Special K

During the pandemic recovery, commentators reached for the letter K to describe an economy moving in two directions at once. One arm of the K rises while the other falls. The label was meant to capture a moment, but it kept describing things long after the moment passed, which is usually a sign that a metaphor has caught something real.

Start with the part of the picture that is not in dispute, because it is measured directly rather than inferred. At the end of 2025, the Federal Reserve's distributional financial accounts put the share of household net worth held by the wealthiest tenth of Americans at about 68 percent. The top one percent held close to 32 percent. The bottom half of all households held 2.5 percent. In financial markets the concentration is sharper still. The top tenth of households owned roughly 87 percent of all corporate equities and mutual fund shares, and the top one percent alone owned about half. The bottom half of households owned around one percent of the stock market.

Layer one more concentration on top of those. The market that the wealthy own is itself unusually concentrated. By the late summer of 2025, the seven largest American technology companies, the group investors call the Magnificent Seven, made up a record 34 to 35 percent of the entire S&P 500 by market value, up from about 12 percent a decade earlier. So the gains of the past few years have flowed to asset owners; ownership is concentrated at the top; and within that ownership, an index that millions hold through their retirement accounts now rides on seven firms. Each layer compounds the one beneath it.

It pays to be careful with the evidence here, because the K-shaped story is now repeated so confidently that its softer joints are easy to miss. The most quoted statistic of the past year is a good example. Moody's Analytics, working from Federal Reserve flow data, estimated that the top tenth of earners accounted for roughly 49 percent of all consumer spending by the middle of 2025, the highest share in a series running back to 1989. It is a startling number, and it has been everywhere.

It also rests on a method that infers spending as a residual rather than measuring it directly. In March 2026 the Federal Reserve Bank of Minneapolis reviewed the available data and concluded that they “do not align to tell a clear, K-shaped story.” The Bureau of Labor Statistics consumer expenditure survey, which asks households directly, puts the top tenth's share of spending closer to 23 percent and shows little of the post-pandemic surge. The New York Fed's household panel finds spending growth since 2020 ranging only from about 29 percent at the bottom to 36 percent at the top, a far gentler spread than the headline implies. The wealth concentration is bedrock. The spending divergence is contested, and a practitioner who leans on the dramatic version of it will eventually be corrected by someone who has read the footnotes.

The labor market tells a cleaner version of the divided story. Through 2025 and into 2026 the United States settled into what Federal Reserve officials began calling a low-hire, low-fire economy. Monthly job gains averaged around 26,000 from January 2025 into the spring of 2026, the slowest pace of hiring in more than a decade, while layoffs also stayed low. Anyone who already held a job tended to keep it, while anyone looking for one found the door nearly shut. Long-term unemployment, the count of people out of work for more than half a year, climbed to roughly 1.9 million by December 2025 and made up close to a quarter of all unemployed. Most striking, recent college graduates carried an unemployment rate above the national average through early 2026, by the New York Fed's measure around 5.7 percent. That is a reversal of a relationship that held for decades, and it points directly at the next force in the story.

The divergence did not begin with artificial intelligence, and it helps to be precise about its causes, because each one implies a different response and several of them are weaker than the conventional account suggests.

The most powerful driver is the simplest. For most of the period since the 2008 crisis, interest rates sat near zero and central banks held vast portfolios of bonds. Low rates lift the value of almost everything that throws off future income, from equities to housing, and those assets are owned overwhelmingly by people who already have them. A Federal Reserve study found that around 80 percent of the wealth Americans accumulated in the pandemic rebound came not from saving but from the revaluation of assets they already held. Academic work tracing four decades of falling rates reaches the same place. When the price of capital drops, the people who own capital pull away. This is the part of the K that monetary policy built, and it is the part least responsive to anything an investment agency can do.

Post-pandemic inflation added a second layer, falling hardest on lower-income households whose budgets are dominated by food, energy and rent. The qualification worth making is that wages at the bottom of the distribution also grew unusually fast in 2021 and 2022, which offset some of the squeeze. The net effect was uneven rather than uniformly regressive, and the Federal Reserve's own regional research has spent two years refining exactly how uneven.

A third cause runs through firms rather than households. Across advanced economies the share of national income going to labor has drifted down for a generation, and the work of David Autor and his colleagues attributes much of that to the rise of what they call superstar firms, highly productive companies that capture large market shares while employing relatively few people for the revenue they generate. Industry concentration has risen and the rate at which new firms are born has fallen. There is a vigorous argument about magnitude here that a serious reader should know about. The widely cited finding that corporate markups rose from 18 percent above cost in 1980 to about 67 percent by the 2010s has been challenged by economists who show that counting overhead differently shrinks the increase substantially. The direction is broadly accepted. The size is not settled.

A fourth cause is the one most likely to be quoted wrongly in this field. For years the standard advice for anyone left behind by the economy was education, on the theory that technology rewards skill and the college wage premium would keep climbing. That premium did climb steeply through the 1980s and 1990s. It then plateaued around 2000 and has barely moved since. Recent work from the San Francisco and Cleveland Federal Reserve banks suggests the demand for college-educated labor stopped accelerating two decades ago. The implication is uncomfortable for development practitioners who still reach reflexively for skills programs as the answer to divergence. Education remains valuable, but it is no longer the rising tide it was, and it does not by itself explain the recent split.

The familiar remainder fills out the picture. The China trade shock of the 2000s eliminated on the order of two million American jobs by the central estimates, and the damage to the regions that lost factories proved unusually persistent, though whether the national net effect was a loss or merely a reshuffle is still argued. Housing costs have driven a record wealth gap between owners and renters. None of these causes acts alone, and the most important point for an agency is that the strongest of them, asset revaluation and housing, sit almost entirely outside the toolkit of investment attraction and trade promotion. The forces widening the K are mostly upstream of the people asked to narrow it.

If the K were only the residue of past decisions, it might slowly correct. Artificial intelligence matters because it pushes in the same direction as the existing divergence rather than against it, and it does so through three channels at once: the people it affects, the firms that adopt it, and the places where its capital lands.

Start with people, and with a caution. The International Monetary Fund estimated in January 2024 that about 60 percent of jobs in advanced economies are exposed to AI, with roughly half of those likely to benefit and half at risk of reduced demand. Exposure is not loss. The figure describes contact, not casualties, and it is routinely misquoted as a forecast of unemployment. What gives it teeth is where the early evidence is pointing. A 2025 study from Stanford's Digital Economy Lab, using payroll records from the largest American payroll processor, found that workers aged 22 to 25 in the occupations most exposed to AI saw their employment fall about 13 percent relative to less-exposed peers from late 2022 onward, while older workers in the very same occupations held steady. The authors call these workers canaries in the coal mine and are careful to label the finding early rather than conclusive. It is a working paper, not settled science. But it lines up with the New York Fed's graduate unemployment data and with corporate behavior: Salesforce told investors it had cut roughly 4,000 customer-support roles in 2025 as its AI agents took over half of routine conversations.

Set those facts beside the structure of a career and the K-shaped implication becomes clear. The roles most exposed to early automation are the junior ones, the first-year analyst's work and the entry-level support desk. Those are precisely the rungs that people without capital or connections have always used to climb. A technology that removes the bottom of the ladder while leaving the top intact does more than redistribute work; it narrows the route upward. That is how an efficiency tool becomes an inequality engine, and it deserves more attention than the louder debate about whether AI will erase whole professions.

The second channel runs through firms, and here the divergence is already measurable. The Census Bureau's business survey found in 2026 that 37 percent of firms with at least 250 employees were using AI, against under 20 percent of the smallest firms, with adoption rising among large firms and flat among tiny ones. This rhymes with a pattern the OECD documented well before AI, which it summarized as the best versus the rest: a widening productivity gap between frontier firms and everyone else, driven less by invention than by the failure of new methods to diffuse. AI threatens to deepen that gap, because the firms best placed to extract value from it are the ones that already have the data, the engineers, and the cash. There is a genuine counter-current worth noting. One widely circulated MIT study claimed that 95 percent of corporate AI pilots showed no measurable effect on profit, and although its method has been fairly criticized, the gap it points to is real: adopting the technology and capturing value from it are different achievements, and the second favors firms with capital and talent to spare.

The third channel is capital, and its concentration in 2025 was extraordinary. The Harvard economist Jason Furman calculated that investment tied to AI, chiefly data centers and the equipment inside them, amounted to about 4 percent of American GDP but accounted for roughly 92 percent of the country's GDP growth in the first half of the year. He added his own caveat, that lower interest rates would have generated some of that growth elsewhere, so the counterfactual is softer than the headline. Still, the scale is hard to overstate. The largest cloud companies spent on the order of 370 billion dollars on this infrastructure in 2025 and signaled well over 600 billion for 2026. And that capital does not spread evenly across the map. Brookings found that two metropolitan areas, San Francisco and San Jose, lead the country on every measure of AI activity, with the Bay Area alone accounting for 13 percent of all AI-related job postings.

It is worth keeping a skeptic in the room. Daron Acemoglu of MIT, in work now published rather than merely circulated, estimates that AI will raise total factor productivity by only about half a percent over a decade and lift GDP by perhaps one percent, gains he calls nontrivial but modest, and concentrated. He may prove closer to the mark than the boosters. But notice that even the modest case is a concentrated case. Whether AI transforms the economy or merely nudges it, the distribution of what it produces tilts toward those who already own the assets and the small number of places where its industry clusters. The slope is the problem, more than the speed.

Notice what connects the two ends of this. The same technology thinning out the junior workforce is the one pulling the era's largest flows of capital into facilities that run with almost no one inside them. The economic signature of AI is output uncoupled from employment, and that uncoupling is about to break something the development profession has leaned on for a hundred years.

Here is the part that rarely gets said from a conference stage. A K-shaped economy is not something that simply happens to an investment-promotion agency. The central tool of the trade, public money offered to attract mobile private capital, is itself a mechanism for moving wealth toward the upper arm of the K. Every agency that competes on incentives is, in the aggregate, helping to run a transfer from a broad tax base to a narrow set of capital owners. For most of the profession's history that was easy to miss, because the transfer bought something visible in return.

It bought fewer jobs than the brochures claim even in the best case. The economist Timothy Bartik, reviewing thirty studies of business incentives, found that for at least three-quarters of subsidized firms, and by the cleaner estimates closer to nine in ten, the company would have made the same location decision with no incentive at all. The public money did not tip the choice. It lowered the price the firm paid to do what it intended to do anyway.

What made the gift defensible was a quiet bargain underneath it. Capital needed workers, and it needed them in a particular place, so the subsidy bought local jobs even when it changed no decision. A factory had to be staffed and a back office had to be filled, so the transfer to capital came bundled with a transfer to the community, paid in wages. Artificial intelligence is pulling those two apart. The data center that anchors the modern investment announcement needs the site's electricity and its tax abatement, and it needs construction crews for a couple of years, but it does not need the town's people once the doors open. The jobs that used to ride along with the capital are no longer in the box.

This is a reversal of fortune for the people doing the attracting, and it has not been widely absorbed. In the old arrangement an agency held a real card, because the firm could not build its plant without a local workforce, and that gave the community something to trade. The terms have flipped. The capital-intensive projects that now dominate the pipeline can place their hardware almost anywhere with cheap power and a generous tax code, so the agency's residents are no longer part of what the investor is buying. What is left to compete on is the size of the subsidy.

The scoreboard makes this worse rather than better. Agencies are judged on capital announced and jobs pledged, and those two numbers now pull in opposite directions. The deal that maximizes the headline investment figure is very often the one that minimizes permanent employment, because the way a project gets capital-intensive is by doing without workers. A measure that rewards the biggest capital number is therefore a measure that quietly selects for the projects that widen the K. Fixing it is not a matter of adding an equity column to the annual report. It is a matter of removing a selection pressure that shapes which deals get chased in the first place.

None of this means incentives never work or that every large project is a poor one. Some incentives do tip marginal decisions, and a semiconductor fabrication plant employs thousands of well-paid people for decades, which is a different animal from a data center. A single agency also cannot stop competing on its own while its neighbors keep offering, which is the familiar trap of a race no one can leave alone. The argument is narrower and harder to dodge. As the capital on offer detaches from employment, the share of the profession's work that amounts to a pure transfer is rising, and the tools were not designed for an economy in which that is true.

If prosperity now travels through the ownership of capital rather than the wages it pays, the implication for economic development is blunt and largely unspoken. The way for a community to ride the rising arm of the K is to own a piece of the capital, not merely to host it. A jurisdiction that hands a developer tens of millions and asks for nothing on the other side of the ledger has chosen to fund the upper arm of the K while standing on the lower one. The same money, structured as equity or a share of the revenue, would put the community on the side the economy is now rewarding.

This is not a thought experiment. In 2025 the United States government did exactly this at the national level. It converted roughly nine billion dollars of semiconductor grants into a stake of about ten percent in Intel, becoming one of the company's largest shareholders, and the Defense Department took a fifteen percent position in the rare-earth producer MP Materials. Whatever one makes of the politics, the principle is now established in the open. When public money de-risks private capital, the public can hold equity instead of writing a check and walking away. A state or city signing a seventy-seven-million-dollar abatement can ask the question Washington asked: what do we own when this is done?

The objections are real and worth stating. Governments are poor stock-pickers, equity carries risk that a tax abatement does not, and most agencies have neither the mandate nor the capacity to run an investment fund. So the realistic version is not that every agency should become a sovereign wealth fund. It is a discipline applied to the deals already being done: stop transferring public capital to private capital without taking a position in return. Claw-backs tied to real employment, revenue-sharing on public land, equity warrants in exchange for large abatements, an ownership stake in the infrastructure the subsidy builds. The unifying rule is that the public should stop being the only party at the table that gives without getting.

The same logic reaches trade and export promotion, where the divergence has its own shape. Exporting has always belonged to a small group of firms. In the United States around 2000, the companies selling five or more products into five or more markets were roughly an eighth of all exporters yet accounted for 92 percent of exports. AI widens the gap further, because a frontier firm can now reach a foreign customer directly and cheaply, while the small supplier risks being reduced to a price-taker on someone else's platform. For a Trade Commissioner this argues for a shift in emphasis from offense to defense. The marginal win is no longer one more small firm nudged into its first export sale. It is keeping capable domestic firms from being disintermediated and absorbed by the frontier, and helping them hold the ground they already occupy in foreign markets.

Spreading capability still matters, but as the second move rather than the headline. The frontier firm will adopt the new tools on its own. The other four-fifths of firms will not, and closing that gap is the kind of work that bends the slope of the curve rather than the height of a single announcement. The OECD's long study of the distance between leading and lagging firms concluded that the gap is mostly a failure of diffusion rather than of invention, which is a problem public institutions are actually built to solve.

A final discipline is to hold the trend apart from the noise. Some of the AI figures repeated with great confidence this year will not survive better data, and the careful skeptics may turn out to be right that the productivity gains are modest. A serious agency keeps the part that is not in doubt, the decades of widening wealth and asset concentration sitting in the Federal Reserve's own accounts, separate from the monthly churn of contested claims, and acts on the first without being whipped around by the second.

The economic development profession was built on a quiet assumption, that capital, once attracted to a place, would pay the place back in wages and jobs. That was the transmission belt that turned investment into broad prosperity, and it is the belt that artificial intelligence is now fraying. When capital can produce output with a fraction of the labor, hosting it no longer guarantees a community much beyond a tax bill and a strained power grid.

What follows is not despair, and it is not only the familiar call for more training and better metrics, useful as those are. It is a harder question about position. If the returns of this economy flow to the owners of capital and the few places where it gathers, then a profession whose central act is to give capital away, for free, to firms that would mostly have come regardless, is optimizing for the wrong arm of the K. The daring move, and perhaps the only one that meets the size of the problem, is to change which side of the ledger the public sits on. The agencies that work this out first will look, in hindsight, as though they saw the shape of the thing while everyone else was still cutting ribbons.