← Back to feed

SK Hynix’s board voted this month to lock its AI memory production roadmap in Korea through 2033. The company framed the decision as executing a structural long-term investment strategy. The decision itself is not surprising. What sits beneath it is.

The US has zero domestic HBM wafer fabrication capacity. None. The country that builds the data centers, runs the frontier model labs, and writes the export control rules for AI chips does not manufacture the memory that makes those chips function. Every NVIDIA Blackwell GPU, every Google TPU, and every accelerator in every American data center runs on high-bandwidth memory wafers made in South Korea.

The 2028 Illusion

The CHIPS Act created a timeline. The honest version of that timeline is a packaging facility in Indiana with a realistic opening no earlier than 2028. That is not the same as wafer production. Packaging is the final assembly step. Wafer fabrication — where the memory itself is made — requires a different class of facility, different equipment, different lead times. No credible US wafer fab plan exists.

The gap between 2026 and 2028 is not the relevant interval. The gap between 2026 and whenever the US develops domestic HBM wafer fabrication capability is the relevant interval, and that number is at minimum a decade under current investment trajectories.

Why This Matters Now

Omdia’s demand projection, cited in SK Hynix’s official release, puts DRAM and NAND compound annual growth at 19% from 2025 through 2030. SK Hynix’s management called this a structural transformation — memory crossing from a component category into core AI infrastructure that determines model performance, not just model storage.

That framing is accurate. HBM is now 63% of the component cost in AI accelerator chips. The memory is not peripheral; it is the bandwidth bottleneck that limits how fast models run. Increasing HBM bandwidth and capacity is the primary variable enabling faster inference and larger context windows at production scale.

The AI memory market is sold out. All three major memory producers — SK Hynix, Samsung, and Micron — have committed their 2027 HBM production to existing customers. The supply constraint is physical and multi-year. The 19% CAGR demand projection means the constraint tightens even as production capacity grows.

The Dependency Structure

The US position: every American AI lab, cloud provider, and GPU manufacturer depends on an uninterrupted supply of Korean-made HBM wafers. SK Hynix holds roughly 60-70% of the HBM supply to NVIDIA for its high-end GPUs. Samsung supplies most of the remainder. Micron’s US-designed HBM is manufactured at facilities in South Korea and Japan.

The supply chain has no domestic US node for wafer production. The export control regime the US applies to restrict China’s access to advanced chips depends on wafer fabrication technology concentrated outside US borders.

SK Hynix’s board vote does not create this dependency. It just makes it visible. A seven-year production commitment to Korean facilities is a data point about where the company sees its supply obligations through 2033. The US AI industry’s timeline is the same curve, whether acknowledged or not.

What Changes

Nothing changes in the near term. Supply disruption from a geopolitical event affecting South Korea is a scenario, not a forecast. But the policy infrastructure for managing that risk — domestic wafer production, diversified supply, strategic reserves for AI memory — does not exist. The CHIPS Act invested in leading-edge logic and some packaging. HBM wafer fabrication was not the target.

The gap between the importance of HBM to US AI infrastructure and the absence of domestic production is, at minimum, something the next round of AI infrastructure policy needs to address. The current framework assumes the chip supply chain is the constraint. The memory supply chain is equally concentrated and considerably less discussed.