Historically, "Share" or "Hoard" Has Always Been a Choice

Ren Ito says history offers a reference for how to handle a leading technology. During the Cold War, the US repeatedly faced this same choice: should an advanced technology be shared with allies, or kept to itself.

He cites two sets of examples. During the Cold War, the US shared advanced cryptographic technology with trusted allies, because secure communications made the whole alliance more resilient. Decades later, fighter jet development followed the same logic — just split across two paths.

Both logics have always coexisted. Ren Ito emphasizes that the strategic question here was never just "how to protect the technology," but "who gets to use it, in a way that strengthens both national security and allied resilience."

But AI Isn't Like Cryptography or Fighter Jets

Ren Ito says frontier AI faces the same strategic question, but with one crucial difference. Cryptography and fighter jets were single technologies that could be settled by one decision to "share or not share." Frontier AI is not like that.

Drawing out this shift makes the difference between the two mental models obvious at a glance.

No single model dominates every task for long, and new entrants keep joining the pool. The question has shifted from "who owns that one model" to "who can always pick the best-suited one at any given moment."

The Key Capability Is On-Demand Access to the Best AI

Precisely because AI is an ecosystem, a new strategic capability has surfaced: AI orchestration. Ren Ito argues this has become a national-level skill.

He runs the numbers: for most countries, trying to replicate every frontier model domestically at home would burn through enormous money and talent, and still might not keep pace with the global frontier. A more realistic goal, which he calls "strategic autonomy."

The Ability to Switch Models Anytime Is Itself a Deterrent

Ren Ito points out this approach also brings a new form of deterrence.

You can try this demo yourself: see how a system capable of orchestrating multiple models responds when access to one model is cut off. Click the tabs above to switch which model gets cut.

That's the difference: bet on one model, and losing access means going dark; keep multiple ready to swap in, and cutting one off just means someone else steps in. The weight that "cutting off access" carries as leverage gets diluted along with it.

Competition Has Shifted from "Building Models" to "Building Systems"

Ren Ito says getting access to frontier AI is only the first half of the story — the second half is building the AI system.

Today's AI applications increasingly rely on assembly: multiple frontier models, plus specialized domain models, software agents, proprietary data, and human expertise, combined into one complete package.

Competitive advantage has shifted accordingly, from any single model to the ability to evaluate, orchestrate, and continuously improve the whole system. He believes Japan is well-positioned for this round: Japan's long-standing strength has been integrating various technologies into trustworthy systems. In a multi-model world, that means investing in independent evaluation, interoperability, and secure deployment — with no need to replicate every frontier model in-house.

The Meaning of "Trust" Has Changed

Until now, discussions of "trustworthy AI" have mostly focused on the technical level: alignment, transparency, bias, misinformation, harmful outputs. Ren Ito says these still matter, but for governments, companies, and society, the meaning of "trust" has now expanded.

Trust is no longer determined solely by one model's technical properties. Ren Ito says a trustworthy AI ecosystem must let a nation count on continued access, truly control its own data, and not be held hostage by decisions made elsewhere. Trust is increasingly a property of the institutions, governance arrangements, and international partnerships surrounding the model.

This Race Will Eventually Spread Beyond AI

Ren Ito predicts this shift's impact will extend far beyond AI itself.

He believes the next round of geopolitical competition will expand from military power to AI-driven critical infrastructure — ports, power grids, telecom networks, data centers. Who builds these systems is only part of it; what matters more is who operates them, who governs them, and whose AI they ultimately depend on. Japan has a distinctive position here too: long skilled at building trustworthy infrastructure, sustaining long-term partnerships through reliability and respect for other nations' sovereignty — qualities that may matter just as much in the AI era.

The article closes by returning to its central argument. This diagram lays the two generations of competition side by side:

Ren Ito places the two generations' scoring criteria side by side: the first AI race rewarded those who built the strongest model; the next AI race rewards those who build the strongest system. He believes AI's trajectory won't be decided solely by who builds the smartest model, but also by who can reliably access frontier AI, integrate it into trustworthy systems, and govern those systems responsibly. This, in his view, is the race Japan is positioned to lead.

Ren Ito's judgment: everyone kept asking "whose model is strongest," but that was the wrong question.

Because AI isn't a single thing like cryptography or fighter jets that can be settled by one "share or hoard" decision — it's an ecosystem: multiple frontier models with their own strengths coexist simultaneously, and new ones keep emerging fast. Since no single one can dominate forever, the question that actually needs asking has changed too.

For most countries, replicating every frontier model in-house is both expensive and unlikely to keep pace; more realistic is preserving the freedom to use whichever AI you want (what the author calls strategic autonomy), while building strong capabilities in selection, integration, and security. This skill is called AI orchestration — the figure below shows how it becomes a form of deterrence.

Ren Ito draws an analogy: a multi-model system is like a diversified supply chain — not betting everything on one supplier. Say XiaoHu's team is running a task and needs the strongest AI — see what happens under "cut off" access with two different setups.

Zooming out, Ren Ito says the whole center of competition is shifting.

Getting access to frontier AI is only the first half; the second half is assembling multiple frontier models, specialized models, software agents, proprietary data, and human experience into one complete system. Even the meaning of "trust" has expanded — from "is the model itself reliable" to "will you always have access, can you truly control your own data." His definitions for the two generations of competition are:

Apart from the June 12 export controls, which are an established fact, everything else in the article is Ren Ito's (Sakana AI co-founder, former Japanese diplomat) judgment and argument — no reproducible measured data.