The gen-AI productivity curve does not turn positive until roughly 64 percent of the economy's value added is included.
That is the single number that should end the debate the Fed thinks it is refereeing. It comes from the Kansas City Fed, which ranked detailed industries from lowest to highest productivity contribution and cumulated them across value-added share. For most of the distribution, the curve sits below zero. Then it climbs almost vertically at the right tail.
In plain terms: two-thirds of the American economy is contributing nothing net to the post-2022 productivity pickup. A few industries at the top are doing all the work.
Two-thirds of the American economy is contributing nothing net to the post-2022 productivity pickup. A few industries at the top are doing all the work.
This is the fact that breaks Kevin Warsh's frame. At the last press conference, the incoming Fed chair described the economy as being in "a race" between efficiency gains and business spending—both driven by AI—whose outcome would determine inflation and, by extension, rates. The Fed built a "productivity and jobs" task force to referee it.
A race implies an aggregate crossing a finish line the officials can watch. Efficiency pulls inflation down; capex pushes prices of land and laptops up; you clock the winner and set policy accordingly.
But there is no aggregate to clock. There is a peloton standing still and three riders sprinting off the front.
Start with magnitude, because the KC Fed is not arguing the boom is fake. The gen-AI-era endpoint is about 2.5 percentage points annualized, more than double the 1.2-point pre-pandemic figure. The pickup is real and it is large.
It is just extraordinarily narrow. The top four contributors—retail trade, information, professional/scientific/technical services, and real estate—nearly doubled their contributions, while wholesale trade flipped from positive to negative. Within "information," the gains aren't even sector-wide: they concentrate in data processing and hosting, not broadcasting or telecom.
So the disinflationary dividend, if it exists, is being generated inside a handful of slices of a handful of industries. That is not how a broad 1990s-style productivity wave lowers the price level across the basket a central bank targets.
Now go up the size distribution, where the concentration gets worse. An NBER paper by Boppart, Klenow, Laski, and Li backs out rents from ideas using price-earnings ratios and finds that the largest listed firms contribute less to productivity growth than their sales share.
Read that against the capex story. The mega-caps are precisely the firms spending hand-over-fist on data-center equipment—the ones setting the capital-demand side of Warsh's race. Their footprint in output is large; their marginal contribution to measured efficiency is smaller than that footprint implies.
The efficiency side of the ledger is thin where the spending side is thick. That is not a race between two comparable forces. It is a large, concentrated demand for capital running against a diffuse, hard-to-measure efficiency gain.
Then close the borders. Eichengreen and coauthors, in NBER work linking AI to geoeconomic fragmentation, ask how fragmentation affects the international diffusion of AI—the magnitude of the gains and their distribution across economies. Fragmentation throttles diffusion.
The historical GPT template—which the Bank of Canada leans on—assumes broad spread. Electrification took roughly 40 years to show up in US manufacturing productivity; ICT took about a decade to appear in industry statistics. Diffusion is the whole mechanism by which a concentrated invention becomes a broad-based boom.
If fragmentation slows that spread across borders while the KC and NBER data show it is already narrow within them, the "boom" stays trapped: in three industries, in a few firms, inside national lines that are closing. A gain that never diffuses never becomes the economy-wide disinflation the Greenspan story requires.
Here is where the mechanism flips against the administration's own analogy. The claim, per the AP, is that Warsh and the Treasury can coax a replay of 1999—productivity up, unemployment down, inflation contained, and therefore rates lower.
The 1990s cut the opposite way. Real rates rose during the last productivity boom, and they rose from a starting point meaningfully higher than today's real yields. Nominal rates across the curve were substantially higher then than now—even though inflation was slower and fiscal policy tighter.
The reason is the part of the story the "immaculate disinflation" narrative omits. A genuine technology boom raises the prospective return on investment, which raises the demand for capital, which pulls the neutral rate up. The Greenspan-as-dove morality tale gets the physics backwards.
So even if you grant the optimists their productivity acceleration—an investor cited by The Overshoot puts market-implied economy-wide acceleration at about 1.5 points a year for 5-10 years—simply re-running the playbook implies a big upward shift in rates from current levels, not cuts. Real rates might need to rise by more than any drop in inflation compensation, driven by capital demand rather than efficiency.
And the concentration makes this sharper, not softer. When the productivity gains are diffuse, they can plausibly slow prices across the whole basket. When they are jammed into a few sectors while capex demand is broad and physical, the capital-demand channel dominates the inflation channel. The neutral rate gets pulled up faster than any efficiency dividend can pull inflation down.
This is the medium-term destination the near-term policy debate is frozen relative to. Whatever the FOMC does at 3.5-3.75% in the coming meetings, the point here is that the r-star it is anchoring against may be higher than its own task force wants to concede—because the boom it is being asked to accommodate loads the capital-demand side and starves the efficiency side.
Which leaves the binding constraint, and it is not monetary. It is physical.
The capex engine—the demand side of Warsh's race—runs on data centers, and Americans have decided they do not want them. A Gallup poll released in May found 71% oppose data center construction in their area. That is higher than the 63% zenith of anti-nuclear-plant opposition in 2001.
The collapse is fast. An August 2025 Heatmap survey had support and opposition roughly even at 43-42; a month later, support cratered to 21% while opposition hit 71%, with 55% strongly opposed.
And it is bipartisan. New York put a moratorium on data-center construction in July. Texas—the release valve, the red state that was supposed to absorb the industry the blue states banned—followed this month.
That is the detail that matters. The comfortable equilibrium was blue states banning physical industry and red states catching it. When Texas pauses, the release valve closes, and the capex buildout faces a NIMBY ceiling that no interest-rate path can lift.
The scarce input, as Noah Smith puts it, is energy—and the fight over data centers is really a fight over who gets it. A buildout that can't physically clear can't deliver the boom, whatever the market has priced.
So the Fed is caught between two things it cannot resolve with a policy rate. On the efficiency side, a productivity gain so concentrated—64% of value added contributing nothing net, the biggest firms underdelivering, diffusion blocked at the border—that there is no aggregate signal to read.
On the capex side, a capital-demand surge large enough to lift the neutral rate, running into a physical constraint that may stall it before it clears.
The honest reading is that the "race" is the wrong metaphor and the 1990s is the wrong comfort. Both point the same direction. The gains are too narrow to disinflate the whole economy, and the capital demand behind them—if it clears at all—pushes real rates up.
The Fed will be asked to price a boom it cannot measure against a buildout that may not physically happen. The Greenspan playbook, run honestly, ends in hikes.