The Crown Jewels and the Roads

Why America needs roads, not only crown jewels, in the age of open-weight AI

America may be making a category error in artificial intelligence.

Not a foolish error. Not an irrational one. Possibly not even an avoidable one. Frontier AI systems are becoming powerful enough that no serious government can treat their release as an ordinary product launch. If a model can materially assist cyber operations, vulnerability discovery, autonomous research, biological reasoning, or military planning, then “just ship it” stops sounding like innovation and starts sounding like negligence.

So the instinct to inspect, delay, gate, and approve is understandable.

But understandable is not the same thing as complete.

The United States appears to be moving toward a crown jewels mindset in AI: protect the most capable models, control preview access, restrict risky release, and keep the very best systems inside trusted circles. This is not inherently wrong. Crown jewels are valuable. If a frontier model crosses dangerous capability thresholds, the government has a legitimate interest in knowing who can use it and under what conditions.

The problem is that the crown jewels may not be the whole prize.

China seems to be moving with a different intuition. You do not always need to own the most prestigious object in a market to shape the market. You can win by becoming the thing other people build on.

That is a different game.

The West often overvalues the prestige tier because it is visible, glamorous, and easy to rank. The fastest car. The best phone. The model at the top of the leaderboard. The demo that makes everyone gasp. This kind of excellence matters, and America remains extremely good at producing it. But prestige can seduce policymakers into thinking that the winner of the future is simply the entity with the most impressive flagship.

History is not so tidy.

Honda and Toyota did not need every driver to want a supercar. Civics and Corollas did just fine. Android did not need to be more prestigious than Apple. It became the global default for billions of people. Linux did not need a beautiful consumer brand to become central to servers, infrastructure, and development. Solar panels did not need to be luxury objects to transform energy economics. Generic drugs did not need cultural cachet to matter.

The default object wins by becoming ordinary.

That is the strategic importance of open-weight AI.

If Chinese models become the Civic, Corolla, Android, Linux, or commodity solar panel of artificial intelligence, that is not a consolation prize. That is a serious form of success. It means their systems become part of the everyday layer on which other people build. Startups use them. Universities teach with them. Governments deploy them locally. Small firms fine-tune them. Robotics companies embed them. Developers become fluent in them. Toolchains form around them. Procurement departments normalize them. A thousand invisible habits accumulate.

The leaderboard winner gets applause.

The substrate gets inherited.

This distinction matters because AI is not only a consumer product. It is closer to infrastructure, language, software, electricity, and labor rolled into one. A slightly weaker but open, cheap, customizable, locally deployable model can beat a more capable closed model across many practical domains. Not because it is smarter in the abstract, but because it fits into the world more easily.

The most important question may not be: who has the best model?

It may be: whose model becomes the thing everyone else assumes is available?

A frontier U.S. model locked behind remote access, compliance terms, political review, export restrictions, customer approval, shifting safety policies, and subscription pricing may remain technically superior. It may also become strategically narrower. Useful, prestigious, and lucrative, yes. But narrower.

A Chinese open-weight model that is good enough can spread into the places where friction matters more than perfection.

Good enough is an underrated force in history. Good enough gets installed. Good enough gets localized. Good enough gets taught in bootcamps. Good enough gets baked into internal workflows. Good enough gets used by institutions that cannot rely on U.S. cloud access. Good enough gets loaded onto local servers by companies worried about data sovereignty. Good enough gets modified by hobbyists, researchers, and regional developers who were never going to get privileged access to the newest American frontier model anyway.

Good enough is how roads get built.

This is where American policy may expose itself to an uncomfortable risk. A government that thinks in crown jewels will focus on controlling the highest-value artifact. It will ask: who gets access to the latest model? Who counts as trusted? Which customers can preview it? Which capabilities must be tested before broader release?

Those are legitimate questions.

But the substrate question is different. It asks: what do developers across the world build on while the crown jewels are being guarded? What models become familiar? What systems become cheap? What APIs, weights, fine-tuning habits, deployment patterns, local accelerators, and integration layers become normal?

The crown jewels mindset assumes leadership flows from possessing the best thing.

The substrate mindset understands that leadership can flow from being the thing underneath everything else.

This is where China’s open-source push becomes interesting without needing to turn it into a cartoon. It does not have to be treated as villainy to be strategically important. It can be seen as a rational response to the shape of the market, the limits of export controls, the needs of domestic industry, and the obvious fact that most users do not need the world’s most powerful model for every task.

Open-weight models lower the cost of experimentation. They invite adaptation. They allow foreign users to deploy locally rather than rely on a remote American provider. They give smaller countries and companies a way to participate without waiting for approval from the U.S. state or from a San Francisco lab. They create familiarity, and familiarity is one of the most underestimated forms of power.

That influence does not need to look dramatic. It can look like convenience.

A developer does not need to declare allegiance to Chinese AI to use a Chinese model. A startup does not need to care about geopolitics if the model is free, capable, documented, and easy to run. A government does not need to prefer Beijing to prefer local deployment over dependence on a foreign cloud provider. A researcher does not need to think about national strategy when the weights are available and the alternative is not.

This is how substrate power works. It does not always announce itself as power. Sometimes it arrives as a download button.

The U.S. answer cannot simply be reckless openness. That would be unserious. Public systems should refuse obvious abuse. Labs should evaluate dangerous capabilities. Cloud platforms should govern tool access, monitor misuse, and maintain real accountability. Some frontier systems may deserve narrower release, especially where the risks involve cyber operations, autonomous action, biological reasoning, or military use.

But that is not the same thing as making the gate the strategy.

This is where the policy imagination can go wrong. Access control is visible. Refusals are visible. Preview approval is visible. A government can point to the gate and say it acted. A company can point to the guardrail and say it governed. The model refused. The model did not refuse. The release was delayed. The approved users were narrowed. Everyone argues over the wrapper.

Meanwhile, the harder question waits underneath.

What can the model touch?

That is the question a downstream safety regime has to ask. A model sitting in a chat window is one thing. A model connected to credentials, code repositories, payment systems, procurement channels, infrastructure dashboards, messaging platforms, lab services, robotics systems, or cloud accounts is something else entirely. The danger does not live only in the sentence the model produces. It lives in the path from sentence to action.

So the serious security challenge is not only controlling access to cognition. It is governing the conditions under which cognition becomes consequence.

This matters because the gate becomes weaker as capability diffuses. Cloud models remain governable in ways local models do not. A cloud model has an operator, an account system, a corporate entity, a deployment environment, and a legal address. Governments can pressure it. Companies can patch it. Access can be suspended. Logs can be reviewed. Tools can be limited.

A local model changes the geometry.

Once useful models can be copied, compressed, modified, merged, quantized, privately run, and connected to tools outside the official perimeter, the policy layer becomes less stable. Refusal behavior can be weakened. Access controls can be routed around. Official channels can close while mirrors, derivatives, local copies, and alternative ecosystems continue moving. The point is not that upstream control becomes useless. It is that upstream control becomes incomplete.

That distinction should be central to American strategy.

If Washington responds to AI risk mostly by narrowing access to prestige systems, it may get the worst of both worlds. It may not stop determined misuse, because motivated actors will route around official gates. And it may slow responsible users, researchers, defenders, startups, universities, and smaller institutions that actually remain inside the compliance perimeter.

Bad policy can weaken defenders faster than attackers.

This is especially true in cybersecurity. The same general capability that helps an attacker reason through misuse can also help a hospital, municipality, school district, utility, or small business understand its own exposure. If strong tools become suspect by default, the institutions most likely to obey the rules may become slower and less capable, while serious adversaries move to open models, domestic alternatives, stolen access, local deployments, or whatever else works.

That is not safety.

It is self-disarmament with paperwork.

The better American response would pair frontier caution with defensive acceleration. Keep refusals for public systems. Evaluate dangerous capabilities. Limit tool access where action becomes consequential. But also build responsible access channels for defenders, protected research pathways, stronger infrastructure, better authentication, safer permissions, audit trails, rapid patching, resilient public systems, and practical downstream controls.

In other words: secure the roads, not only the palace.

This is where the substrate argument becomes more than an adoption story. If China’s open-weight systems become widely used because they are available, adaptable, local, and cheap, they will not merely gain market share. They will shape habits, workflows, defensive practices, deployment assumptions, and developer culture. They will become part of the road network.

A U.S. strategy focused too narrowly on guarding the crown jewels risks misunderstanding both sides of the problem. It treats safety as if the decisive layer is access to the best model. It treats leadership as if the decisive layer is possession of the best model. But in a world of diffusing capability, neither assumption is enough.

Safety moves downstream.

Strategy does too.

This matters especially outside the wealthy core of the West. In the United States, Canada, Japan, Britain, and parts of Europe, many people may continue to use the best proprietary systems because they can afford them and because they remain inside the trust perimeter. But much of the world will care about different things: cost, deployability, language support, local control, hardware efficiency, and freedom from permission structures.

If Washington treats AI leadership as a luxury race, it may misunderstand the mass market. If it treats frontier models as crown jewels but neglects the default layer, it may win the showroom and lose the factory floor.

There is an older American habit at work here. The United States is very good at apex technology. It knows how to produce the best chips, the best labs, the best aircraft, the best software firms, the best universities, and the most dazzling launches. That culture is powerful. It should not be dismissed.

But AI may not reward only apex thinking.

There is also no reason America has to remain trapped in a single lane. The early AI market made frontier access feel like the whole market because anything far below the frontier was often too weak to matter. When the technology was immature, not frontier often meant not useful enough. But that condition may not last. As models become broadly competent, the market can bifurcate. One lane can remain the Ferrari lane: maximum capability, maximum compute, maximum security scrutiny, and maximum strategic sensitivity. Another lane can become the grocery-getter lane: older, cheaper, less glamorous systems that are still perfectly serviceable for schools, small businesses, municipalities, internal tools, local deployments, customer support, coding assistance, document work, and ordinary institutional automation.

Something similar happened in games. For decades, the console wars trained people to talk as if technical progress meant one thing: better graphics, bigger worlds, more realism, more spectacle. From the 8-bit era through the PlayStation 4 generation, that made a certain amount of sense. The medium was still climbing toward a mature baseline. But once that baseline became good enough, the market fractured into lanes. Some players still want the game that melts a high-end graphics card. Others spend hundreds of hours in Stardew Valley, Hollow Knight, pixel art roguelikes, cozy farms, handheld indies, and stylized worlds that barely care about photorealism at all.

AI may follow a similar pattern. While the technology was immature, anything far below the frontier often felt useless. But once the baseline becomes broadly competent, the question changes. Not every task needs the latest Ferrari model. Some institutions will want the safest, fastest, most capable system available. Others will want the reliable station wagon: cheap to run, easy to host, good enough for ordinary work, and available without drama. At maturity, the market does not only ask what is most powerful. It asks what fits.

That second lane could become strategically important. Older data centers, cheaper inference clusters, smaller models, distilled systems, and non-frontier deployments may all gain second lives as the frontier buildout continues. The question is whether the United States recognizes this as strategy rather than leftovers. A mature AI ecosystem does not need every workload to run on the newest model. It needs the right capability at the right cost, with the right permissions, in the right place. America can still build that layer. But it has to want the roads, not only the showroom.

AI may reward distribution, adaptation, habit formation, local embedding, and ecosystem gravity. It may reward the model that becomes boring because everyone has already built around it. The future may not be decided by which model gets the loudest applause on launch day. It may be decided by which model becomes the assumed component in a thousand ordinary systems.

There is a reason Android matters.

Android did not need to make every affluent American abandon the iPhone. It became the default mobile operating system for much of the planet. That gave it enormous ecosystem power. App developers had to care. Hardware makers had to care. Accessory makers had to care. Users across income levels and regions entered the smartphone era through it.

A Chinese open-weight AI ecosystem could do something similar. Not as a phone operating system, but as an intelligence operating layer. The model that teaches millions of developers how to build with AI becomes more than a model. It becomes a habit. The model that runs locally in schools, municipal systems, small factories, and regional startups becomes part of institutional muscle memory. The model that everyone fine-tunes becomes the model everyone understands.

This is not the glamorous version of victory.

It may be the durable one.

The United States still has enormous advantages: frontier labs, capital markets, chip ecosystems, cloud infrastructure, elite talent, enterprise trust, and deep integration with allied economies. It can still lead. But leading will require more than guarding the most advanced systems. It will require a serious answer to the substrate problem and a serious answer to the downstream safety problem.

Those are connected.

If the official American answer to AI risk becomes slower access, narrower previews, trusted customer lists, and increasingly political gates, then the rest of the world will not simply wait. Developers will build with what they can touch. Companies will deploy what they can afford. Governments will choose what they can control. Defenders will use what they are allowed to use, or what they can find elsewhere. Toolchains will form wherever friction is lowest.

That does not mean openness is automatically wise.

It means gatekeeping is not automatically strategy.

America needs Civics, not only Ferraris. It needs station wagons, not only supercars. It needs roads, not only palaces. It needs defensive acceleration, not only restriction. It needs to understand that the future of AI leadership may be decided not only by who has the most powerful model, but by whose systems become ordinary enough to build on and safe enough to live with.

The country that owns the flashiest model owns the headline.

The country that owns the substrate owns the habits.

And the country that hardens the world around intelligence owns the safer future.

America cannot afford to confuse the three.

- Iarmhar

June 9, 2026

This essay is part of the Compute, Geopolitics, and Civilizational Strategy Cluster