Notes
The AI chip supply chain is narrow and specialized, making it feasible to monitor production. The vast majority of AI chips are designed by NVIDIA. The most advanced logic chips (the main processor) used in AI chips are almost all fabricated by TSMC — accounting for around 90 percent of market share. Most AI chips are fabricated on versions of TSMC’s five-nanometer process node, a node likely only supported by two or three manufacturing plants. EUV lithography machines, a critical component in advanced logic chip fabrication, are made exclusively by ASML. High-bandwidth memory (HBM), another key component to AI chips, is dominated by two or three companies. This narrow and technical supply chain would be relatively easy to monitor and hard to clandestinely replicate.
Monitoring AI chip production would have relatively small spillover effects. While some of the same processes also produce other chips (e.g., smartphone chips), the chips themselves are easily differentiated. Chip design would change over time, but as a snapshot, current AI chips would be identifiable via their large high-bandwidth memory (HBM) capacity and specialized matrix-multiply components, among other factors.
When it comes to monitoring the AI chip supply chain, based on existing bottlenecks, a good start might be to monitor HBM production, logic die fabrication, and subsequent steps (e.g., packaging, testing, server assembly), along with key inputs such as EUV lithography machines.
Our Article states that sales of AI chips within Party states will have a presumption of approval, but does not indicate this presumption for AI chip manufacturing equipment. Chip sales are likely to have a relatively short-term effect on AI development capacity, as the lifecycle of AI chips is typically only a few years. By contrast, chip manufacturing capacity could lead to significant chip production for many years to come, and it would be especially concerning if a country became a Party, built up an AI chip supply chain, and then withdrew from the treaty. Therefore, we suggest more conservative restrictions on chip manufacturing equipment than on chips themselves.
Paragraphs 4 and 5 of this article permit the sale of AI chips and chip manufacturing equipment to Treaty Parties but not to non-Party States or other entities. That is, parties accept risks from chip manufacturing and concentration, but only in cases where the chips are subject to monitoring. The ability to manufacture and possess chips without a protective response from other states thus emerges as a positive incentive to join the treaty.
On its own, this does not prevent non-Treaty Parties from accessing AI chips in Treaty Parties remotely (i.e., cloud computing, or Infrastructure-as-a-Service), but such chips would be under ISIA monitoring to ensure they are not being used in violation of Article IV.
Restrictions on non-parties could go further, if need be. For example, non-parties could be banned from remote access to AI chips (i.e., from renting AI chips in Treaty countries via the cloud) or from accessing AI models via APIs.
If monitoring chip production and preventing smuggling were infeasible, another approach would be to ban all production of new AI chips. This approach would run less of a risk of chips being diverted, but it has the cost of losing the value that these chips could have produced in non-research, non-development AI applications. It would still rely on some monitoring of chip production facilities — e.g., to ensure they are only producing non-AI chips or that they are decommissioned. Our treaty design invites chip production to continue due to the large benefits their use might bring, with the dangers mitigated by monitoring (Article VII).