Anthropic’s advanced AI model Methuselah is hindered by strict US export controls that prevent international collaboration and revenue generation, slowing innovation and allowing competitors like Japan’s Sakana AI and Chinese labs to catch up. These regulations, intended to protect US AI leadership, may instead weaken it by restricting access and favoring more accessible, compliant models preferred by global customers.
Anthropic developed a groundbreaking AI model called Methuselah, which is so advanced that it falls under strict US export controls. While this might seem like a competitive advantage, it actually creates significant challenges. Since Methuselah cannot be shipped to paying customers abroad, it generates no revenue despite the high costs involved in training and running such frontier models. This makes Methuselah both Anthropic’s greatest asset and its biggest liability.
A more critical issue arises from US deemed export rules, which treat giving foreign nationals access to controlled technology as equivalent to exporting it to their home countries. Anthropic’s research team is highly international, meaning many of the engineers who build and improve Methuselah are legally barred from working with it. This creates a paradox where the protective moat around the technology actually hampers the very people who are supposed to advance it, slowing innovation and iteration.
Meanwhile, competitors like Japan’s Sakana AI are making strides with models like Fugu, which claims to rival Methuselah on key benchmarks such as coding, reasoning, and scientific tasks. Fugu achieves this not by being a single massive model but by orchestrating a team of publicly available models to work together dynamically. Although independent verification is pending and Fugu currently struggles with long-running tasks, this approach challenges the notion that a single secretive model is the only path to frontier AI capabilities.
The broader implication is that US export controls may inadvertently weaken American AI leadership. Chinese labs, such as CIA with their GLM 5.2 model, are rapidly advancing outside the US regulatory framework and are expected to reach frontier levels soon. This means that restrictions designed to protect US dominance might instead slow domestic labs while allowing foreign competitors to catch up or surpass them without similar constraints.
Finally, demand is shifting toward AI infrastructure that complies with local laws and regulations, including GDPR and data privacy rules, and that avoids US government backdoors. For many European and Asian customers, models that are frontier-level yet compliant are more attractive than those that are cutting-edge but restricted and inaccessible. In this evolving landscape, Anthropic’s protective moat may have backfired, potentially undermining the dominance of US AI labs and opening the door for international competitors.