Treasury Secretary Scott Bessent and Federal Reserve Chairman Kevin Warsh—the two leaders of the American economy—say the AI productivity boom will soon make America richer, so much so that it will be deflationary, and rich enough that the country can stop worrying about its $40 trillion debt.

But analysts are beginning to ask: richer for whom? Workers' share of U.S. income has already fallen to its lowest level on record, even as corporate profit margins set new records quarter after quarter.

According to Gregory Daco, chief economist at EY-Parthenon, the productivity gains that explain that divergence largely predate the AI boom. "Productivity growth protects margins, not income," Daco wrote in a note Thursday.

Economic output grew 1.7% in the second quarter on just 0.3% more hours worked. Compensation rose 2.6%, which, measured against a spring and summer of oil-driven inflation, amounts to "flat to slight contraction" in real terms, Daco told Fortune in an interview.

Profit margins reached a record 14.9% of GDP, while the labor share fell to 52.8%—the lowest since the government began tracking the figure in 1947. The labor share measures how much of national income goes to workers as wages and benefits rather than to owners of capital as profits and investment returns, which is why the split matters for how the gains of any boom are distributed. Daco said 50% is not a floor. "As long as you continue to see concentrated gains on the capital side, and within a certain number of firms," labor's share could keep plummeting, he said.

The productivity behind those numbers stems from a decade of conventional automation, cost discipline (hiring pulled back after post-pandemic bloat), and capital spending—not from AI. What AI has delivered so far is further concentration.

"You tend to have greater concentration and more of a winner-takes-all type of environment when you have these technological advances," Daco said. In nearly every technological revolution—the late-19th-century railroad boom or the 1990s dot-com revolution—large, vertically integrated firms initially capture the gains while smaller firms face "persistent cost pressures, persistent policy uncertainty, higher interest rates," Daco noted.

In the 1990s, a handful of companies at the technological frontier front-loaded capital investment and reaped the capital gains, but productivity growth from cheaper software spread quickly throughout the economy, and wage growth followed. There is no guarantee AI will follow the same timetable.

The AI boom is, after all, uniquely and historically capital intensive. Data center investment is expected to reach $31 trillion by 2050—nearly the size of current U.S. GDP—according to analyst firm PricewaterhouseCoopers LLP. While the engine sputters in other sectors, construction and manufacturing are roaring because of data centers; without them, the industry would be in recession, a Chicago manager said in the Federal Reserve's Beige Book this week.

That looks like growth: companies spending hundreds of billions on equipment that should eventually let the economy produce far more with less. But much of that equipment is not made in America.

Imports of the large computers used in AI servers have exploded over the past year. Net imports of "large computers"—the Census category covering GPU servers—hit a $450 billion annualized pace last month, a startling rise from roughly $50 billion a year through 2023, according to Census data compiled by economist Joseph Politano. GDP accounting treats an imported server as adding to investment and subtracting as an import in equal measure, so the net contribution to GDP is zero.

That helps explain the shape of the AI economy so far. Capital spending is booming, productivity is improving, corporate margins are enormous, and conditions are "loose," as Warsh points out—yet hiring is weak, housing is struggling under tight rates, and workers' share of income keeps shrinking.

"While U.S. investment is booming, growth in gross domestic product has been modest," wrote Jon Hilsenrath, the former Wall Street Journal Fed reporter who now advises hedge funds at Serpa Pinto Advisory.

That poses a hard question for Warsh and Bessent: let it rip, or do something about it?

Growth is not the same as broadly distributed income. If every dollar of output increasingly accrues to the owner of a data center—passively collecting checks—or to a shareholder who owns the data center operator, the fiscal math gets complicated. The economy may be getting richer while the tax base and political constituency policymakers usually associate with a boom grow far more slowly. It is not hard to imagine that fanning suspicion of the AI buildout and its benefits.

The investment itself also carries costs and risks. Hundreds of billions of dollars in AI spending compete for capital in an economy where borrowing is getting more expensive. Higher long-term rates make mortgages costly and suppress homebuilding, as the Wall Street Journal demonstrated in a striking chart this week.

None of this means the AI productivity boom will fail; it simply means it is not immediately obvious how it raises labor's share of income. "I don't think there's a floor," Daco said.

This story was originally featured on Fortune.com.