Regulating Global AI via Lithography and Memory Export Controls

This article explores how multilateral export controls on extreme ultraviolet (EUV) lithography equipment and high-bandwidth memory (HBM) chips are used to regulate global access to frontier artificial intelligence training. By restricting the distribution of advanced semiconductor manufacturing tools and critical memory components, allied nations aim to bottleneck the computational power required to build next-generation AI models. The following sections break down the specific technologies involved, the mechanics of these international restrictions, and their direct impact on the global AI landscape.

The Role of EUV Lithography in AI Hardware

Training frontier AI models requires thousands of advanced graphics processing units (GPUs) and specialized AI accelerators. Manufacturing these cutting-edge chips relies entirely on extreme ultraviolet (EUV) lithography.

EUV lithography uses extremely short wavelengths of light to print microscopic transistor patterns onto silicon wafers. Currently, only one company in the world, the Dutch firm ASML, produces commercial EUV lithography machines. Without access to these machines, semiconductor fabrication plants cannot efficiently produce logic chips at the 7-nanometer node or smaller, which are the foundational building blocks of modern AI data centers.

The High-Bandwidth Memory (HBM) Bottleneck

Raw processing power is useless for AI training if data cannot be fed into the processor fast enough. This is the exact problem solved by high-bandwidth memory (HBM).

HBM vertically stacks memory dies and connects them using microscopic wiring called through-silicon vias (TSVs). This architecture allows massive amounts of data to transfer between the memory and the processor at speeds far exceeding traditional memory. Frontier AI models have billions or trillions of parameters; training them requires HBM to prevent data bottlenecks. The HBM market is heavily concentrated in South Korea and the United States, specifically among three manufacturers: SK Hynix, Samsung, and Micron.

Mechanics of Multilateral Export Controls

Because the supply chains for both EUV machines and HBM chips are concentrated across a few allied nations, governments have established multilateral export controls to restrict their proliferation.

Led by the United States, and in coordination with the Netherlands, Japan, and South Korea, these controls operate on two primary fronts:

• Choking Manufacturing Capabilities: The United States and the Netherlands have restricted the sale and servicing of EUV lithography equipment to specific nations of concern, primarily China. By denying access to the tools required to print advanced circuits, these controls prevent targeted nations from domestically manufacturing their own frontier AI chips.
• Restricting Component Access: Recent updates to export rules directly target the finished components needed for AI training. The U.S. Department of Commerce utilizes the Foreign Direct Product Rule (FDPR) to restrict the global shipment of advanced HBM chips—and the logic chips paired with them—if they are made using U.S. software or equipment. This limits the ability of restricted entities to import the memory required to assemble AI accelerators locally.

Impact on Frontier AI Training

These targeted bottlenecks directly regulate who can train frontier AI models by limiting access to compute infrastructure.

Frontier AI training requires immense clusters of interconnected chips running continuously for months. Without EUV lithography, targeted entities must rely on older, less efficient deep ultraviolet (DUV) technology. While DUV can be pushed to create somewhat advanced chips through complex workarounds, it suffers from low manufacturing yields and sky-high costs, making it nearly impossible to produce AI processors at the scale required for massive data centers.

Simultaneously, the HBM restrictions ensure that even if a restricted organization designs a workaround processor, it cannot easily acquire the memory necessary to run large-scale AI workloads efficiently. By capping both the domestic production of logic chips and the importation of advanced memory, multilateral export controls force targeted nations to train AI on slower, less capable hardware. This drastically increases the time, energy, and financial cost required to achieve breakthroughs in frontier artificial intelligence.