SK Hynix Was Already Sold Out. Then the Fires Started.
SK Hynix's Cheongju Campus 4 was already operating with zero margin before May 27. HBM3E fully booked through 2027, DDR5 up 5x in under a year, and a facility running at 100% capacity around the clock. Then three safety incidents hit in under three weeks.
- Between May 27 and June 10, 2026, SK Hynix's Cheongju Campus 4 experienced three separate safety incidents: a spark evacuation, a fire with hydrogen fluoride leak, and a toxic chemical exposure.
- The M15X fab runs at 100% capacity, 24 hours a day. In those conditions, the frequency of human error and safety incidents increases, and the pattern here reflects that reality.
- HBM memory, specifically HBM3E, is fully booked through 2026 and into 2027, directly limiting AI accelerator production for NVIDIA and AMD.
- 32GB DDR5 module prices have risen nearly 5x since mid-2025, driven by data center demand that now consumes roughly 70% of global DRAM output.
What Happened at SK Hynix Cheongju?
Between late May and mid-June 2026, SK Hynix's Cheongju Campus 4 experienced a series of safety incidents in rapid succession, not a single isolated event.
May 27: Sparks from semiconductor processing equipment at the M11 factory triggered an evacuation lasting approximately one hour.
June 1: A fire broke out at 10:32 a.m. in a sixth-floor gas room connecting the M15 and M15X plants. Sprinklers quickly extinguished the fire, but hydrogen fluoride spread inside the enclosed room. All 3,600 workers were evacuated from the M15 and M15X plants, and seven workers were taken to a company-affiliated hospital after complaining of eye irritation. SK Hynix stated production equipment was unaffected and operations would continue.
June 10: Two workers moving equipment at Cheongju Campus 4 came into contact with a liquid suspected to be tetramethylammonium hydroxide (TMAH), a highly toxic chemical widely used in semiconductor wafer cleaning and developing processes. The workers were transported to hospital.
In each case, SK Hynix reported that production was not significantly interrupted and that operations continued or resumed promptly. The fires were contained quickly. But that framing misses the more important point.
The real issue is not whether any single incident halted production. It's that this fab is running at 100% capacity, 24 hours a day, under extreme demand pressure, and in those conditions, the frequency of human error and safety incidents increases significantly. Three separate events at the same campus in under three weeks is a pattern, not a coincidence.
Why Memory Supply Was Already Under Pressure Before the Incidents
The Cheongju incidents are not the root cause of the current shortage. They are a risk multiplier on top of a structural problem that predates them. But they matter precisely because of how little margin exists in this supply chain.
The M15 and M15X plants at Cheongju Campus 4 hold a critical position in the global DRAM and NAND flash supply chains. Although SK Hynix has clarified that production will not be interrupted, the global memory chip supply is already in a state of tight balance, and any production disruption could trigger a chain reaction in the market.
The global pivot to AI infrastructure has created a memory supercycle unlike anything the industry has seen before. Data centers now consume approximately 70% of global DRAM output, compared to the historical norm of 20 to 30%. SK Hynix, the leading HBM memory manufacturer among a small group of HBM memory manufacturers worldwide, had its HBM capacity fully booked through 2026 before any of these incidents occurred.
The core issue: producing HBM requires the same wafer capacity used for conventional DRAM. Every wafer allocated to HBM3E for an AI accelerator is a wafer that cannot produce DDR5 for a server rack or LPDDR5X for a smartphone. That trade-off is now playing out across the entire electronics supply chain. For a deeper look at how CoWoS packaging constraints are compounding the HBM bottleneck, see our analysis.
Which Memory Products Are Most Affected?
High-Bandwidth Memory (HBM3E / HBM4)
What it is: HBM is the memory stack used in AI accelerators. The HBM memory GPU interface is what enables the extreme bandwidth that modern AI workloads require. NVIDIA's Blackwell Ultra (B300) platform requires up to 288GB of HBM3E per GPU using 12-Hi stack configurations.
Supply status: SK Hynix HBM capacity is fully booked through 2026, with shortages expected to extend into 2027. There is effectively no spot availability through standard channels.
Who is affected: Any hyperscaler or cloud provider scaling AI infrastructure, including AWS, Microsoft Azure, and Google Cloud, is competing for the same constrained pool of HBM allocation. AI GPU rollouts are directly gated by HBM memory chip availability.
High-Capacity DDR5 Server Memory (including 128GB RDIMMs)
What it is: DDR5 is the standard memory for modern data center servers. 128GB RDIMMs are required for the highest-performance workloads, including AI inference and large-scale data processing. For more on how 128GB RDIMM shortages are affecting server build timelines, see: The 128GB RDIMM Shortage: What Server OEMs Need to Know.
Supply status: Acute shortages. 32GB DDR5 modules rose from approximately $95 in mid-2025 to a projected $550 to $600 by Q2 2026, close to a 5x increase in under 12 months.
Who is affected: Server OEMs and hyperscalers building or expanding data center capacity. Organizations planning server refreshes or new infrastructure deployments are facing substantially higher costs than they may have budgeted.
|
Module |
Mid-2025 Price |
Q2 2026 Projected Price |
Change |
|
32GB DDR5 |
~$95 |
~$550 to $600 |
~5x increase |
LPDDR5X Mobile and Thin-Client Memory
What it is: LPDDR5X is the standard memory for smartphones, tablets, and ultrathin laptops. The 16Gb die supports 4GB, 8GB, and 16GB device configurations.
Supply status: Constrained. Global DRAM supply growth in 2026 is projected at approximately 16% year-on-year, well below the historical 20 to 30% range.
Who is affected: Consumer electronics manufacturers including major Android OEM vendors, Apple supply chain partners, and PC manufacturers. Dell, HP, and Lenovo have already signaled price increases of 15 to 20% across notebook lines. Smartphone vendors face a three-way choice: raise retail prices by 10 to 20%, reduce memory specs, or delay launches.
How Long Will the HBM Memory Shortage Last?
SK Hynix and its competitors are investing in new packaging capacity and fab expansion. However, semiconductor fab construction and qualification cycles are measured in years, not months. No meaningful relief from new capacity is expected before 2027 at the earliest.
Goldman Sachs significantly raised its 12-month price target for SK Hynix in June 2026 and reiterated a Buy rating, with the firm expecting memory supply tightness to persist through 2028. That outlook was formed before the June incidents, assuming normal operating conditions at facilities like M15X.
The current shortage is structural, not cyclical. Until wafer capacity can be meaningfully expanded or demand growth moderates, the industry will operate under tight allocation, elevated prices, and intense competition for available supply.
Strategic Implications for Procurement and Supply Chain Teams
The pattern of incidents at M15X is a reminder that the semiconductor supply chain operates with very little buffer, and that sustained maximum utilization carries its own risks. A facility processing 90,000 wafers per month, running 24/7, producing memory that is already sold out years forward, has no capacity to absorb downtime. When incidents do cause disruption, the effects propagate immediately.
For procurement and supply chain teams, the near-term reality is this: conventional allocation channels will remain under pressure, spot market prices will reflect scarcity, and the organizations that secured supply early are in the strongest position.
Standard allocation channels for HBM are exhausted. If your roadmap depends on HBM3E or HBM4 availability, waiting on your primary vendor is not a strategy.
Fusion Worldwide maintains access to high-bandwidth memory and other critical DRAM components through a global independent distribution network. Search available HBM inventory, submit a request for quote, or speak with a sourcing specialist directly.
Fusion also sources DDR5 server memory, LPDDR5X, and other components under allocation pressure. If you have a bill of materials with constrained parts, our team can assess availability across the full list.
Did the SK Hynix Cheongju incidents damage production equipment?
SK Hynix reported that production equipment was not directly damaged and that operations continued after each incident. The June 1 fire was extinguished by sprinklers; the June 10 TMAH exposure involved workers handling equipment in transit. However, three safety incidents at the same fab in under three weeks, at a facility running at 100% utilization around the clock, raises legitimate questions about whether the pace of operations is sustainable without increasing risk to workers and, eventually, to production continuity.
What is HBM memory and why does it matter for AI?
High-Bandwidth Memory (HBM) is a stacked DRAM architecture that delivers significantly more bandwidth than conventional DDR memory. It is the standard memory for AI accelerators including NVIDIA and AMD GPUs. Without sufficient HBM supply, AI infrastructure cannot scale.
What memory types are produced at SK Hynix's M15X facility?
The M15X fab primarily produces HBM3E and HBM4 memory, high-capacity DDR5 server memory including 128GB RDIMMs, and LPDDR5X mobile DRAM. It processes approximately 90,000 wafers per month and operates at full capacity.
Why are DDR5 prices rising so sharply in 2026?
Data center demand for AI infrastructure has pushed data centers to consume approximately 70% of global DRAM output, far above the historical 20 to 30% range. This has reduced available supply for conventional server and consumer memory, driving prices up significantly.
When will HBM memory supply constraints ease?
SK Hynix has confirmed HBM allocation is fully booked through 2026, with shortages expected into 2027. Goldman Sachs projects supply tightness persisting through 2028. New capacity from fab expansion investments is unlikely to reach volume production before 2027.
How are PC and smartphone OEMs responding to LPDDR5X shortages?
Major OEMs including Dell, HP, and Lenovo have signaled PC price increases of 15 to 20%. Smartphone vendors are weighing retail price increases of 10 to 20%, memory specification reductions, or delayed product launches.
What can procurement teams do to manage memory supply risk?
The most effective strategies include securing allocation earlier in the planning cycle rather than relying on spot availability, diversifying sourcing relationships, and working with distributors that have visibility into secondary market supply. Waiting until a shortage is widely reported typically means paying a significant premium.
Sources: Data Center Dynamics (June 1 fire and HF leak); The Asia Business Daily (June 10 TMAH exposure); TrendForce DRAM pricing data; Goldman Sachs SK Hynix price target, June 2026; industry analyst estimates for supply growth projections.