The high-bandwidth memory race, known as HBM, is no longer just a competition between manufacturers. It has become a strategic issue for artificial intelligence, data centers, and technological autonomy among major powers. South Korea continues to dominate this market with SK Hynix and Samsung, but China is beginning to close the gap at a rapidly increasing pace, raising industry concerns.
According to Seoul Economic Daily, the technological gap between Korean manufacturers and ChangXin Memory Technologies (CXMT), China’s leading DRAM producer, has narrowed to around three years. Reports indicate that CXMT has already achieved a level comparable to HBM3, although manufacturing yield issues still limit its ability to produce a high volume of good chips efficiently.
This nuance is important. China has not yet caught up with SK Hynix or Samsung in the most advanced HBM technology. Korean leaders are already working on HBM4, HBM4E, and future roadmap generations, while CXMT is focused on stabilizing HBM3. Nonetheless, this Chinese progress confirms that Beijing is not resigned to depending on foreign memory for powering its AI accelerators, servers, and domestic models.
CXMT and China’s Mobilization for AI Memory
CXMT’s movement responds to direct pressure from the Chinese government to reinforce the national semiconductor supply chain. U.S. restrictions on advanced chips, manufacturing equipment, and critical components have pushed China to accelerate the development of its own suppliers. In the case of HBM, this urgency is even greater because designing an AI accelerator isn’t enough; memory capable of moving data at enormous speeds alongside the processor is essential.
HBM is a vertically stacked DRAM memory connected via high-density interconnects, placed very close to GPUs and accelerators to provide significantly higher bandwidth than conventional memory. It is critical in chips from NVIDIA, AMD, and other AI providers because modern models require reading and writing vast amounts of data constantly.
According to Pulse/Maeil Business Newspaper, CXMT plans to expand its monthly 12-inch wafer capacity to approximately 300,000 units this year, with nearly 20% allocated to HBM3. This would equate to roughly 60,000 wafers per month dedicated to this technology, though more conservative estimates put initial HBM capacity around 30,000 wafers monthly.
The critical point will be the difference between theoretical capacity and actual production. Low yields will allow CXMT to demonstrate technical progress but won’t yet compete on volume or cost. If manufacturing yields improve, China will have a domestic HBM source sufficient to power parts of its AI industry, especially systems designed for the domestic market.
South Korea Keeps the Lead but Can No Longer Relax
SK Hynix remains the dominant player. Reuters, citing data from Counterpoint Research, reports that the company held 58% of the global HBM market in Q1, followed by Samsung and Micron, each with 21%. Additionally, SK Hynix is a key supplier for NVIDIA and expects to maintain a significant role in future architectures like Vera Rubin.
Samsung is also working to strengthen its position. The company has begun sending HBM4E samples to clients and demonstrated thermal management technologies for future generations. Meanwhile, SK Hynix plans to double its wafer capacity over the next five years to meet AI demand. Chey Tae-won, president of SK Group, has warned that memory bottlenecks could persist until 2030.
This context explains why Chinese advances are being closely watched. South Korea maintains technological leadership, with global customers, packaging expertise, manufacturing performance, and a privileged relationship with NVIDIA. However, the history of the semiconductor industry shows that a three-year gap can be closed with massive investment, growing domestic demand, and ongoing government support.
China doesn’t need to surpass SK Hynix tomorrow to reshape the market. Producing enough HBM3 to supply their national accelerators, reducing reliance on imports, and alleviating pressure on their AI supply chain can already have geopolitical and commercial impacts.
NVIDIA Spark Adds Pressure on LPDDR and Advanced Memory
This development coincides with another significant movement: NVIDIA’s release of RTX Spark, a new AI PC platform developed in collaboration with Microsoft, aimed at personal agents. NVIDIA describes it as a “petaflop superchip” with Blackwell RTX GPU, Grace 20-core CPU, up to 128 GB of unified memory, and the ability to run models with up to 120 billion parameters and context of one million tokens locally.
Although RTX Spark targets the PC market rather than data center accelerators, it underlines the same trend: AI is increasing memory demands across all layers. Not only is HBM necessary for large cluster training, but also vast amounts of LPDDR, efficient DRAM, and unified memory are required for workstations, laptops, mini-PCs, edge devices, and personal systems capable of running agents locally.
For Samsung and SK Hynix, this broadens opportunities. Low-power memory chips like LPDDR5X could gain prominence in a new wave of AI-focused computers. The PC is shifting from a simple productivity or entertainment device to becoming a machine capable of running assistants, models, and creative workloads locally, reducing reliance on cloud services.
However, this also escalates the competition. Demand for HBM in data centers, DRAM for servers, LPDDR for AI devices, and NAND for enterprise storage now compete for investment, manufacturing capacity, and materials. Memory has become one of the primary bottlenecks in the current AI cycle.
An Increasingly Political Race
The overarching perspective is that memory is now a key part of technological sovereignty. The U.S. seeks to restrict Chinese access to advanced chips. China responds by developing its own accelerators and pushing CXMT, YMTC, and other domestic players. South Korea protects its strategic advantage in HBM. Taiwan maintains its critical role in advanced manufacturing and packaging. Japan and the U.S. reinforce alliances for materials, equipment, and memory supply chains.
Within this landscape, CXMT is more than just a DRAM manufacturer. It is a component of China’s industrial strategy. Its potential IPO, investments in capacity, and progress toward HBM3 aim to fund and accelerate a supply chain less dependent on imports. Reuters reported plans for CXMT to list in Shanghai with an estimated valuation of $42 billion and to build an HBM packaging plant there.
The technical challenges remain enormous. HBM requires advanced DRAM, precise stacking, TSV technology, thermal management, logic integration, high-performance packaging, and validation with specific accelerators. Moreover, the market is moving rapidly: while CXMT pursues HBM3, industry leaders are already negotiating HBM4 and developing HBM4E.
The direction is clear. China has moved from viewing HBM as a distant technology to considering it a national priority. When a country combines internal demand, public funding, geopolitical pressure, and a more mature industrial base, timelines can be shortened.
South Korea remains at the forefront. China still trails, but the gap no longer feels as comfortable as before.
Frequently Asked Questions
What is HBM memory?
HBM stands for High Bandwidth Memory. It is vertically stacked DRAM placed close to GPUs or accelerators to provide extremely high bandwidth, essential for AI systems.
What has CXMT achieved?
According to South Korean reports, CXMT has reached a technological level close to HBM3, though it still faces low manufacturing yields and has not demonstrated large-scale stable production comparable to Korean leaders.
Has China matched Korea in HBM?
No. SK Hynix and Samsung remain ahead in more advanced generations such as HBM3E, HBM4, and HBM4E. The recent development is that China appears to be reducing the gap to about three years in HBM technology.
Why is NVIDIA RTX Spark relevant here?
Because it illustrates that demand for memory in AI is no longer limited to data centers. It will also grow in PCs, personal devices, and edge systems capable of running agents and models locally.
