Background & Context§
The rapid escalation of US AI export controls has created an unexpected vacuum in global frontier model access. On June 13, 2026, the Trump administration imposed a ban preventing Anthropic from distributing its most advanced models—Mythos and Fable 5—to non-US entities, citing national security concerns. Within two weeks, two Asian startups have seized the opportunity: Japan's Sakana AI launched Fugu, an agentic orchestration model claiming parity with Fable 5 and Mythos Preview, while China's 360 unveiled Tulongfeng, a cybersecurity tool designed to rival Mythos specifically in vulnerability discovery. These launches signal a potential realignment in the global AI supply chain, where local alternatives are filling gaps left by US export restrictions.
The News: What Happened Exactly§
On June 25, 2026, Sakana AI—a Tokyo-based startup founded by Google alumni David Ha and Llion Jones, and former Mercari/Stability AI executive Ren Ito—announced Fugu, a frontier model named after the Japanese blowfish. Sakana claims Fugu "stands shoulder-to-shoulder with leading models like Anthropic’s Fable 5 and Mythos Preview." The model is designed for agentic use cases, with built-in capabilities to orchestrate access to other models via APIs. A spokesperson emphasized that the launch was "entirely coincidental" with the export ban, stating that research for Fugu had been presented at ICLR earlier in the spring. However, Sakana's website prominently advertises "delivering frontier capability without the risk of export controls," directly appealing to customers concerned about supply disruptions.
Two days earlier, Chinese cybersecurity firm 360 unveiled Tulongfeng, an AI tool for automatic software vulnerability discovery, which it claims can go head-to-head with Anthropic’s Mythos. A companion tool, Yitianzhen (autonomous cyber defense and incident response), was also launched. According to Reuters, 360’s founder Zhou Hongyi described vulnerability-finding AI as a "national strategic asset" and warned of "one-way transparency"—a scenario where some actors possess advanced vulnerability detection while others do not. 360 did not respond to a request for comment.
Agentic Orchestration: Fugu's Technical Differentiator§
Sakana's Fugu is not merely a larger language model but an orchestration model—a meta-controller that coordinates multiple models via APIs. David Ha wrote on X: "Orchestration Models are the next frontier, beyond bigger models." This architecture addresses a critical risk highlighted by the export ban: reliance on a single provider. Ha argued that "access to top models can disappear overnight" and that "collective intelligence is the practical hedge against this concentration of power." Fugu is optimized for Japanese language and culture, targeting businesses and government agencies seeking to reduce exposure to tightening export controls while maintaining frontier AI capabilities.
Market Impact and Strategic Positioning§
Sakana's spokesperson acknowledged that US models remain important to Asia, characterizing the moment as a temporary "realignment toward any one set of players" rather than a permanent shift. Co-founder Ren Ito published an op-ed in Project Syndicate urging the US to prioritize preserving access for its closest allies, arguing that "AI should not become a technology that is hoarded; it should be one that is developed together." Meanwhile, Anthropic's run-rate revenue had crossed $47 billion in May 2026, though its dependence on Asian enterprise customers is unknown. Since the ban, at least two companies—one in Tokyo, one in Beijing—have stepped into the gap, with local models trained to better understand local language and nuance.
Historical Parallels & Similar Incidents§
This situation echoes the Huawei ban of 2019, when the US restricted Chinese telecom giant from accessing Google Mobile Services and semiconductor supplies. In response, Huawei launched HarmonyOS and developed its own app ecosystem, initially positioning it as a hedge but eventually pivoting to a full-scale alternative. Similarly, Sakana is framing Fugu as a hedge, not a replacement, but the underlying dynamic—where export controls accelerate local innovation—is identical. The key difference: Huawei's self-sufficiency took years, whereas Sakana and 360 are releasing viable frontier models within two weeks, suggesting the barrier to competitive model development has lowered significantly since 2019.
Another parallel is the Chinese AI model boom following US GPU export restrictions in October 2022. When the US banned sales of NVIDIA A100 and H100 chips to China, Chinese firms like Baidu (ERNIE Bot), Alibaba (Qwen), and SenseTime developed their own models trained on alternative hardware (e.g., Huawei's Ascend chips) or using more efficient architectures like mixture-of-experts. The current export ban targeting Anthropic's models mirrors that hardware ban but focuses on software. The speed of adaptation—two weeks versus months—reveals how quickly the ecosystem has matured. However, unlike hardware bottlenecks, software bans are easier to circumvent; Anthropic's models remain available via VPNs or proxies, but enterprise customers face legal risks. This creates a distinct market for certified, compliant alternatives like Fugu and Tulongfeng.
A third instructive case is OpenAI's API ban in China in 2024, which led to the rapid emergence of domestic alternatives such as Baidu's ERNIE API and Alibaba's Tongyi Qianwen API. Within weeks, Chinese developers migrated to local providers, effectively decoupling from US AI infrastructure. Sakana and 360 are now playing an analogous role for Asian enterprises that rely on Anthropic's models for national cybersecurity or agentic workflows. The lesson from 2024: once alternatives achieve comparable quality, switching costs drop, and the market may not revert even if bans are lifted. As Sakana's spokesperson noted, U.S. models remain important, but "local alternatives...are already filling the gap." The question is whether this is a temporary workaround or the beginning of a balkanized AI landscape.