The Question Before the Crisis: AI and the Next American Order

The most revealing move at any AI and national security gathering right now is the reach for a historical analogy. In fireside chats and sidebar conversations, senior lab leadership and government officials are asking the same question: what is this like? Skunk Works? Bell Labs? The canals and railways of the industrial age? The Manhattan Project? The question is shorthand for something larger: how to engineer a generational bet, a strategic undertaking that coordinates public and private power at a scale neither can reach alone, and that transforms national power if it succeeds.

Senior leaders are asking because they need to move quickly. China is already running an integrated model for AI: hardware, software, education, research labs, and contracts coordinated as a single stack. The United States has not organized itself that way in living memory.

The search for a technological analogy is not about catching up with China. It is about finding a model that would let the United States set the agenda and shape the system others will have to operate inside. The search is failing because the most consequential seam in AI is not between the United States and China. It is between the U.S. government and the frontier labs — two actors who have mistaken proximity for alignment and are losing ground on the assumptions that make the next generational bet possible.

The debate over technology analogies is stuck for a structural reason. Each analogy on offer is one-and-done. Bell Labs was a thirty-year arrangement under monopoly conditions that no longer apply. Skunk Works solved a specific Cold War problem. The canals and railways were a single infrastructure buildout. None is portable to the next forty years of AI, because the next forty years will not be governed by a single arrangement. They will be a succession of arrangements; strategic operating systems that hold under specific conditions, then give way to what comes next.

The real question is not which analogy is correct. It is what assumptions underpin the current system, and whether the United States directs the coming succession or merely undergoes it.

The end of Pax Britannica shows us what it costs to undergo a succession rather than direct it. The signals were not missed. They were misread; interpreted in the vocabulary of the order that was ending rather than the one that was forming. British tariff reform treated lost competitiveness as a problem to be tariffed away behind imperial preference, rather than as the sign that industrial primacy was passing to firms rebuilding on the new fields such as electricity and chemicals, where American and German industry had pulled ahead. British naval expansion treated the German challenge as a naval problem to out-build rather than a question of global ambition requiring Britain to commit to new diplomatic alliances. At the highest levels of strategy, the succession stalled. The war came, and the next order was dictated by a different power.

Underpinning any strategic order are assumptions about the environment and about other actors. The system runs on those assumptions. When they shift, inertia keeps the system moving long after it has stopped producing the outcomes it was designed for. The signals appear before failure. They are misread because they are interpreted through the language of the past.

The current debate over AI and national security cannot define the next strategic order because it cannot see the assumptions that underpin the present one. The usual framings each miss something: counting inputs like semiconductor capacity and PhDs; reducing AI to a capability that can be ordered or bought; treating it as an oracle to be consulted but not understood. What none of those framings reveals is that "AI" is not one thing. Pull it apart and five elements sit inside it:

·      Capability — what the model does

·      Apparatus — the compute, data, and infrastructure that make it possible

·      Makers — the institutions producing it, their governance and constraints

·      Ecology — the talent, culture, and tacit knowledge behind frontier work

·      Operations — how and where it is deployed

Capability, apparatus, makers, ecology, operations — CAMEO. The procurement system collapses all five into capability and then treats capability as a thing: something specifiable, contractible, deliverable. From inside that language the collapse is invisible. The three signals that follow show it from outside.

Signal one. Frontier AI labs are declining to participate in standard DoD procurement vehicles — OTAs, SBIRs, traditional contracts — that would have been obvious wins for defense-tech vendors five years ago. The vehicles work. That is not the issue. They are built for vendors selling into government. Frontier labs are not behaving as vendors. They are acting as independent strategic actors with their own views on what to build, who should have access, and under what conditions. The procurement system has no vocabulary for this.

Signal two. DoD program offices are writing AI requirements in the language of a prior order. Specifications describe model characteristics such as parameter counts, accuracy thresholds, and latency targets, and frame deliverables as if AI were a product. The labs are not engaging. Not because the requirements are too demanding, but because the requirements cannot describe what is being bought. Apparatus and ecology — infrastructure, data systems, talent density, research culture — are conditions of production, not deliverables. They cannot be specified into existence.

Signal three. Senior defense officials describe AI using commodity metaphors: electricity, IT, horizontal capability. Frontier labs use a different vocabulary. Frontier systems do not behave like commodity software. They reshape the institutions that deploy them. Access itself becomes strategic. The difference between frontier and commodity models is not incremental. It is categorical. The two sides are operating with different ontologies in the same conversation.

Each signal looks like a procurement issue in isolation. Together, they indicate something else.

The mismatch is not primarily about policy disagreement.

It is about ontology; what each side believes the object really is.

In irregular warfare, seams are structural gaps that adversaries exploit. The seams here are different. They emerge between actors on the same side who are operating from incompatible assumptions. The U.S. government and frontier AI labs are not adversaries in the way that the United States and China are — their interests are not fundamentally opposed. But they are misaligned at a foundational level, and the seam between them is widening.

This kind of seam is more dangerous because it can rupture before either side recognizes it. Each side reads the breakdown as a failure of execution. The government side sees reluctance, evasion, obstruction. The labs see bureaucratic incompetence. Neither sees the shift in assumptions underneath. The gap widens while both are looking at the wrong thing

If the relationship between frontier labs and the U.S. government is held together by assumptions that are already shifting, then fixing procurement is not the first-order problem. Procurement belongs to the previous order, built for a different object and a different industrial relationship. Patching it improves the machinery of that order. It does not produce the next one.

The current debate about AI is focused on analogies. The analogies are not the issue. The issue is succession and the arrangements that make that possible.

Reading the present system correctly is the precondition for the generational bet itself, for a real chance to shape the next world order rather than have it dictated by another power. Most of that work remains to be done.

— Andrea