Every serious conversation about AI and the state reaches for the same fix — Bell Labs, the Manhattan Project, the railroads — as if naming the analogy hands over the solved problem underneath it. It doesn't. What the cases actually share sits one level down, in the arrangements that connected public purpose to the capability the state never owned itself. No such compact exists yet for AI. It's being assembled right now, by people who haven't agreed they're assembling anything. Each arrangement in the record had someone who noticed early and moved. This one is still waiting on its architect.

Walk into any serious conversation about AI and national security and you will hear the same question asked in different rooms. Is this Bell Labs? The Manhattan Project? Skunk Works? The railroads? The reach for a historical analogy is an attempt to borrow a solved problem. Name what this is like, the thinking goes, and you can borrow the arrangement that made it work.

The instinct is sound. The execution fails, because the analogies quickly become stand-ins for comparisons between technologies, and at that level no two frontier technologies behave alike. They become useful only when read as comparisons between arrangements: the institutional bargains that connected public purpose to capability the state did not itself build.

There is a harder fact under the search. No finished arrangement is waiting to be borrowed. The founding compact between the government and the frontier labs is being written now, by the people in the room, and they are present at the creation, ready or not.

The history of strategic technology is less a history of invention than a history of arrangements. Bell Labs was not the government. Neither were the semiconductor firms, the railway companies, or the yards and arsenals that armed the great powers. Even the Manhattan Project ran on universities and industrial contractors that sat outside the formal machinery of the state. What linked their work to national purpose was never ownership. It was an arrangement that gave public authority a practical way to shape what private institutions produced.

The mechanism changed with the era. Bell Labs did long-horizon research inside the economics of a regulated monopoly; AT&T’s protected position paid for the patience, and regulation set the terms. The early integrated-circuit industry matured on government demand, as defense and space procurement bought down the cost of a technology the commercial market could not yet absorb. The railways became strategic through law rather than engineering, when charters, land grants, and rights-of-way turned private track into national mobility. The Manhattan Project fused objective, funding, secrecy, science, and production under a single wartime authority built to expire.

One case is worth slowing down for, because it describes a condition that looks like the present. Meiji Japan is usually remembered as state-led modernization, which is true and incomplete. The Meiji state had ambition and authority. It did not begin with the industrial, naval, and engineering capability its ambitions required. Its first problem was not production but access. It hired foreign advisors, sent the Iwakura Mission to study foreign institutions, dispatched students abroad, and bought ships from foreign yards. What kept this from being simple dependence was the discipline behind it. Imported knowledge was routed into Japanese schools, ministries, arsenals, and dockyards, so that access became absorption and absorption became domestic capacity. The state stayed the organizing actor while the expertise it needed still lived somewhere else.

Then there is Armstrong. By the end of the nineteenth century, W. G. Armstrong’s Elswick works on the Tyne had become one of the most advanced military-industrial enterprises in the world, designing and building naval guns, ordnance, and complete warships. Britain helped make the firm possible, through naval demand, imperial scale, and the prestige of British power, and the Admiralty was among its most important customers. It was not the only one. When it suited the business, Armstrong sold abroad. In the decades before the Russo-Japanese War, the firm helped modernize the Imperial Japanese Navy with ships built at Elswick and the guns to arm them.

This did not make Armstrong disloyal or Britain weak. Britain remained the dominant naval power, and the Admiralty kept real leverage over the firm. It could place orders, confer prestige, regulate access, and set much of the environment the firm depended on. What it could not do was determine Armstrong’s commercial choices, its foreign clients, or the direction of its technology. The Admiralty had influence. It did not have control.

A state can remain powerful and strategically dominant while the capabilities that define the frontier sit inside institutions it does not own and cannot command. The test is whether authority converts into practical influence over institutions that have additional, and sometimes competing, incentives.

The first essay in this series argued that “AI” is not one object, and that much of the confusion in the American debate comes from arguing about different parts of the system with the same word. The relationship between governments and frontier labs has the same defect. Regulation, procurement, competitiveness, safety, industrial policy, partnership: all of it gets filed under governance, as though it were a single question. It is several questions, and they do not resolve the same way.

Even in an era of AI, none of the old instruments of national power have vanished. Governments still shape energy, capital, export controls, immigration, supply chains, and the terms of deployment. The novelty sits elsewhere. Frontier labs hold more than technical skill. They hold concentrations of talent, research judgment, evaluation capacity, and the accumulated learning of operating at the edge, none of it easily separated from the institution that grew it. Compute can be financed and models can be accessed. The judgment about which directions are promising and which risks are real is harder to buy from outside. For most of modern history the state did not need to build the technology itself. It needed a practical way to shape the institution that did. That is the capacity now in question.

It is tempting to read the relationship as lopsided in the labs’ favor. They hold capabilities the state cannot reproduce and increasingly set the direction of the field, so why spend so much effort on government at all? Because dependence has not disappeared. It has changed shape. Frontier development requires reliable power, large-scale construction, deep capital markets, stable property rights, secure facilities, semiconductor supply, and open markets. None of that is produced by the labs. It is sustained by political and legal order and by state capacity.

This is why the relationship cannot be read through procurement alone. A defense contract matters as revenue, but revenue alone does not explain the depth of engagement. Understanding how a new arrangement could take shape means knowing what is at stake for the labs. The relationship supplies consequential use cases, operational environments, and a seat in the conversations that will set the conditions for frontier work. The state is not only a customer, and the laboratory is not only a vendor. Each holds something the other cannot easily replace. The relationship is mutual dependence expressed through different forms of power.

This is also why no familiar category fits. Contractor is too narrow, partner too vague, regulated entity incomplete, national asset misleading. Governments make law, allocate resources, conduct diplomacy, provide security, and answer for national outcomes. Labs decide what gets built, what gets prioritized, and which risks are acceptable. For most of modern history those activities stayed close together. Now they sit in different institutions, and both have grown more important at once. Governments attend more closely because frontier AI reaches economic competitiveness, military capability, intelligence, and critical infrastructure. Labs matter more because the decisions that shape the technology are made inside a handful of organizations at the edge of the possible. The old arrangements held because the institutions deciding the frontier and the institutions answerable for it stayed connected to each other. The assumptions underneath them may no longer describe the relationship taking shape. What is forming, named plainly, is a founding compact: the terms on which public purpose and frontier capability will be joined in this era. It is being assembled piece by piece, by parties who have not agreed that they are assembling anything.

Read this way, the debate over which analogy fits has been answering the wrong question. Bell Labs, the railroads, the Manhattan Project, Meiji Japan, and Armstrong are not blueprints to copy. They are evidence of one recurring move: successful powers found arrangements that connected frontier capability to strategic purpose without requiring that capability to live inside the state. The mechanism changed every time. The requirement did not.

China makes the stakes concrete. There, hardware, software, research, capital, and state direction are meant to move as a single stack, and the apparent coherence is the point. The American answer cannot be imitation. The American pattern has been mixed. The state built its own arsenals and laboratories when it had to, but its distinctive strength has been the share of the frontier held by institutions it does not own and cannot command. The United States does not lack private frontier capacity. It already has it. The open question is whether that capacity can still be connected to public purpose at the speed and scale the frontier now demands, or whether the connection is being left to form on its own.

That is not a procurement question, nor a regulatory one, nor a question of who is ahead. It is a question of grand strategy.

Every arrangement in the record had architects, people who saw that the old compact had stopped describing the world and set about founding the next.

This one is being assembled now, without agreement that it is being assembled, and no one has taken up that work. Acheson had a name for the people who did: present at the creation.

— Andrea