SK Group Chairman Chey Tae-won warned this week that demand for AI memory chips — specifically high-bandwidth memory (HBM) — is outrunning supply by 60% to 100%, and governments are beginning to treat memory access as a national competitiveness issue. With SK Hynix controlling 58% of the global HBM market, a geopolitical contest over memory could reshape AI hardware supply chains.
For two years the AI hardware conversation has been about one thing: GPUs. Who has them, who can't get them, and what that means for the balance of power in artificial intelligence. SK Group Chairman Chey Tae-won says that conversation is about to shift — and the next bottleneck might hit harder. Speaking this week, Chey warned that demand for AI memory chips is growing so fast it could pull governments into what was once a straightforward supply-chain negotiation. The component in question is HBM — high-bandwidth memory, a specialized form of stacked DRAM that sits next to AI processors and feeds them data fast enough to keep up. Without enough HBM, even the most advanced GPU sits idle.
"Customers are requesting between 60% and 100% more AI memory for next year compared to current demand," Chey said. And the capacity to make it isn't arriving fast enough.
Why Memory Became a National Security Issue #
The physics of AI hardware creates an awkward dependency. Frontier model training requires massive parallel computation — thousands of GPUs working together. Each GPU needs HBM stacked next to it to handle the torrent of data flowing through transformer architectures. More compute means more memory. There's no way around it.
SK Hynix controls roughly 58% of global HBM revenue. Samsung and Micron split most of the rest. That's three companies — two in South Korea, one in the US — supplying a component that every AI lab and cloud provider on earth depends on.
When demand outstrips supply, countries start treating components differently. Chey said he's already seeing governments lock up domestic HBM supply for their own industries, and he cautioned that nations may soon begin pressuring one another over access to chipmakers. Memory is becoming a resource, not just a component.
The Factory Problem #
Building HBM capacity isn't like spinning up more cloud instances. A single high-end memory fab costs tens of billions of dollars and takes years to reach production volumes. The manufacturing requires advanced packaging that stacks memory dies vertically while managing heat — a technique only the big three have mastered at scale.
Chey's warning lands as the US continues restricting China's access to advanced AI chips and China pours state investment into domestic semiconductor production. If memory supply becomes a third front in that contest — alongside processors and manufacturing equipment — the hardware constraints on AI development get significantly more complicated.
What This Means #
The AI race is expanding beyond models and into supply chains. Companies that can secure HBM supply will train bigger models. Companies that can't will hit a ceiling that software can't fix.
Chey didn't predict full-blown trade restrictions over memory. But he didn't rule them out either. When a company that supplies more than half the world's AI memory says governments are circling, the smart move is to pay attention.
The GPU shortage taught the industry how fast hardware constraints reshape competitive dynamics. The memory shortage could teach a harder lesson — because unlike GPUs, there are even fewer suppliers.
Q: What is HBM and why does it matter for AI?
Q: How much memory demand growth is expected?
Q: Which companies control the HBM market?
Q: Can governments just build their own memory fabs?
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained #
Artificial Intelligence The science of creating machines that can perform tasks requiring human-like intelligence — reasoning, learning, perception, language understanding, and decision-making.
Attention A mechanism that lets neural networks focus on the most relevant parts of their input when producing output.
Compute The processing power needed to train and run AI models.
GPU Graphics Processing Unit.