US policy makers are most likely operating off of incomplete data. Recently I've seen numerous citations of this report by the think tank Epoch AI (which shows that the US controls 75% of "global GPU cluster performance" while China controls only 15%):
. The report appears to be well-researched at first glance (particularly for US GPU clusters), but if you read the actual article they published, it will indicate that their estimates for China are flawed, as their research criteria for China consisted of "two Chinese-language analysts conducted targeted searches of systems in China and Hong Kong" and a redacted Approaches section:
B.3. Approach for finding Chinese AI supercomputers
We decided to redact our approach to finding Chinese AI supercomputers and avoid providing identifying information about them throughout the paper to preserve data sources. We take this step as a precautionary measure because Chinese websites cited in public reports have been redacted or replaced with malware in the past (Wei, 2023). If you would like to request access to our methodology for Chinese AI supercomputers, please contact Konstantin at
[email protected]
This is a garbage in, garbage out situation. US policy makers are convinced their sanctions are working because data sets like these bolster their claims. Even if, in the back of their heads, they're aware that the estimates could be wrong, from a career perspective it is much better for politicians to shift the blame to the data providers and to say that they did the best they could, given the data they had, than to really question the data.