██████╗ ██████╗ ██╗ ██╗ █████╗ ╚════██╗██╔═══██╗██║ ██║██╔══██╗ █████╔╝██║ ██║███████║███████║ ╚═══██╗██║ ██║██╔══██║██╔══██║ ██████╔╝╚██████╔╝██║ ██║██║ ██║ ╚═════╝ ╚═════╝ ╚═╝ ╚═╝╚═╝ ╚═╝
Welcome to 3OHA, a place for random notes, thoughts, and factoids that I want to share or remember
13 June 2026
"History doesn't repeat itself, but it often rhymes"
—Attributed to Mark Twain
Anthropic has just disabled access to its newest AI models, Fable 5 and Mythos 5, after receiving a US government export control directive that restricts use by foreign nationals due to national security concerns. The order forces the company to block access globally, including for users and employees outside the US, while leaving its other models unaffected. What strikes me is that this news is only unexpected to those unfamiliar with the recent history of controls over strategic computing systems. In particular, US export controls on high-performance computing have always followed a stable objective: limit access to scalable compute with military relevance, which closely fits this case. The target has chenged over time (from machines to metrics to users to supply chains) but the constraint remains the same.
In the 1980s, export controls were built around Cold War logic. Advanced computing systems were treated as strategic military assets due to their use in simulation, cryptography, and weapons design. Control was coordinated through CoCom (Coordinating Committee for Multilateral Export Controls), which restricted exports to the Soviet bloc and aligned states. Policy was direct: low performance thresholds, case-by-case licensing, and frequent outright denial. If a system was powerful enough to matter, it was likely restricted.The assumption was simple: computing capacity maps directly to military capability.
After the collapse of the Soviet Union, the strategic logic weakened slightly. The US shifted toward selective control rather than blanket denial. Export policy began to differentiate between allies and higher-risk states, with more focus on commercial competitiveness. In 1996, CoCom was replaced by the Wassenaar Arrangement, a regime that coordinated export control lists rather than enforcing strict bloc containment This caused a shift toward performance-based regulation. Export controls were formalized using MTOPS (Millions of Theoretical Operations Per Second), a numerical performance metric. Systems below thresholds could be exported more freely. Systems above thresholds required licenses, especially for higher-risk countries such as China and Russia. The system was intended to be structured, but performance thresholds became outdated quickly due to rapid hardware progress (something similar to what is currently happening to AI models). As computing performance scaled rapidly, MTOPS thresholds required constant revision. Political pressure also increased following reports of diverted HPC systems used in foreign military research. This caused a constantly shifting boundary between commercial and strategic computing.
By the 2000s, computing moved from single machines to distributed clusters and cloud systems. Hardware-based classification became insufficient, and policy shifted toward end-use and end-user controls (i.e., what the system is used for and who operates it became central). In other words, the assumption changed: computing is dual-use by default and context determines risk. This report (February 2001) provides a brilliant summary of the situation at the time and the need to shift to software protection.
The expansion of export controls beyond supercomputers and towards the AI ecosystem, including GPUs, AI accelerators, or semiconductor manufacturing equipment, seems to be a very natural move. As AI becomes more and more strategically important, both the compute capacity itself and the models will likely become a controlled resource.