Dominating 80% of the HBM Market but Lacking AI’s Soul: South Korea’s Semiconductor Powerhouse Risks Becoming a High-End Foundry — BigGo Finance

South Korea’s memory titans, Samsung Electronics and SK Hynix, leveraging their High Bandwidth Memory (HBM) technology, collectively command over 80% of the global market share, positioning themselves as the most critical hardware suppliers for AI computing infrastructure. However, in core AI software and system domains such as GPU design and large language models, South Korea lags significantly behind the United States, fueling industry concerns that this semiconductor powerhouse may be devolving into a top-tier foundry rather than a rule-maker.

This contradiction has been thrust into the spotlight ahead of South Korean President Lee Jae-myung’s visit to Silicon Valley. President Lee will attend an AI summit in San Francisco and meet with tech leaders including NVIDIA CEO Jensen Huang, OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and Broadcom CEO Hock Tan. Samsung Electronics Chairman Lee Jae-yong, SK Group Chairman Chey Tae-won, Hyundai Motor Group Chairman Chung Eui-sun, and Naver founder Lee Hae-jin will also attend, marking what is arguably the highest-profile AI investment promotion trip in South Korean history.

Kim Yong-beom, President Lee’s policy chief, revealed that Samsung and SK Hynix are preparing to announce major agreements with U.S. tech giants during the visit, including long-term memory supply contracts, strategic investment partnerships, and multiple memorandums of understanding. While Kim did not disclose specific amounts, he emphasized that the presidential visit helped accelerate the negotiation process. These agreements are expected to provide investors with a clearer picture of future AI memory demand.

Structural Concerns Behind Hardware Dominance

SK Hynix became the first to mass-produce HBM4 this February, securing two-thirds of the orders for NVIDIA’s Vera Rubin platform. Its high-end HBM market share surged to 90% in the first quarter, with HBM revenue more than tripling year-over-year and gross margins approaching 80%. Samsung also completed HBM4 mass production, shipped 12-layer HBM4E samples ahead of schedule, and secured an exclusive order for OpenAI’s Titan chip, breaking SK Hynix’s monopoly on high-end customers.

However, structural risks lurk beneath South Korea’s hardware dominance. Domestic large language model companies like Naver and Upstage rely entirely on NVIDIA GPUs for training, meaning the core computing discourse remains firmly in American hands. The lack of a software ecosystem is even more critical—South Korea’s AI applications are mostly lightweight consumer products, with no general-purpose foundation models capable of competing with international leaders. Most companies depend on open-source frameworks for secondary development, with limited original technology.

Furthermore, the industrial structure is overly concentrated in the single domain of memory, heavily dependent on major overseas clients such as NVIDIA, Google, and Microsoft. Should geopolitical circumstances shift or competitors significantly expand production, South Korea’s AI hardware industry would face direct impact.

AMD Joins the Fray, Adding Fuel to HBM Demand

Just as concerns mount over South Korea’s hardware dominance lacking a software soul, AMD officially launched its next-generation AI server GPU, the “MI455X,” and its first rack-scale AI system, “Helios,” at the “AMD Advancing AI 2026” annual technology conference in San Francisco, injecting a new wave of growth momentum into the HBM market.

The MI455X is equipped with 432GB of HBM4, 50% more capacity than the 288GB on NVIDIA’s next-generation Vera Rubin GPU. The Helios system integrates 72 MI455X GPUs, delivering a total HBM4 capacity of approximately 31TB per rack, equivalent to 864 stacks of 36GB HBM4. Throughout the presentation, AMD consistently compared Helios to NVIDIA’s Vera Rubin NVL72, highlighting its competitive advantages in memory capacity and bandwidth.

AMD CEO Lisa Su stated that the AI accelerator market will reach $1.4 trillion by 2030, emphasizing that “no single company can solve all problems; ecosystem-level collaboration is crucial.”

South Korea’s semiconductor industry is closely watching AMD’s official entry into the rack-scale AI system market. AI server racks, which integrate dozens of GPUs, CPUs, and networking devices, are the fundamental building blocks of AI data centers. Previously monopolized by NVIDIA, AMD’s entry into the competition is certain to further drive HBM demand. For SK Hynix and Samsung, which already dominate the HBM market, this translates to more stable long-term demand.

Notably, AMD’s HBM4 is primarily supplied by Samsung. Samsung not only was the first to mass-produce HBM4 but also signed a memorandum of understanding with AMD in March covering next-generation AI memory, including HBM4. Anthropic also announced a strategic partnership with AMD, planning to use Helios systems to build 2GW of AI infrastructure. Previously, OpenAI, Meta, and Oracle had also announced AI infrastructure expansion plans with AMD.

HBM Market Landscape Reshaped, Demand Structure Diversifies

UBS predicts a significant shift in the HBM market’s customer structure. NVIDIA’s share of demand is expected to drop from 60% this year to 41% next year, while Google’s share climbs from 18% to 30%, and AMD’s grows from 7% to 13%. The HBM market is transitioning from NVIDIA’s monopoly to a landscape where multiple AI platforms flourish.

In terms of overall market size, UBS estimates the HBM market measured by capacity will grow from 33.14 billion GB this year to 58.65 billion GB next year, a surge of 77%. Among this, AMD’s HBM demand is projected to skyrocket from 2.17 billion GB to 7.52 billion GB, a 3.5-fold increase, the highest growth rate among all customers.

The surge in HBM demand will also have a ripple effect on the standard DRAM market. As memory manufacturers expand HBM production capacity, standard DRAM production space will inevitably be squeezed. HBM employs a multi-layer DRAM stacking structure, consuming significantly more wafers than standard DRAM, which has become a key factor driving up overall DRAM prices.

Earnings Season Kicks Off, Market Scrutinizes AI Demand

The optimistic outlook for the memory industry is about to be tested by heavyweight earnings reports. SK Hynix will report on July 29, followed by Samsung Electronics and Japan’s Kioxia Holdings on July 30 and 31, respectively.

HSBC analysts noted that rising HBM prices and the commencement of HBM4 shipments will support Samsung and SK Hynix’s performance, helping to alleviate market concerns about slowing profit growth in the second half of the year. SK Hynix CEO Kwak Noh-jung further expects that memory supply shortages could persist beyond 2030 amid surging AI demand.

Samsung Electronics already released preliminary results earlier this month, reporting second-quarter operating profit of ₩89.4 trillion (approximately $61.3 billion), beating market expectations by about 6%. However, its stock price plunged over 7%, reflecting investor concerns about the return on AI infrastructure investments. Kioxia’s experience was even more dramatic—the NAND flash memory giant briefly claimed the title of Japan’s most valuable stock in June, only to see its share price halved in July.

Despite this, Wall Street and Japanese brokerages maintain their bullish stance, with some even raising target prices. Analysts broadly believe that company fundamentals have not deteriorated, and this correction is primarily driven by market sentiment and technical selling pressure rather than a reversal of industry trends. Bloomberg Intelligence further pointed out that since building new memory fabs takes years, the severe supply-demand imbalance could persist until 2028 or beyond, forecasting NAND prices to rise over 300% this year, with further gains likely next year.

The Long-Term Challenge: Hardware Giants Lacking AI’s Soul

Returning to the structural issue, the South Korean government announced last month an investment plan supported by Samsung Electronics, SK Group, and Naver, totaling at least $880 billion, aimed at expanding chip manufacturing, developing AI data centers, and strengthening South Korea’s position in semiconductors and physical AI. President Lee is also expected to issue the “San Francisco AI Declaration” during this trip, outlining a blueprint for expanding AI economic applications in South Korea.

Kim Yong-beom revealed that more than half of the approximately 8GW of AI data center capacity planned in the first phase by the South Korean government could advance into concrete projects during the presidential visit, including securing customers, site selection, and South Korean engineering partners. He emphasized that U.S. tech companies account for 80% to 90% of orders in South Korea’s latest expansion plan, indicating that U.S. demand remains the key driver of South Korea’s semiconductor and AI infrastructure investments.

However, in the long run, while South Korea’s semiconductor industry holds the most critical storage chips for AI computing, it lacks the AI brain and soul that drive the entire ecosystem. With U.S. companies controlling the core discourse in GPU design, large language models, and cloud computing platforms, if South Korea cannot achieve breakthroughs at the software and system level, it may be destined to remain a top-tier supplier rather than an ecosystem rule-maker.

Industry insiders bluntly state that if the “strong hardware, weak software” structure of South Korea’s AI industry does not change, even with booming HBM demand and stellar earnings, it will be difficult to escape the fate of being a “high-end foundry” in the long term. How to cultivate AI platforms and foundation models with original technology on the foundation of memory dominance will be the key challenge in transforming South Korea’s semiconductor powerhouse.