SK Hynix plans major DRAM expansion as AI keeps squeezing the memory market

SK Hynix is reportedly preparing a major memory production expansion that could nearly double its DRAM wafer capacity by 2030. The plan was already in motion before NVIDIA CEO Jensen Huang publicly asked the company to “please make more” memory during Computex, showing that SK Hynix had already seen where demand was heading.

The expansion matters because memory has become one of the biggest bottlenecks in the AI hardware market. GPUs get most of the attention, but advanced AI accelerators depend heavily on memory, especially high bandwidth memory. As AI data centers grow, demand for DRAM and HBM has surged, putting pressure on supply across the wider PC and server markets.

SK Hynix is already one of the most important suppliers in this space, and the company appears to be preparing for demand that will stay high for years. According to supply chain reports, SK Hynix wants to raise its DRAM wafer input capacity from about 550,000 wafers per month today to around one million wafers per month by 2030.

Yongin and Wuxi will play key roles in SK Hynix’s memory push

The center of the expansion is SK Hynix’s Yongin semiconductor cluster in South Korea. The first phase of that site is expected to include six cleanrooms, with equipment installation starting in February 2027. The company is expected to add 60,000 wafers of capacity at first, then continue adding similar capacity every six months.

By 2030, the first phase of Yongin could reach 360,000 wafers per month. SK Hynix is also expected to add 80,000 wafers of capacity at its M15X facility in Cheongju Technopolis.

The company’s Wuxi plant in China will also remain important. That facility is responsible for a large share of SK Hynix’s current DRAM production, reportedly close to half. Its role could become even more important as the company tries to balance demand from AI customers with broader memory supply needs.

The timing is important for the whole technology industry. AI demand has pushed memory prices higher, and PC builders are already feeling the impact through more expensive DDR5 and SSDs. Console makers, laptop companies, handheld gaming PC brands, and server customers are all competing for parts in a tighter market.

SK Hynix’s expansion could help ease some of that pressure, but not immediately. Semiconductor fabs take years to build, equip, qualify, and ramp. Even if equipment installation begins in 2027, the full effect of the new capacity may not be felt until closer to 2030.

There is also some caution in the supply chain. Memory markets are cyclical, and manufacturers have been burned before by large capital spending plans followed by sudden demand changes. Equipment suppliers are reportedly watching carefully because past expansion cycles have sometimes been followed by order reductions.

That risk is real. If AI demand slows, or if customers shift to different memory architectures, expanded capacity could create oversupply later. But right now, the market is moving in the opposite direction. AI companies, cloud providers, GPU makers, and server builders all need more memory, not less.

For NVIDIA, more SK Hynix capacity is especially important because HBM remains central to its AI accelerator roadmap. Jensen Huang’s public message may have been playful, but the business reality behind it is serious. NVIDIA needs memory partners to scale if it wants to keep shipping high end AI hardware at the pace customers expect.

For consumers, the expansion is a longer term positive sign. More DRAM capacity could eventually help stabilize prices and improve supply. But it will not quickly fix today’s memory crunch. DDR5 prices may remain high for some time, especially while AI demand continues to absorb large amounts of production.

SK Hynix’s plan shows how deeply AI has reshaped the memory business. DRAM is no longer just a PC component story. It is now part of the core infrastructure behind AI, cloud computing, GPUs, and future data centers. If SK Hynix reaches its one million wafer per month target by 2030, it could become even more central to the next phase of the AI hardware race.