South Korea's HBM4 Moment: How Samsung and SK Hynix Became the Gatekeepers of AI

Two companies in South Korea control the future of artificial intelligence infrastructure. Samsung Electronics and SK Hynix produce 90% of global High Bandwidth Memory, the specialized DRAM that enables every frontier AI model to function.¹ With HBM4 mass production launching in February 2026, server DRAM prices surging 60-70%, and a 900,000 wafer-per-month commitment to OpenAI's Stargate project, memory has transformed from commodity component to strategic national asset.²³⁴

The implications extend far beyond quarterly earnings. Organizations planning GPU deployments now face memory-constrained timelines stretching into 2027. NVIDIA's next-generation Rubin architecture depends entirely on HBM4 availability. And South Korea has positioned itself as the irreplaceable chokepoint in global AI infrastructure.⁵

The February Revolution: HBM4 Arrives

February 2026 marks the most significant architectural overhaul in memory technology history. Both Samsung and SK Hynix will begin mass production of HBM4 simultaneously, ending months of speculation about which company would reach volume first.⁶

Technical Specifications

The JEDEC JESD270-4 standard, published in April 2025, defines HBM4's capabilities.⁷ The architecture doubles interface width from 1024 bits to 2048 bits, enabling bandwidth that previous generations could not approach.⁸

The channel architecture expansion from 16 to 32 independent channels enables greater access flexibility and parallelism in memory operations.⁹ For AI workloads that depend on massive parallel data access, this architectural change proves more significant than raw bandwidth numbers suggest.

Power efficiency improvements accompany the performance gains. JESD270-4 supports vendor-specific voltage levels for VDDQ (0.7V, 0.75V, 0.8V, or 0.9V) and VDDC (1.0V or 1.05V), enabling lower power consumption.¹⁰ Micron claims its HBM4 delivers over 20% better power efficiency compared to HBM3E products.¹¹

Samsung's Approach

Samsung adopted a turnkey approach for HBM4, handling chip design, manufacturing, and advanced microfabrication in-house.¹² The company uses a 10nm-class fabrication process for the base die of its HBM4 chips, which Samsung claims delivers better performance than competitors.¹³

Production begins at Samsung's Pyeongtaek campus in February 2026.¹⁴ The company plans to increase HBM production capacity by 50% during 2026, targeting approximately 250,000 wafers per month by year-end.¹⁵

Samsung's HBM4 chips reportedly achieved 11.7 Gbps in internal evaluations—an industry-leading figure if validated in production.¹⁶ Most chips will be installed in NVIDIA's Vera Rubin AI accelerator system, launching in H2 2026, with additional supply going to Google for seventh-generation Tensor Processing Units.¹⁷

SK Hynix's Strategy

SK Hynix takes a different approach, outsourcing the logic base die to TSMC through its "One-Team" alliance.¹⁸ This collaboration ensures HBM4 synchronization with TSMC's manufacturing processes for NVIDIA GPUs, potentially reducing integration complexity.

Production launches at SK Hynix's M16 plant in Icheon and the M15X fab in Cheongju.¹⁹ The company uses a 12nm fabrication process for the base die and adopted the Advanced MR-MUF (Mass Reflow Molded Underfill) process to minimize production risk.²⁰

SK Hynix maintains market leadership with an estimated 53-60% share of HBM production.²¹ The company completed HBM4 development first, showcasing working samples in April 2025, and delivered paid samples to NVIDIA ahead of Q1 2026 contract finalization.²²²³

The Stargate Commitment: 40% of Global DRAM Output

OpenAI's Stargate infrastructure initiative secured unprecedented memory commitments from both Korean manufacturers. The agreements, announced in October 2025, locked in supply through 2029 at volumes that will reshape global memory markets.²⁴

Scale of Commitment

Samsung and SK Hynix confirmed that OpenAI's anticipated demand could grow to 900,000 DRAM wafers monthly.²⁵ This volume represents approximately 40% of total global DRAM output—more than double current global HBM production capacity.²⁶²⁷

Analysts estimate the commitment amounts to more than 100 trillion won ($72 billion) of incremental demand for Korean chipmakers over four years.²⁸

Data Center Construction

The partnership extends beyond chip supply to physical infrastructure. OpenAI plans to construct Stargate data centers in South Korea, with Samsung and SK Hynix participating in development.²⁹

These facilities will serve dual purposes: demonstrating Korean memory technology in production AI workloads while establishing local infrastructure for OpenAI's global expansion. The Korean government views the data centers as strategic assets that reinforce the country's position in AI value chains.³⁰

The 60-70% Price Surge

Samsung and SK Hynix are raising server DRAM prices by 60-70% for Q1 2026 compared to Q4 2025.³¹ The increases reflect a fundamental shift in market power that began with HBM demand and now affects all server memory categories.

Pricing Dynamics

Both companies have rejected long-term agreements of two to three years, instead requiring quarterly contracts that allow stepwise price increases.³² Major customers including Microsoft and Google face the same terms—there are no volume discounts sufficient to offset the seller's market conditions.³³

Combined with 50% increases through 2025, cumulative price increases could nearly double server DRAM costs by mid-2026.³⁴ Industry consensus suggests tight supply conditions will persist through 2026, with Micron CEO Sanjay Mehrotra stating the shortage could extend beyond that year.³⁵

Downstream Effects

The server DRAM shortage cascades through consumer markets. DDR5 retail prices surged two to three times within weeks of the Stargate announcement, with some modules increasing by more than 150% in under a month.³⁶

Enterprise buyers face delivery timelines stretching to 10-14 months, with certain orders pushed into late 2026.³⁷ Organizations planning AI infrastructure deployments must now factor memory availability into project timelines with the same rigor applied to GPU allocation.

Supply Allocation Strategy

Both manufacturers prioritize HBM production over conventional DRAM, squeezing output for standard server memory.³⁸ This strategic choice maximizes revenue per wafer but exacerbates shortages for customers requiring traditional memory products.

SK Hynix's CFO previously stated the company has "already sold out our entire 2026 HBM supply."³⁹ Micron confirmed similar constraints, with HBM capacity for 2025 and 2026 fully booked.⁴⁰ New capacity will not meaningfully impact availability until 2027.

NVIDIA Rubin: The Demand Driver

NVIDIA's Rubin architecture, entering production in Q1 2026, demonstrates why HBM4 has become strategically critical.⁴¹ The platform's specifications require memory performance that only HBM4 can deliver.

Rubin Specifications

The first Rubin GPU features 288 GB of 6.4 GT/s HBM4 memory arranged in eight stacks, providing approximately 13 TB/s of aggregate bandwidth.⁴² This represents nearly triple the memory bandwidth of Blackwell, the current-generation architecture.⁴³

NVIDIA quotes up to 50 PFLOPS of NVFP4 inference and 35 PFLOPS of training performance.⁴⁴ Rubin Ultra, targeted for 2027, will double performance by moving from two compute chiplets to four, with memory capacity expanding to 1 TB of HBM4E delivering approximately 32 TB/s bandwidth.⁴⁵

Supply Chain Dependencies

Rubin's production depends on TSMC's 3nm process for compute dies and Korean HBM4 for memory.⁴⁶ Both represent potential bottlenecks. NVIDIA has secured HBM4 samples from all major DRAM manufacturers, but volume production allocation remains competitive.⁴⁷

The accelerated February 2026 HBM4 production timeline directly responds to NVIDIA's Rubin requirements.⁴⁸ Both Samsung and SK Hynix adjusted schedules to align with NVIDIA's platform launch, demonstrating the customer's leverage in partnership negotiations.

The $600 Trillion Yongin Cluster

SK Hynix announced plans to increase investment in its Yongin semiconductor cluster from 128 trillion won ($85.5 billion) to 600 trillion won ($410 billion).⁴⁹ The expansion transforms Yongin into the world's largest HBM production hub.

Cluster Specifications

The first fab at Yongin will focus on next-generation HBM products including HBM4 and HBM4E, alongside 321-layer NAND—the world's highest-density NAND flash.⁵⁰ Operations begin in 2027.

Samsung operates complementary facilities at Pyeongtaek, where a new P5 fab will come online in 2028.⁵¹ Samsung's $14.3 billion Research & Development Center in Yongin, operational since mid-2025, focuses exclusively on next-generation AI semiconductors.⁵²

National Investment Framework

The Korean government accelerated construction of the Yongin Semiconductor National Industrial Complex to begin in December 2026.⁵³ When complete, the complex will attract up to 360 trillion won ($246.4 billion) in private investment from major industry players.⁵⁴

President Lee Jae Myung announced a blueprint to invest more than 700 trillion won ($534 billion) in the semiconductor sector through 2047.⁵⁵ Near-term allocations include:

The government's 150 trillion won ($110 billion) National Growth Fund targets SK Hynix's Yongin cluster as its first investment destination.⁵⁶ This public-private coordination represents one of the largest industrial policy commitments in Korean history.

Strategic Implications

South Korea's dominance in HBM production creates advantages and vulnerabilities that will shape global AI infrastructure development.

National Security Dimensions

Memory has become a strategic asset comparable to energy resources or critical minerals.⁵⁷ Control over HBM production provides leverage in trade negotiations and geopolitical disputes. The Korean government increasingly treats semiconductor manufacturing as a national security priority.

The geographic concentration of production creates risk. Samsung and SK Hynix facilities in Korea represent single points of failure for global AI infrastructure.⁵⁸ Natural disasters, military conflict, or even industrial accidents could disrupt supply chains worldwide.

Both companies maintain operations in other countries, but advanced HBM production remains concentrated in Korea. Expanding this capability to other regions would require years of investment and technology transfer that neither government nor company has prioritized.

For AI Infrastructure Operators

Organizations planning large-scale AI deployments face new constraints:⁵⁹

Memory-first planning: GPU allocation depends on HBM availability. Secure memory commitments before finalizing compute purchases.

Extended timelines: Plan for 12-18 month lead times on server DRAM and even longer for HBM-equipped systems.

Budget adjustments: Factor 60-100% memory cost increases into infrastructure projections.

Supplier relationships: Direct relationships with memory manufacturers may prove as valuable as GPU allocations.

For the AI Industry

The HBM bottleneck challenges assumptions about AI scaling. Compute density improvements driven by architectural advances (like those demonstrated by DeepSeek) may prove more practical than raw GPU accumulation when memory constrains deployment.⁶⁰

Inference optimization gains strategic importance. Models that require less memory bandwidth per token can serve more users on constrained hardware. Architecture decisions that seemed like engineering details become competitive advantages.

For Korea

Success creates its own challenges. The concentration of AI infrastructure value in Korean memory manufacturing invites regulatory scrutiny, potential trade restrictions, and competitive responses from other nations seeking to develop domestic capacity.⁶¹

China's memory manufacturers continue development despite export controls. The United States has increased investment in domestic semiconductor manufacturing. Neither represents near-term threats to Korean HBM dominance, but both signal long-term strategic responses to concentration risk.⁶²

What Comes Next

The HBM shortage will ease—eventually. New fabrication plants planned between 2026 and 2030 will begin reaching scale, and memory manufacturers will expand capacity in response to sustained demand.⁶³ But relief remains years away.

2026 Outlook

HBM4 production ramps through 2026, but demand from Rubin deployments will absorb most output. Server DRAM prices will remain elevated through the year, with quarterly increases likely to continue until new capacity comes online.

The February production launch marks the beginning of the HBM4 era, not the solution to supply constraints. Expect allocation battles between hyperscalers, with smaller buyers facing extended delivery windows.

2027 and Beyond

SK Hynix's Yongin cluster begins operations in 2027, potentially adding meaningful HBM4 capacity. Samsung's P5 fab reaches production in 2028. These additions should improve availability, though NVIDIA's Rubin Ultra and subsequent architectures will drive continued demand growth.

HBM4E stacks reaching 64GB will enable new AI system configurations, but will also require additional production capacity that may not exist at launch.⁶⁴ The cycle of demand outpacing supply shows no signs of breaking.

The Longer View

South Korea's semiconductor industry has achieved a position that few countries can challenge. The combination of manufacturing capability, R&D investment, and government support creates durable competitive advantages that will persist for the foreseeable future.

For AI infrastructure operators, this reality demands strategic adaptation. Memory is no longer a commodity to be purchased on demand—it is a constrained resource requiring long-term planning and supplier relationships. Those who recognize this shift early will secure positions in the next generation of AI systems. Those who don't will wait.

References