AI 가속기 설계사 아스트라가 HBM4E 메모리 검증 일정을 앞당기며 차세대 패키징 계약에서 삼성전자와 SK하이닉스 간 경쟁이 격화되고 있다.
“AGI has arrived,” declared Nvidia CEO Jensen Huang as he congratulated OpenAI on the release of GPT-6 Astra, which the company calls its most intelligent and aligned model yet. AGI, or artificial general intelligence, refers loosely to AI capable of performing intellectual tasks across a broad range of domains at or beyond human levels. Whether Astra has actually crossed that threshold remains disputed, but the race toward it among frontier developers including OpenAI, Anthropic, Meta, and Google is intensifying. Huang noted that Astra was trained using more than 100,000 Nvidia Grace Blackwell GPUs across NVL72 systems, with another 400,000 GPUs coming online.
Regardless of whether the AGI label sticks, the economic implications for the semiconductor industry are clear. More capable AI models, increasingly autonomous agents, and heavier competition for users and enterprise workloads are driving demand for computing power. Greater computing power requires more GPUs—and consequently, more high-bandwidth memory (HBM) to feed data into them. South Korea’s two memory giants and their investors reacted positively Monday: Samsung Electronics shares jumped 5.68 percent to close at 270,000 won, while SK hynix surged 8.26 percent to 1,783,000 won.
The next battlefield for HBM supremacy is taking shape in Gyeonggi Province, home to the advanced chipmaking complexes of both companies. Samsung and SK hynix are preparing to go head-to-head using sixth-generation 10-nanometer-class, or 1c, DRAM in HBM4E after pursuing sharply different process strategies for HBM4. The confrontation is growing more consequential as Samsung closes the gap with the longtime HBM leader. According to Counterpoint Research, Samsung’s share of global HBM revenue jumped to 33 percent in the second quarter from 21 percent three months earlier, while SK hynix’s share fell to 50 percent from 58 percent. The gap between the two Korean makers narrowed to 17 percentage points from 37 points in just three months. A year earlier, SK hynix held 64 percent of the market against Samsung’s 15 percent, leaving a 49-point gap. Micron ranked third with an 18 percent share in the second quarter, down from 21 percent in the first. These figures do not yet reflect the balance of power in HBM4, as most revenue still comes from HBM3E, with HBM4 shipments expected to become increasingly visible in market revenue during the second half of this year.
Despite the narrowing gap, Samsung and SK hynix pursued markedly different manufacturing strategies for the new generation. Samsung took the more aggressive route in DRAM process technology, beginning mass production and commercial shipping of HBM4 in February using 1c DRAM—a break from the industry’s conventional practice of applying a more mature process to HBM products. Samsung’s HBM4 combines 1c DRAM with a 4-nanometer logic base die and delivers a sustained data-transfer speed of 11.7 gigabits per second per pin, with speeds reaching up to 13 Gbps. A Samsung official stated that yields for its 1c-based HBM4 have stabilized and that the product has been supplied smoothly to customers since mass production began in February.
SK hynix took the opposite approach. Rather than immediately adopting 1c for HBM4, the company retained its proven fifth-generation 10-nanometer-class, or 1b, process while initially applying 1c to conventional DRAM products. An SK hynix official explained that the company deliberately reversed Samsung’s strategy, deploying 1c first in higher-value HBM products while retaining an older process for conventional DRAM. Using the already proven 1b process for HBM4 offered an advantage in securing stable yields, the official said, describing the decision as part of a broader process strategy rather than the result of a single technical factor.
The distinction matters because a more advanced DRAM process alone does not determine HBM performance. Lee Jong-hwan, professor of system semiconductor engineering at Sangmyung University, noted that while 1c holds a performance advantage over 1b, the difference between the two processes is only one component of the overall HBM architecture. “DRAM performance is important, but the connection technology that links each DRAM die is also critical,” Lee said. “The DRAM itself and packaging technologies all have to perform well together to achieve strong overall HBM performance.” He added that the difference between 1b and 1c alone is not large enough to determine the competitiveness of the finished HBM product, as both are successive refinements within the same 10-nanometer-class generation.
The two companies’ strategies are now set to converge with HBM4E. SK hynix is moving its broader DRAM production toward 1c as it prepares to use the newer process in HBM4E, bringing its core DRAM process in line with Samsung’s. The company confirmed that shipments of products based on its 1c process began in earnest in the second quarter. Moving to the newer process gives chipmakers room to improve performance and power efficiency, though the economic payoff depends heavily on stabilizing yields as manufacturing complexity increases. “Better performance makes it harder to secure yields,” Lee said. “In semiconductors, yield directly translates into money. If yields are poor, you cannot supply the volume customers want.” “Performance and yield have to improve together. That is what creates real competitiveness,” he added. This trade-off is particularly critical for HBM, which stacks multiple DRAM dies and requires sophisticated bonding, packaging, and thermal-management technologies alongside the underlying memory process. HBM4E will put those capabilities to a tougher test.
Samsung shipped samples of its 12-layer HBM4E to major global customers in May. The product combines 1c DRAM with a 4-nanometer logic base die and supports data-transfer speeds of up to 16 Gbps per pin. The company argues that its experience mass-producing HBM4 with the same 1c DRAM and 4-nanometer base-die combination could give it an advantage as HBM4E moves toward commercial production. “Because we have already mass-produced HBM4 based on 1c and 4-nanometer technology, we expect to provide HBM4E with more stable performance and quality,” a Samsung official said. Samsung emphasized that its key advantage lies less in headline specifications than in having already validated the underlying manufacturing combination through HBM4 mass production.
SK hynix supplied samples of its own 12-layer HBM4E to major customers in June. The product also supports speeds of up to 16 Gbps per pin and offers more than 20 percent higher power efficiency than the preceding generation, according to the company. HBM4E remains at the sample-supply stage for SK hynix, with mass production dependent on customer schedules. The company aims to move forward within this year, though precise timing has not been fixed. SK hynix noted that the biggest improvement expected from moving to the newer DRAM node is performance, particularly data-transfer speed.
The convergence on 1c fundamentally shifts the competitive dynamic. The HBM4E contest is unlikely to be decided simply by which company uses the more advanced DRAM process. Both must prove they can combine high-performance 1c DRAM with stable yields, advanced packaging and bonding, efficient thermal management, and large-volume production aligned with customer schedules. Consequently, the two companies enter the HBM4E era from different positions. SK hynix brings the larger HBM market share and customer and manufacturing experience accumulated during its HBM3E leadership. Samsung brings experience mass-producing 1c-based HBM one generation earlier and a sharply rising HBM revenue share.
Investor expectations are rising alongside the technological competition. Consensus estimates for Samsung Electronics’ third-quarter operating profit stood at 114.2 trillion won ($84.8 billion) as of Sept. 4, up 12 percent from three months earlier, according to FnGuide. SK hynix’s third-quarter operating-profit estimate rose 4 percent over the same period to 78.8 trillion won.