TAIPEI (TVBS News) — When Samsung, SK Hynix and Micron pivoted their factories toward high-end AI chips last year, they left behind a gap in the global memory market. Taiwan's smaller manufacturers rushed to fill it. Now, an industry report projects that strategic opening will help drive memory chip sales to more than double in 2026 — a "super cycle" that analysts say is reshaping the competitive landscape.
The Market Intelligence and Consulting Institute (產業情報研究所), a Taiwan-based research organization affiliated with the Institute for Information Industry (資訊工業策進會), released the analysis Wednesday (April 30) as demand for high-bandwidth memory, or HBM, and DDR5 chips continues to outpace supply. The institute projects continued growth through 2028, though historical memory cycles suggest an eventual correction.
"The global memory industry is all about AI," Chris Hung (洪春暉), vice president and director general at MIC, said at the institute's spring forum Tuesday. DRAM serves as temporary data storage for processing tasks, he explained, and the three major global players have redirected most of their production toward high-bandwidth memory and DDR5 chips to meet surging AI demand.
Taiwan's memory chip makers Nanya Technology (南亞科技) and Winbond (華邦電子) have seen revenue growth exceeding 40 percent in 2025, according to the report. The gains came as the three major players shifted production capacity toward high-end AI memory, leaving gaps in the market for older products. Because HBM and DDR5 use similar process technology, "they are competing for the same production capacity," Hung said, pushing market prices higher across the board.
The AI boom has also spurred development of new memory technologies that could reshape the market again. Sandisk Corporation and SK Hynix announced in August 2025 a collaboration to develop High Bandwidth Flash, or HBF, a technology designed to store AI models directly alongside processing chips. Unlike AI training, which creates large language models, inference runs continuously as models generate answers across millions of devices — making these workloads hungry for both memory bandwidth and capacity.
"We're not just defining a new industry standard – we are setting the bar for the next era of AI computing," Alper Ilkbahar, chief technology officer at Sandisk, said at a February standardization event at the company's Milpitas, California headquarters. "Through our workstream under the Open Compute Project, we're laying the foundation for HBF to become a core building block of next-generation AI systems."
HBF technology is designed to deliver comparable bandwidth to HBM while providing eight to 16 times the capacity at a similar cost, according to Sandisk. The company targets delivery of the first HBF samples in the second half of 2026, with the first AI-inference devices using the technology expected in early 2027.
Cheng Kai-An (鄭凱安), an industry consultant at MIC, explained the distinction between the two memory types at Wednesday's forum. "HBM is basically composed of DRAMs," Cheng said. "When the power goes off, the memory goes off, so it is used for temporary access of computing data." HBF, by contrast, uses NAND flash technology that retains data when power is off, making it suitable for storing AI inference models near the GPU chip.
Taiwan Semiconductor Manufacturing Co. plays a critical role in the memory supply chain through its advanced packaging technology. The company's Chip-on-Wafer-on-Substrate, or CoWoS, packaging is essential for integrating HBM with AI processors, according to the MIC report.
However, Taiwan faces limitations in the most advanced memory segments. "It could be a challenge because Taiwan doesn't currently have the capability to produce HBM and HBF," Cheng said. Taiwanese companies are instead pursuing customized memory solutions for AI chips, which he described as "a niche market" with growing demand.
The MIC report cited revenue figures for Chinese memory manufacturers CXMT (長鑫存儲) and YMTC (長江存儲) but did not provide detailed information on their strategies or how U.S. export restrictions may affect competitive dynamics.
The window may not stay open long. Industry analysts caution that memory markets have historically been cyclical, and the giants will eventually expand capacity. Cheng noted that DDR4 demand could peak this year as system designs transition to DDR5. For Taiwan's smaller memory makers, the strategic question is not whether the super cycle will end, but what position they will hold when it does. The gap that Samsung, SK Hynix and Micron left behind is already beginning to close. ◼