AI Memory Demand Is Growing Exponentially — and Doubling Production Still Won't Catch Up

SK Hynix began trading American Depositary Receipts on the Nasdaq on Friday, in what was already the largest foreign IPO in U.S. history — raising $26.5 billion and surpassing Alibaba’s record. But the company’s chairman had a more sobering message for investors than the celebration might suggest.

Chey Taewon, chairman of SK Group (SK Hynix’s parent company), sat down with Bloomberg and CNBC after the listing ceremony. His central argument: the memory industry has entered a structural growth phase driven by AI, and the old rules no longer apply.

For decades, memory demand tracked two things: how many people existed, and how many smartphones and PCs they bought. Those were the variables that mattered. When phone shipments dipped, memory prices followed. When PC sales slowed, so did the entire DRAM market. It was a predictable, cyclical business.

That cycle is breaking. Chey pointed to four forces that have fundamentally changed the demand structure: AI agents, the key-value cache generated during inference (known as KV cache, a memory-hungry component of large language models), physical AI, and robotics. Each one consumes memory at a scale the industry has never seen before.

“Future memory demand will grow exponentially,” Chey told Bloomberg. “Even if we double production capacity in five years, our customers tell us it’s still not enough. Some customers say they need five to six times what we can currently supply.”

He added that demand for massive-scale memory will persist until AGI — artificial general intelligence — becomes widely adopted across society and the demand structure stabilizes.

Chey’s comments point to a tension that’s quietly reshaping the semiconductor industry. Memory used to be a commodity business — you built fabs, you produced DRAM and NAND, and you hoped prices didn’t crash. Now it’s a strategic bottleneck. Every AI company building larger models and deploying more inference infrastructure needs exponentially more memory bandwidth and capacity, and the fabs that used to produce consumer DRAM are being repurposed for high-bandwidth memory and AI accelerators.

SK Hynix is already the dominant player in HBM (high-bandwidth memory), the specialized memory used in AI GPUs. The company’s ADR listing gives it access to U.S. capital markets, which will help fund the capacity expansion Chey described. But even with that war chest, the supply gap may take years to close.