SK hynix has begun equipment installation at M15X, their new factory in Cheongju, located 111 kilometers south of Seoul, with the goal of commencing operations in 2026. Designed as a center for high-bandwidth memory (HBM), the site enhances the company’s roadmap to lead the next wave of AI memory solutions amid a sustained increase in global demand.
Sources in the industry report that the company is already working inside the cleanroom of M15X and has relocated part of its workforce from its Icheon headquarters to Cheongju to prepare for ramp-up. M15X is an extension of the current M15 and is poised to become SK hynix’s Main HBM plant, facilitating the transition to HBM4 and supporting other group investments in South Korea.
A surgical move for the AI cycle
The decision to advance equipment installation and turn M15X into a HBM hub is strategic. The explosive growth of AI models—from foundational system training to large-scale inference—has elevated HBM to a critical component: offering higher bandwidth, greater efficiency per watt, and lower access latency compared to conventional DRAM. Controlling volume and yield of HBM will be key to commanding a significant part of the AI value chain.
Within this landscape, SK hynix is now one of the leading suppliers for major AI GPU manufacturers. It was an early mover with HBM3/HBM3E, and aims to repeat this success with HBM4. The company claims to have prepared for its industrial deployment following the completion of internal certifications. The Cheongju site guarantees incremental capacity just as demand for next-generation accelerators (with wider buses, taller stacks, and more demanding thermal requirements) is increasing supply pressures.
How M15X complements SK hynix’s manufacturing footprint
Cheongju previously housed M15, where essential TSV capabilities (Through-Silicon Via) have been deployed for reliable stacking and encapsulation of HBM with good thermal performance. The M15X is designed to absorb demand surges with an optimized design for HBM: density per square meter, material flow, and thermal dissipation planned for 12-Hi stacks and beyond, as well as for advanced molds and next-gen underfills.
The relocation of personnel from Icheon aims to ensure a rapid learning curve during ramp-up, spanning process engineering to maintenance and quality. This approach—bringing key equipment and personnel together—seeks to minimize downtimes between installation, line qualification, and commercial volumes, which is critical as the industry competes month-to-month for additional capacity.
Project timeline: from installation to production
The industrial schedule is clear: with civil works ending in 2025, the target is to qualify lines and start production in 2026. Meanwhile, the company continues to optimize M15 and prepare its packaging ecosystem for HBM3E and HBM4, ensuring M15X doesn’t start from scratch but is supported by proven processes and materials.
In fact, throughout 2025, the group has indicated that HBM4 has passed its internal validations and is close to final negotiations with NVIDIA for supply agreements. Essentially, the commercial pipeline is in place; what remained was physical capacity and operational alignment to serve large-scale orders.
HBM4: higher bandwidth, more layers, and increased heat
If HBM3E represented a significant leap, HBM4 raises the bar:
- Wider interface with buses up to 2,048 bits and higher speeds, bringing memory closer to the traffic peaks demanded by next-gen AI accelerators.
- Stacks of 12 layers (with higher configurations on the roadmap), which densify capacity but increase thermal management and signal integrity challenges.
- Custom base logic per client, complicating substitutions and making vendor relationships more sticky.
M15X is built to address these exactly: from ovens and presses to metrology and AOI (automated optical inspection), as well as molds and encapsulation materials that can handle increasing power and densities. The battle for leadership isn’t just about nanometers; it’s about processability, performance, and reliability at volume.
Competitive implications
The early equipment deployment at Cheongju comes at a time when all major players—SK hynix, Samsung, and Micron—are pushing to secure HBM capacity for 2026. Short-term, the advantage will go to those capable of multiplying useful wafers of HBM3E and HBM4 with competitive yields and closed contracts with leading GPU/ASIC designers.
In this race, SK hynix leverages history (being the first to deliver HBM3E to a major customer), portfolio (migration to HBM4 already underway), and capacity (M15X and ongoing optimizations at M15/Icheon), all within South Korea, simplifying logistics and supplier coordination.
Employment, supply chain, and ripple effects
The relocation of equipment and talent from Icheon to Cheongju is more than an internal move: it impacts the supply chain (equipment, chemicals, materials) and service providers (maintenance, construction, utilities, engineering). Once line qualification is complete, hiring of specialized technicians and operators will grow alongside production ramp-up, covering roles from process engineering and quality to factory logistics and safety.
For South Korea, this strengthens its position as a semiconductors power in the AI era, with Cheongju serving as a DRAM/HBM hub and Yongin as a major cluster expected to drive future growth.
Risks and challenges: from installation to yield
Not everything follows a straight path. The timelines from installation to production depend on factors like:
- Delivery and acceptance of equipment, process qualification, and initial batch variability.
- Availability of critical materials (resins, underfills, interposers, substrates) and their consistency.
- Thermal management and mechanical stress in tall stacks at higher frequencies.
- Integration with customer logistics (coordinated ramp-ups, part validations, joint testing).
The schedule pressure from AI—driven by increasingly shorter GPU cycles—pushes to overlap phases. M15X was conceived precisely to absorb this overlap without compromising quality or operational security.
What’s next
With equipment already entering clean rooms, staff relocated, and HBM4 ready for mass production according to the company, 2026 is shaping up to be the year Cheongju becomes a primary HBM node. The missing piece—beyond execution—is the solidification of multiyear contracts with major buyers of HBM4, with the company indicating they are in the final stages of negotiations.
For AI customers—from hyperscalers to system builders—each useful wafer of HBM is vital. The start-up of M15X will not only add volume but also reduce supply risks and provide an ecosystem cushion to boost density and performance in future accelerators arriving in 2026–2027.
Frequently Asked Questions
What is HBM and why is it key for AI?
HBM (High-Bandwidth Memory) is a vertically stacked DRAM using TSV technology, connected to GPU/ASICs via interposers with very wide buses. It provides massive transfer rates and low latency, with better power efficiency than traditional DRAM, making it essential for training and inference of large models.
What improvements does HBM4 offer over HBM3E?
Wider bandwidth per channel, with up to 2,048 bits interfaces, higher speeds, and taller stacks (e.g., 12-Hi), along with custom logical dies per client. The goal is to sustain performance jumps for next-gen accelerators and reduce energy consumption per processed token.
Why choose Cheongju instead of Icheon or Yongin?
Cheongju hosts the M15 (already equipped with TSV capacity) and the new M15X, specifically designed for HBM. Icheon continues playing a vital role in DRAM and other lines, while Yongin is set to become a major cluster facility, enabling faster ramp-up and better supplier coordination based on specialized regional strengths.
What impact could this have on supply for NVIDIA and others?
If the M15X ramp proceeds as planned, SK hynix will expand its supply of HBM3E/HBM4 right when new-generation accelerators go volumetric. The company states it’s in the final negotiation phase with NVIDIA for HBM4, which, if finalized, would provide stable volumes by the second half of 2026.
via: en.yna.co.kr