A forecast said Korea will maintain its lead in the global memory chip market through 2031, backed by facility investments by Samsung Electronics and SK hynix. However, demand for high-bandwidth memory (HBM) is expected to rapidly spread from Nvidia's graphics processing units (GPUs) to in-house artificial intelligence (AI) accelerators from big tech such as Google.
Josephine Louw, an industry analyst at French market research firm Yole Group, said on the 4th (local time) at Future of Memory and Storage (FMS) 2026 in Santa Clara, California, that Korea will keep its lead in the memory industry through 2031.
Louw cited Korea's plan to double wafer production capacity over the next five years and expanded investments by Samsung Electronics and SK hynix. With Samsung Electronics and SK hynix simultaneously expanding production capacity and technology development in next-generation DRAM, including HBM, and NAND, Korea's advantage will continue, the analysis said.
Memory consumption by region is rapidly being reorganized around the United States. According to Louw, the U.S. share of global memory consumption rose 15 percentage points, from 37% in 2021 to 52% last year. Over the same period, China's share fell 6 percentage points, from 35% to 29%.
The combined share of the United States and China increased from 72% to 81%. Investments by U.S. big tech in AI data centers have driven memory demand, while U.S. semiconductor restrictions on China have limited production and consumption of advanced chips in China.
Louw assessed that the memory market is moving from "AI 1.0," centered on training AI models, to "AI 2.0," where inference demand is expanding in earnest. As AI services are applied to real industries and consumer products, requirements for data throughput and memory capacity and bandwidth are growing simultaneously.
The speed of product development has also emerged as a key competitive edge. Previously, next-generation memory development took about 18 months, but as AI accelerator release cycles shorten, memory companies are under pressure to supply new products in about 12 months.
In the HBM market, Nvidia's influence is expected to gradually wane. Avril Wu, senior research vice president at Taiwan market research firm TrendForce, projected Nvidia's share of total HBM demand will fall from 66% last year to 58% this year.
That is because cloud service providers (CSPs), including Google's tensor processing unit (TPU) as well as Amazon and Meta, are increasing adoption of application-specific integrated circuits (ASICs) they developed in-house. An ASIC is a chip designed to match a specific corporation's AI services and computing methods.
TrendForce forecast that ASIC-based products will account for 27.8% of AI server shipments this year. GPUs will still take 69.7%, but with Google and Meta expanding their in-house AI chips, the growth rate of ASIC servers will outpace that of GPU servers.
HBM demand itself is expected to expand regardless of Nvidia's declining share. TrendForce said that while Nvidia will maintain its position as the largest HBM customer, in-house AI accelerators from big tech, such as Google's TPU, are rapidly growing as new demand sources.
At FMS, Wu projected a "structural supercycle" from 2026 to 2028, as consumer memory production capacity shifts to HBM4, HBM4E and high-capacity NAND, prolonging supply bottlenecks. The meaning is that high-value AI products will absorb wafer input, potentially constraining supply of commodity DRAM and NAND.
The technological catch-up by Chinese memory corporations was also cited as a variable. Yole Group, after analyzing DRAM from Changxin Memory Technologies (CXMT), assessed it has reached the 1z process, the third generation in the 10-nanometer class. That is the technology level industry-leading companies secured in 2019.