H200 exports would substantially reduce America's AI compute advantage over China

This analysis combines estimates of AI chip production by US and Chinese chip design firms with estimates of AI chip smuggling. Impacts are measured in terms of the percentage of 2026-produced AI compute owned by different actors.

A key dynamic affecting relative compute share is whether developers will take advantage of higher-performance FP4 features on Blackwells in 2026, or just use FP8. Here, we assume a 50/50 mix. See full report for further scenarios and analysis.

* Assumes no export of US data center AI chips to China, including the H20
** Assumes NVIDIA has the same share of its revenue from China as it had pre-2022 chip export controls
† Assumes Chinese companies pursue a strategy consistent with stockpiling behavior already seen for SME, and equivalent chips from other US chip design companies are also appproved for export.
Chart: AuthorsSource: IFP analysis