Building Out of Coercion

Lincoln signed an industrial policy in 1862. We need one for the AI economy now.

In my first post here, I laid out the case that supply coercion will be the central economic vulnerability of the next decade. The follow-up question I’ve been getting from audiences and reporters since is the right one. What do you actually do about it?

The honest answer is that you can’t negotiate or sanction your way out. Both responses come after the leverage has already been applied. The only durable answer is to need less of what an adversary controls, which means producing more of it at home, or close enough to home that nobody can credibly squeeze us.

Which makes this a workforce question, well before it’s a trade question.

The chip fabs we want to onshore don’t run themselves. None of what we’re trying to build domestically does, whether it’s rare earth processing, advanced manufacturing, or the data center build-out for AI. All of it requires people who can design and operate the equipment, and we don’t have enough of those people. We’re not training anywhere near enough of them, either. Community colleges are thin in the places we’d need them deepest. Four-year programs produce graduates in fields that don’t match what we’re trying to build. The visa system trains the engineers we want, then sends them back to Taiwan, India, and Korea.

What’s worth pausing on is the scale of what’s actually happening on the capital side. Bridgewater estimates AI capex is adding roughly 140 basis points to US GDP growth this year. Renaissance Macro Research found that in recent quarters, AI investment contributed more to growth than all consumer spending combined. The hyperscalers, meaning Microsoft, Google, Meta, Amazon, and Oracle, have collectively committed something on the order of half a trillion dollars for 2026 alone. Data center construction is running at record levels, and the electric grid build-out behind it is the largest expansion of utility capital plans in decades.

That’s real, and on its own terms, it’s good news. We’re putting the physical infrastructure of the next economy in place. The chips, the racks, the cooling, the gigawatts of power.

But here’s the catch. The infrastructure isn’t the source of the productivity gain. Productivity shows up only when industry actually uses the technology, and “uses” means workers who can take an AI capability off the shelf and translate it into operating decisions inside a real business. A data center running at sixty percent utilization improves nothing. Pharma companies are buying AI platforms right now that they can’t integrate into their discovery workflows because the people who would do that integration don’t yet exist in their organizations. Banks are signing enterprise AI contracts and finding that without redesigning their credit processes, the contracts cost more than they save. The capex builds capacity. People convert capacity into productivity.

And the capex itself doesn’t create that workforce. Data centers, once built, employ surprisingly few people. The jobs that actually move productivity are the deployment jobs in every other industry, the workers who use new technology in the course of doing something else. We don’t train them in college today. The Land Grant institutions could.

We still have time to build that workforce. We don’t have a lot of time. The capacity going into the ground in 2026 will come online in 2028 or 2029. The question is whether we’ll have anyone ready to use it.

Meanwhile, the fiscal picture isn’t helping. Net interest on the federal debt is now larger than the defense budget. It’s the fastest-growing line in the federal accounts. The stock of debt keeps rising, the rates we roll over at are well above the lows of the last decade, and the trajectory looks worse before it looks better. Every dollar that goes to servicing debt is a dollar that doesn’t build anything.

This is what crowding out actually looks like. Budget arithmetic, not theory.

We’ve solved this kind of problem before. In 1862, in the middle of the Civil War, Lincoln signed the Morrill Act. Federal land was sold to endow colleges in every state, where they would teach agriculture, mechanical arts, and military tactics. The Hatch Act of 1887 added experiment stations to the same institutions. The Smith-Lever Act in 1914 established cooperative extension, a service that took research out of the university and brought it to farms and small towns where it was needed.

What looks in hindsight like education policy was actually industrial policy executed through institutions. It built workforce capacity, an applied research base, and a mechanism for getting both out to where the work was actually being done. In every state. The institutions it created compounded for more than a century.

Last fall, Mark Hagerott at North Dakota State, Ramayya Krishnan at Carnegie Mellon, and I wrote a piece for Fortune proposing a Digital-AI Land-Grant Act. Same institutional logic, different century. The argument was straightforward. Take the model that built the agricultural and industrial workforce of nineteenth- and twentieth-century America and apply it to the technology that will define the twenty-first. The piece is in the Fortune archive. I won’t recapitulate it here.

This is what addressing supply coercion at its root looks like. We get less coercible when we can produce more of what we need ourselves. Tariffs don’t substitute for that. Sanctions don’t substitute for that. The capacity has to actually exist.

There’s a second problem the Act would also address, and it’s the one I keep getting asked about in the Q&A after every talk.

The labor market for recent college graduates is in trouble. Unemployment for graduates aged 22 to 27 is at 5.7 percent. The overall rate is 4.2. Underemployment, meaning the share working in jobs that don’t require a degree, is 41.5 percent. Oxford Economics estimates that recent graduates, who make up about five percent of the workforce, have accounted for roughly twelve percent of the rise in unemployment since 2023. That’s not how it usually goes. Recent-graduate unemployment normally runs below the overall rate, not above it.

The anxiety I hear from audiences is some version of the same question. AI is eating the bottom rung of the career ladder. The work that junior associates used to do at law firms. The work that young analysts used to do at banks. Entry-level work used to be how people learned to do the work above it. If AI is doing that work now, where do the senior partners and managing directors of the next generation come from?

It’s a fair question, and the answer isn’t to throw our hands up. The answer is to raise the bottom rung. AI now does the work that used to be the first step on the ladder, so the work of the next decade is making sure graduates have the skills to step onto what’s become the new first rung, the work that uses AI rather than being replaced by it.

This is exactly the kind of skill formation that the Land-Grant institutions were built to do, and they’re already in every state and region. A Digital-AI Land Grant Act would put that work at the center of the next decade. Without something like it, those unemployment numbers among recent graduates are a leading indicator of a generation that gets credentialed but can’t find the rung.

There’s an obvious objection. We can’t afford it. The debt’s already too high. The deficit’s already too large.

I’ve heard that objection inside the Federal Reserve. I’ve made it myself. It’s not wrong as far as it goes. What it misses is that our budget process treats consumption spending and capacity-building investment as if they were the same thing. They aren’t. Borrowing to consume is one kind of decision. Borrowing to build institutions that will compound for a century is a fundamentally different kind. The Morrill Act, for what it’s worth, was financed largely with federal land that wasn’t being used productively. The institutions it endowed paid that investment back many, many times over.

I’m not arguing that we should ignore the deficit. I’m arguing that the budget process needs to distinguish between spending that buys us another year and spending that builds us another century. If we can’t make that distinction, every dollar looks the same on the page, and the dollars that build capacity keep losing to the dollars that buy time.

There’s a harder version of the same objection, and I hear it from people whose judgment I respect. They look at the same numbers I do and conclude that the fiscal trajectory is past the point where any of this matters. The math, in their reading, has already been decided. I understand how they got there. I don’t agree.

Here’s what I’d say to them. The belief itself becomes a constraint. If enough of us conclude that we can’t build, we don’t build, and the capacity doesn’t get built. That’s a supply problem too. It just doesn’t come from outside.

Supply coercion occurs when the things that hold us up are controlled by someone who can choose to let go. Usually, that someone is a foreign government, a cartel, or the bond market. But sometimes it’s us.

We can’t retaliate our way out of that. We have to build out of it. We’ve done it before. The template’s right there. And the political coalition for workforce, infrastructure, and capacity building is broader than most Washington debates would suggest.

What we’re missing is the discipline to treat capacity-building as what it actually is.

Load-bearing.