Economists arguing about AI and growth disagree by close to an order of magnitude, from a fraction of a percent of added productivity to something near transformative. Much of that gap comes from something other than the technology itself: how much of the capability actually gets deployed, into which parts of the economy, and how fast. The careful work on the question models the firm and the task. Task-based accounting, the productivity J-curve, and the reading of real interest rates each discipline the debate. They do not yet price a second path that runs through the state. Whether frontier capability reaches energy, biology, public infrastructure, and the security systems on which the rest of the economy depends will depend heavily on the arrangement between the government and the labs that hold it. That arrangement is the input most visibly still being decided. It is the swing variable, and this essay sets it on the table for the models that will have to price it.

The argument about AI and economic growth has become unusually good, and anything added to it has to be measured against what is already defined.

Daron Acemoglu’s task-based accounting gives the discipline a near-term ceiling under current evidence: the share of work AI touches, times the saving on each task, bounds the aggregate gain. On current evidence the number is modest, no more than about two-thirds of a percent of total factor productivity over a decade. Erik Brynjolfsson’s productivity J-curve explains why even real gains can arrive late, because measured output waits on the slow and costly reorganization that lets a firm actually use a general-purpose tool. The real-rate reading set out by Basil Halperin and his coauthors comes at the question from the other end: if markets expected growth to accelerate the way the most dramatic forecasts imply, long-term real interest rates should already reflect that expectation, and they do not. All three come at the question with a different instrument, the structure of work, the productivity statistics, the price of money, and each is more disciplined than the forecasts it was built to check.

And yet the estimates do not agree. Acemoglu’s modest decade sits close to an order of magnitude below the scenarios Anton Korinek and others have mapped for a genuine transition to advanced AI, which run to growth rates well above anything in the modern record. The people at both ends are careful, using the field’s best tools, and they still land far apart.

The gap will not close on its own, because it comes from outside the models. The estimates can rest on roughly the same path of capability and still land far apart, because what divides them is deployment: how much of the capability reaches use, into which parts of the economy, and how fast. The models bound the ceiling well. They say far less about the path to it, and the path runs through the conditions around the technology.

The growth debate, for all its range, converges on deployment. From Acemoglu’s low end to the scenarios Korinek has mapped, every estimate treats AI as something that diffuses through firms and tasks: a model is adopted, work is reorganized around it, and the productivity shows up, or it does not. Tyler Cowen’s version of the point is that intelligence is not the only constraint; the humans and institutions around it become more binding as capability improves, and Brynjolfsson’s J-curve is the same claim in the productivity data, where the gains wait on organizational change rather than on the model. On that frame the open question is how fast firms can remake themselves around the tool.

There is a second path the technology has to travel, and it does not run through the firm. The largest deployments, the ones that would move a national growth rate rather than a company’s margin, reach into energy, security, biology, and public infrastructure. Those do not arrive because a business redesigned a workflow. They arrive, if they arrive, through an arrangement between the government and the small number of institutions that hold the AI frontier: who gets access to the most capable systems, on what terms, under what constraints, and toward which ends. The state cannot simply purchase these deployments into existence, the way it could once use demand to shape warships or early chips, because the capability sits in private hands, and what the state holds over it is influence rather than control. Earlier essays in this series called the thing being negotiated a founding compact. This channel is slower than adoption inside firms, because its terms must be negotiated between parties who do not answer to each other. Its ceiling is higher, because it opens the sectors large enough to move a national growth rate. And it is barely built. It is also a case of what Cowen calls state capacity, the pairing of free markets with a state able to build and coordinate, in the form that capacity has to take when the frontier sits inside institutions the state does not own.

The arrangement is a different kind of input from the others feeding that order-of-magnitude gap. Task structure is close to fixed; it is a fact about the work, and it moves slowly. Task structure is relatively slow-moving; it is a fact about the work as currently organized. Organizational reinvention is slow too, but it has precedents the models can use. The arrangement between the state and the producers is neither fixed nor on a schedule. It is being decided now, by specific people in specific rooms, and it could go several ways, from one that moves frontier capability into public use at scale to one that constrains it for years.

That is what makes it the swing variable. Growth at the high end of the estimate range cannot come out of marginal sectors; rates above anything in the modern record require gains in the economy’s largest ones, energy, biology, public infrastructure, and the security systems on which the rest of the economy depends. Those are also the most regulated sectors in the economy, and the ones where deployment runs through the arrangement. The one input still undecided sits exactly where the biggest numbers would have to come from, and the growth estimates do not price it. Yet.

Once the arrangement is in view, several of the field’s open puzzles read differently.

Halperin’s rate signal is the clearest case. On the standard reading, flat real rates say markets doubt transformative AI is close, since those expectations should already show up in long rates, and they do not. But one number can hold more than one belief. The same flat rate is what you would see if markets expected the capability and doubted it would diffuse, slowed by the long work of building everything around it. It is also what you would see if markets expected the capability, expected that it could diffuse, and were betting that governments will not let a technology this consequential run unchecked. The first reading judges the technology; the political reading judges the terms under which the technology will be allowed to diffuse; the rate cannot tell them apart. That is the part the standard reading skips: a number taken as a verdict on whether the technology is coming may be, as much as anything, a verdict on whether the state will let it through.

Acemoglu’s argument leaves a similar institutional question open. He holds not only that the measured gains are modest, but that the larger ones depend on steering AI toward work that complements people, and that the steering is not happening. That is a claim about direction, and it leaves unanswered questions: who steers, through what institution, and why the steering is absent. Those are questions about an arrangement, and his case for steering depends on one existing to steer with.

Brynjolfsson’s J-curve extends the same way. It says output waits until the complementary capital is built, and at the level of the firm he has measured what that capital is: the reorganization, the new processes, the intangible investment. The deployments with the highest ceiling need complementary capital of a different kind, built at the level of the state rather than the firm: the access rules, security frameworks, evaluation capacity, and procurement paths that let a frontier system into public infrastructure. Little of that capital currently exists. His curve reaches into territory the firm-level data cannot see.

This essay does not put a number on the arrangement, because the number is the modelers' to produce: how many tenths of a percent it adds or subtracts, and under which path. Naming the institution the number depends on is the strategist's share of the work, and it is the kind of problem I work on. Acemoglu and Lensman have modeled the regulation of transformative technologies, so governance already has a place inside economics. The claim here is narrower: the arrangement between Washington and the frontier labs does not yet enter the models that produce the growth estimates.

So, the question to leave open is the one the models are built to answer, once the object is in front of them. How much does the arrangement between the state and the frontier move the growth path, and in which direction?

The honest answer today is that no one knows, in part because the arrangement is still being written. The economists are arguing about how big and how fast. A decisive part of that answer is an institution still taking shape, or stalling, while the argument runs. It is the one input still most open to choice, which is what makes it the swing variable, and the one the models will want as soon as someone has set it on the page.

—Andrea