicates it is heading toward the peak of a memory supercycle.

The outlook for HBM market share expansion is also striking. From 17% in the second quarter of 2025, Samsung’s share is estimated to surge to around 35% in 2026, with some institutions projecting above 30%. If market share expands to this extent, Samsung is expected to move beyond its second-place position and enter into full-scale competitive dynamics with SK hynix. Since SK hynix currently maintains an oligopolistic share of more than 50%, Samsung’s 30% share is seen as the threshold at which a “two-top competition” can emerge.

What is important is that Samsung’s technological capabilities in HBM4 are expected to continue to strengthen. HBM4 has already established a full-scale mass production system in the first half of 2026, and is projected to achieve a yield target of 70% for volume production around April 2026. With parallel development underway for next-generation technologies such as HBM5 and HBM6, some industry analysts see the possibility that the technology gap could be fully closed.

The possibility of recreating overwhelming technological leadership

Whether Samsung Electronics can recreate overwhelming technological leadership in HBM, the key memory of the AI era, now appears to be less a question of technology roadmap and more one of organizational determination and execution. The performance of Samsung’s HBM4 is already at a level that surpasses competitors, and its integrated capabilities in design, manufacturing, and packaging are among the best in the world. The key question is how quickly this technological edge can be commercialized and translated into supply stability that satisfies customers.

2026 is expected to be the turning point when these changes enter full swing. Whether Samsung can triple its HBM revenue, achieve a market share above 30%, and successfully secure customers in the ASIC market is likely to determine the next decade. As the saying in the semiconductor industry goes, “Two years of preparation, ten years of growth,” the question of whether current technological advances will evolve into genuine overwhelming leadership depends on execution over the coming years.

Popular News