• SK Hynix aims to roughly double memory-wafer capacity by 2030 to meet surging AI demand for high-bandwidth memory.
  • NVIDIA CEO Jensen Huang pushed back on the plan, saying more supply is needed to avoid bottlenecks in AI compute pipelines.
  • The exchange underscores a growing tension between AI chip makers and memory suppliers as demand outstrips supply.

Capacity Expansion Falls Short

SK Hynix Inc., the world's leading supplier of high-bandwidth memory (HBM) used in AI accelerators, has unveiled plans to roughly double its total wafer output by 2030. The move marks a strategic shift from SK Group Chairman Chey Tae-won's earlier cautious stance on adding new fabs, reflecting the urgent need to capitalize on AI-driven demand. However, NVIDIA Corp. CEO Jensen Huang publicly stated that the expansion is insufficient, according to people familiar with his remarks at a recent industry gathering.

"Doubling capacity sounds ambitious, but when you consider the trajectory of AI compute, it won't keep pace with demand," Huang said, according to a person who heard the comments. NVIDIA is a major customer of SK Hynix's HBM, which is critical for its graphics processing units used in data centers. The memory supply-demand gap has been a persistent bottleneck in AI hardware pipelines.

A Tightening Supply Chain

The stakes are high. SK Hynix controls roughly a majority of the HBM market as of early 2026, according to industry trackers, making its capacity plans crucial for the entire AI ecosystem. The company intends to invest heavily in new fabrication facilities and improve yields at existing plants. Yet, Huang's pushback highlights concerns that memory supply will remain constrained even with a doubling of output. Without additional capacity, AI developers and data-center operators could face higher costs and delayed deployments, analysts warn.

Other memory makers, including Samsung Electronics Co. and Micron Technology Inc., are also ramping up their HBM production, but the sector-wide expansion may still lag demand. The capacity race is not just about volume; it also involves complex manufacturing processes for HBM, which stacks multiple DRAM dies vertically, requiring advanced packaging and testing.

Implications for the AI Boom

NVIDIA's public pressure on SK Hynix is unusual but underscores the critical nature of memory supply. "We have a constant dialogue with our suppliers," an NVIDIA spokesperson said, declining to comment directly on Huang's remarks. SK Hynix did not immediately respond to requests for comment.

Analysts say that without a significant acceleration in capacity, memory prices could remain elevated, squeezing margins for AI hardware makers and potentially slowing the pace of AI innovation. On longer timescales, if capacity expansions materialize as planned, the memory ecosystem could see improved resilience, but near-term bottlenecks are expected to persist through the end of the decade.

Correction: This article has been updated to clarify that Huang's remarks were made at an industry event, not in a formal statement.