Memory bandwidth is the physical ceiling on AI inference throughput, which places the companies supplying it at the front of any rotation tied to AI capital spending. Samsung and SK Hynix both saw their shares slide on South Korea's Kospi as investors extended a broader sell-off in AI-exposed chipmakers. A trading halt was triggered on the exchange.
Where memory suppliers fit in the AI hardware stack
The constraint driving chipmaker sensitivity to AI sentiment is architectural. A GPU cluster's utilization rate depends on how fast data moves between the processor and memory. When that bandwidth falls short, compute sits idle. The chipmakers filling that role carry the AI infrastructure bet in concentrated form.
Samsung and SK Hynix both supply memory at scale for AI workloads. That position means their order books tighten when AI capex intentions pull back, and their share prices often move ahead of that tightening. The Kospi declines in both names fit that pattern.
Reading the trading halt
The sell-off "deepens" by the framing attached to this session, placing it inside a move already in progress rather than one beginning here. Investors had been rotating out of chipmaker exposure before South Korean markets registered the pressure.
A trading halt on the Kospi indicates the pace of decline hit exchange-level intervention thresholds. That kind of response is triggered by speed, not direction. For Samsung and SK Hynix, the slide reflects the same thesis playing out: when AI capex sentiment weakens, memory suppliers are the most direct expression of it among publicly traded names, and they absorb the exit accordingly.