The global artificial intelligence landscape is currently undergoing a seismic shift as geopolitical tensions rewrite the rules of technology export. Simon Glairy, a distinguished authority in risk management and AI-driven assessment, joins us to dissect the fallout from recent U.S. export bans on frontier models. With his deep background in navigating complex regulatory environments, Glairy provides a unique perspective on how regional players in Tokyo and Beijing are maneuvering to fill the vacuum left by American silicon giants.
Our discussion explores the strategic implications of the U.S. government’s restrictions on Anthropic’s most powerful tools and how these bans have inadvertently accelerated the development of localized “sovereign” AI. We delve into the rise of defensive and offensive cybersecurity AI in China, the emergence of orchestration models in Japan that prioritize collective intelligence over single-provider reliance, and the broader economic consequences for a multi-billion-dollar industry suddenly facing fragmented markets and a crisis of trust.
The Trump Administration recently enforced an export ban on Anthropic’s Mythos and Fable 5 models, effectively cutting off non-Americans from some of the world’s most advanced AI. From a risk management perspective, how do you see this sudden restriction altering the strategic landscape for international enterprises that were previously dependent on U.S. infrastructure?
The immediate fallout of this ban is a palpable sense of vulnerability among global enterprises that once viewed U.S. AI as a stable utility. When you consider that Anthropic’s run-rate revenue crossed a staggering $47 billion in May 2026, it becomes clear just how much capital and operational trust was flowing into these frontier models before the order took effect only two weeks ago. For a risk manager, the realization that access to top-tier models can disappear overnight is a nightmare scenario that forces an immediate pivot toward diversification. We are seeing a frantic re-evaluation of “concentration risk,” where relying on a single provider for national or corporate infrastructure is now seen as a critical failure point. This isn’t just about losing a tool; it’s about the sudden evaporation of a competitive edge, leaving companies to scramble for local alternatives that don’t carry the same “kill switch” baggage.
In the wake of these bans, Tokyo-based Sakana AI launched Fugu, a model they claim stands shoulder-to-shoulder with Mythos and Fable 5. Given their focus on Japanese language and culture, do you believe these localized frontier models can truly provide a viable hedge against U.S. export controls?
Sakana AI is playing a very sophisticated game by positioning Fugu not just as a replacement, but as a strategic hedge against the volatility of international politics. While their spokesperson insists the launch timing was coincidental, the fact that their website explicitly advertises “frontier capability without the risk of export controls” tells you exactly where the market’s head is at. Founded in 2023 by former Google and Stability AI heavyweights, Sakana is leaning into their research presented at ICLR to prove that you can deliver high-level value without the massive datasets typically associated with U.S. giants. Fugu, named after the Japanese blowfish, represents a specialized type of resilience; it is optimized for local nuance and designed to orchestrate access to other models through APIs. This approach suggests that for Japanese businesses and government agencies, the goal is no longer just “the biggest model,” but the most reliable one that remains firmly within their jurisdictional grasp.
While Japan seeks a “hedge,” China’s cybersecurity firm 360 has taken a more aggressive stance with the unveiling of Tulongfeng and Yitianzhen. What are the implications of treating vulnerability-finding AI as a “national strategic asset,” as described by Zhou Hongyi?
The move by 360 represents a hardening of the digital iron curtain, where AI for cybersecurity is now viewed through the same lens as kinetic weaponry. By launching Tulongfeng for vulnerability discovery and Yitianzhen for automated incident response, China is signaling that they will not accept what Zhou Hongyi calls “one-way transparency.” This is a deeply emotive concept in the tech world; it’s the fear that one nation could possess the “digital X-ray vision” to find flaws in everyone else’s software while keeping their own defenses opaque. From a global risk standpoint, this accelerates an AI arms race where the focus shifts from generative creativity to systemic exploitation and defense. When a firm like 360, which has been at the forefront of Chinese cybersecurity, frames these tools as strategic assets, it means the era of collaborative global security research is effectively over, replaced by siloed, nationalistic development.
David Ha of Sakana AI mentioned that “orchestration models” are the next frontier, moving beyond the race for bigger and bigger models. How does the concept of “collective intelligence” change the way we think about AI power concentration?
David Ha’s vision for orchestration is a direct response to the “bigger is better” philosophy that has dominated Silicon Valley, and it’s a brilliant tactical shift for smaller players. Instead of trying to build a singular, monolithic god-model that requires thousands of H100 GPUs, orchestration focuses on the ability to coordinate many specialized models to solve complex tasks. This “collective intelligence” acts as a practical hedge because it prevents a user from being held hostage by the export whims of a single government or the price hikes of a single provider. It’s a move toward a more modular, resilient ecosystem where the intelligence is distributed rather than centralized. If you can orchestrate a dozen smaller, localized models to match the performance of a banned frontier model like Mythos, you’ve essentially neutralized the power of the export control.
Despite the rise of these local alternatives, some industry leaders still argue that U.S. models remain essential to Asia. Do you think we are seeing a permanent realignment of the AI market, or is this just a temporary detour?
It is a mistake to think that U.S. influence will vanish overnight, but the trust that underpinned the “global AI” dream has been fundamentally fractured. Even though Sakana’s Ren Ito argued at the G7 summit that AI should be developed together and not “hoarded,” the reality on the ground is that local alternatives are already filling the gap. These regional models are being trained with a better grasp of local language and cultural nuances, which provides an organic advantage that even the most powerful U.S. model struggles to replicate. When a company can choose between a powerful but “risky” American model and a slightly less powerful but “safe” local model, the risk-averse choice is increasingly the local one. We may not see a total decoupling, but we are certainly entering an era of “multi-aligned” AI where the dominance of a few California labs is replaced by a more fragmented, competitive, and culturally specific global market.
What is your forecast for the future of sovereign AI?
I expect that within the next three years, we will see the emergence of “Sovereign AI Stacks” in every major economic bloc, where the hardware, data, and orchestration layers are entirely decoupled from foreign oversight. The $47 billion revenue milestone achieved by Anthropic was a peak of the old world; the new world will be defined by smaller, more efficient models like Fugu that prioritize jurisdictional security over raw parameter count. We will likely see a surge in government-backed “national AI refineries” that treat compute power and model weights as a utility as essential as electricity or water. Ultimately, the quest for AI supremacy will shift from “who has the largest model” to “who has the most unshakeable access,” making orchestration and localized fine-tuning the most valuable skills in the global tech economy.
