Sovereign AI After Fugu: From Vertical Dominance to Horizontal Resilience
The AI industry appears to be entering a period in which “vertical” frontier-model dominance will increasingly be recombined through “horizontal” orchestration. Up to now, the leading paradigm was that a small number of “frontier labs” sought to own the full capability stack through ever-larger monolithic models. But the emerging alternative changes the technical landscape. Fugu shows that AI orchestrators can now dynamically compose specialized agents and models from multiple providers, allowing users to access a single interface while the system itself coordinates diverse capabilities.
The “vertical/horizontal” framing is useful because it illuminates the core power dynamics at play. AI competition is no longer reducible to the question of who builds the single best model; it increasingly turns on who controls the interfaces, routing logic, and trusted environments that connect multiple layers of capability. For Japan and other middle powers that recognize the necessity of sovereign AI, the implications are especially important. Fugu suggests that sovereign AI may not require replicating the full vertical stack of US or Chinese frontier labs—it may also be pursued through trusted orchestration layers that combine, verify, and substitute among multiple models while preserving operational resilience.
The Vertical-to-Horizontal Shift
For the past several years, frontier AI development has been defined by vertical integration. Leading US labs such as OpenAI, Anthropic, Google, and others pursued massive scale in single models. Billions of dollars in compute investment, breakthroughs in scaling laws, and proprietary post-training techniques produced broad “God Models” capable of handling a wide range of tasks. These systems offered impressive coherence and performance, and they created strong competitive moats rooted in proprietary weights, data, infrastructure, and brand.
Yet the same vertical model also concentrated risk. A single policy or regulatory decision, outage, pricing change, safety restriction, or performance regression could disrupt entire workflows, leaving users—and even nations—dependent on a handful of mostly US-based providers for critical operational capabilities. The economic parallel is a highly integrated manufacturer owning every stage from raw materials to final assembly: efficient when stable, but fragile when any key link is severed.
Horizontal orchestration changes this calculus. Systems like Fugu present users with a single, convenient interface while operating as intelligent coordinators behind the scenes. They route subtasks to specialized agents optimized for different functions, from reasoning and coding to verification and domain expertise. These agents can draw on frontier, open-weight, domestic, or lower-cost models, while the orchestrator manages allocation, redundancy, and final synthesis. The result may approximate—or, in some contexts, exceed—single-model performance with greater resilience and adaptability.
“Horizontal” is a useful shorthand, though the underlying architecture is more dynamic than the term implies. Real orchestration is closer to networked coordination than side-by-side aggregation. The key point is that capability can increasingly be produced through composition across a networked stack, reducing reliance on any single vertically integrated model.
Still, the vertical/horizontal lens captures the strategic, business, and geopolitical stakes involved: orchestration means that the competitive moats begin to erode. Frontier models, compute, talent, and data remain central, but they no longer exhaust the sources of advantage. As orchestrators learn which models perform best in particular workflows, how tasks should be routed, and how outputs should be verified, they can generate a new layer of strategic intelligence. This reduces the dominance of individual model laboratories at the margin, creating new opportunities for systems integrators and enterprise platforms—as well as middle powers.
In short, vertical capability supplies many of the strongest components, while horizontal orchestration determines how those components are assembled, substituted, managed, and then embedded into real workflows. The strategic stack is therefore becoming more layered.
Sovereign AI and Chokepoint Navigation
The urgency of this shift became evident earlier this month when US export-control action prompted Anthropic to suspend foreign access to its latest Fable 5 and Mythos models. Governments, enterprises, and researchers outside the US suddenly faced uncertainty over access to frontier capability. The episode underscored a core vulnerability of vertical dependence: intelligence itself can become a chokepoint asset subject to geopolitical leverage.
Sovereign AI—national efforts to develop, operate, or reliably access advanced AI systems under trusted conditions—has therefore become strategically necessary, not just aspirational. No responsible government wants core economic, security, or governance functions to depend entirely on a single foreign provider whose access can be curtailed by policy decisions made elsewhere. This is especially true when AI systems are increasingly embedded in cybersecurity, industrial planning, infrastructure management, scientific research, and public administration.
Horizontal orchestration directly addresses part of this problem by enabling dynamic rerouting. If one model or provider becomes restricted, unavailable, or commercially unattractive, the system can substitute alternatives without collapsing overall performance. Domestic models, open-weight leaders, accessible international components, and specialized domain agents can be combined intelligently. The result is a form of “technological optionality”: users retain access to advanced capability even when the composition of the underlying stack changes.
This does not make export controls obsolete—they can still have a significant impact. A country cut off from advanced compute may struggle to train or run the most capable systems at scale, and a firm denied access to a leading proprietary model may still face a meaningful performance loss. Orchestration reduces exposure to any single model chokepoint, but it does not eliminate dependence on the broader AI ecosystem.
The strategic effect is therefore quite subtle. Horizontal systems make model-access controls less decisive when users can combine available systems into more capable composites. Strategic competition would then shift from denying discrete frontier assets toward shaping the broader environment in which models are accessed, evaluated, and finally deployed.
Fugu makes this point especially relevant for Japan, since Sakana AI exemplifies how a middle power can contribute meaningfully in that environment. Japan is unlikely to reproduce the full scale of either the US or Chinese AI ecosystems. It can, however, build trusted orchestration and integration layers that increase national and allied resilience. In this sense, Fugu points toward a more realistic form of sovereign AI: control over strategic interfaces and integration capacity rather than self-sufficiency across every layer.
Implications for the US-China AI Race
This transition also reframes the US-China competition over AI leadership, as both governments deploy resources to control a technology that will be critical to both economic growth and military power.
The US retains formidable vertical strengths: frontier-model innovation, hyperscaler compute, advanced semiconductor access, and the talent and software ecosystems that sustain them. These advantages produced the current generation of leading models and underpin US leverage through export controls and cloud infrastructure.
However, excessive concentration on protecting premium vertical assets carries risks if the Trump administration underestimates the emerging horizontal layers. If capability increasingly depends on orchestration, deployment speed, integration depth, cost efficiency, and real-world feedback, then the basis of competition becomes more multidimensional. Owning the strongest model remains a major advantage, but the strategic value of that model depends on how effectively it is embedded into workflows, combined with other tools, and translated into productivity gains.
China may be structurally well-positioned for parts of this emerging horizontal environment. Chinese firms have shown strength in turning AI systems into lower-cost, rapidly deployed applications. That aligns with broader features of China’s political economy: supply-chain dominance, dense industrial ecosystems, and the capacity to scale technologies quickly into real-world use cases. If orchestration rewards adaptive assembly and practical deployment, some of China’s existing industrial advantages may become more relevant.
That being said, China still faces constraints in original frontier breakthroughs, access to the most advanced chips, and the governance of complex AI systems. Multi-agent architectures can also reveal reliability problems when tasks require judgment, verification, safety, or cross-domain synthesis. And even sophisticated orchestration remains partly dependent on advanced components that may be subject to international controls.
The more plausible scenario is a broadened contest rather than a decisive shift in favor of either side. Crucially, both powers will likely seek to combine vertical innovation with horizontal resilience. The US-China race could therefore move from a single contest over frontier-model performance toward a broader struggle to turn AI capability into reliable, affordable, and widely deployed systems.
A useful historical parallel may be the US-Japan semiconductor competition of the 1980s, which shows how leadership in one layer of a strategic technology does not automatically translate into durable dominance across the whole stack. Japan’s strengths in manufacturing quality and process integration were formidable, but the US eventually adapted through trade pressure, alliance coordination, and the rise of the fabless/foundry model enabled by TSMC.
The lesson for the current AI competition is not that US vertical leadership is destined to erode; rather, leadership must be renewed when the decisive terrain shifts. For the US, that means translating frontier-model strength into robust orchestration, deployment, and alliance ecosystems. For China, it means ensuring that integration advantages do not become rigidities if the technological frontier shifts again.
However, for middle powers, the US-China context creates a risk of binary dependence: relying too heavily on either US-controlled frontier models or China-centered open ecosystems. Yet as AI capability becomes more composable, the space between the two AI great powers may widen. Countries with strong engineering, institutional trust, and domain expertise can build valuable horizontal positions without vertically owning every foundational layer.
Japan and Middle-Power Optionality
Japan’s opportunity lies in making horizontal composability operational. If advanced AI capability can be assembled from multiple models and deployment environments, then Japan’s strategic task is to control the points where those components are made reliable, trusted, and ultimately most useful. Fugu illustrates how integration itself can become a source of sovereign capability—from an economic security standpoint, that would allow Japan to remain connected to global AI innovation while retaining optionality under geopolitical uncertainty.
The same logic applies to allied cooperation. From the perspective of middle powers, sovereign AI will be difficult to sustain if national systems fragment into incompatible silos, and this could generate a shared interest in keeping trusted AI ecosystems interoperable across the model and infrastructure layers. Indeed, interoperability would widen the room for Japan and likeminded partners to maneuver by reducing the pressure to choose between dependence on a single foreign provider and costly technological self-sufficiency.
Europe faces a similar dilemma. Like Japan, it must avoid excessive dependence on US frontier providers without drifting into costly technological isolation. Orchestration may offer one way to manage that tension: advanced economies outside the US-China core can remain connected to global AI capability while focusing on building trusted integration systems for domestic deployment. In that sense, the middle-power logic of Fugu may also apply to other technologically advanced actors that seek some measure of AI autonomy.
Another important arena is the Global South. If China can offer low-cost AI systems bundled with cloud services and industrial applications, it may gain influence through practical deployment. But Japan and its partners might find a vast market for credible and attractive alternatives that allow countries to combine models and tools from multiple sources, rather than becoming locked into one provider’s stack.
Fugu should therefore be understood as more than a Japanese AI product—it is a signal of where middle-power strategy may be heading. The most important capabilities may lie in integrating and governing intelligence across a distributed ecosystem. For middle powers seeking sovereign AI capabilities, that may become both a plausible objective and a strategically meaningful role.
Conclusion
The emergence of horizontal AI orchestration marks an important evolution in a technology that increasingly underpins economic and national power. By moving beyond reliance on single vertical stacks toward resilient, composable systems, the field becomes more distributed and adaptive. But it will also become more strategically contested.
Fugu does not prove that frontier-model dominance is over. Nevertheless, it does suggest that the strategic value of frontier models will depend on how they are orchestrated, embedded, secured, and translated into real-world outcomes. In a world of sovereign AI, where capability is distributed across multiple layers, actors that combine vertical innovation with horizontal resilience may find the competitive landscape more favorable.
For the US and China, this means the AI race will be fought across the full stack rather than at the model layer alone. And for Japan and other middle powers, it creates a wider field of strategic action. In short, against the backdrop of US-China AI competition, countries that cannot vertically dominate every layer can still potentially build meaningful autonomy and influence by controlling trusted interfaces and high-value domains of application.
Sovereign AI after Fugu will likely be less about technological self-sufficiency than strategic optionality. Japan may find itself well placed to help define that agenda.
(Photo Credit: Shutterstock)
Disclaimer: The views expressed in this IOG Commentary do not necessarily reflect those of the Institute of Geoeconomics (IOG) or any other organizations to which the author belongs.
