Who Keeps the Dividend

The productivity shock is not a forecast anymore. It is in the data, and the most important number in it is not the productivity number.

Nonfarm business labor productivity rose 2.2% year over year in the second quarter of 2026. Unit labor costs rose 1.4% over the same four quarters.¹ That is the combination everyone has been waiting for: output per hour rising faster than the cost of producing it.

Then read one line further down the same release.

The labor share — the portion of output that goes to workers as compensation — was 52.9% in Q2. That is the lowest reading in a series that starts in 1947.¹ Real hourly compensation over the last four quarters: down 0.1%. Unit nonlabor payments over the same period: up 9.0%.¹

The dividend is real. It is not being shared.

The mechanism is cost collapse, not headcount

The naive version of the AI productivity story is that companies fire people. That is not what is happening and not what needs to happen for this to matter.

Search, synthesis, drafting, coding, analysis, and coordination all carried a per-unit labor cost. That cost is collapsing. A company that needed 100 knowledge workers to produce $20 million of output does not have to cut 30 of them. If the same 100 produce $24 million, revenue per employee rises while most of the cost structure sits still.

Revenue grows faster than headcount. Operating leverage increases. Margins expand. EPS follows.

That chain is correct as far as it goes, but it just stops one step short of the question that determines who actually gets paid.

Competition is the missing term

The operating leverage argument assumes the firm keeps the surplus. Nothing guarantees that.

If AI diffuses broadly across an information-intensive sector, competitors use the same tools to produce the same output at lower cost, and they bid the savings away. The productivity gain shows up as price deflation in that sector rather than margin at any firm inside it. The customer keeps the dividend. The producer keeps the same margin on a smaller revenue base.

This is not a hypothetical failure mode. It is the ordinary outcome of a cost reduction available to everyone at roughly the same time.

So the screen is not "which companies get the productivity dividend without financing the infrastructure that creates it." The screen is narrower and harder: which companies have enough pricing power to keep a dividend that all of their competitors also received.

Those are different lists, and the second one is much shorter. It favors switching costs, regulatory moats, distribution locks, proprietary data, brand-priced services, and contracted revenue. It disfavors any business where the buyer can run a competitive bid.

Some of the winners are casualties

The list of AI beneficiaries always includes professional services. Half of that list belongs in the other column.

A law firm, an accounting practice, or an agency that bills by the hour and doubles output per hour has just cut its own revenue in half. The productivity gain is a direct attack on the pricing model. The only escape is repackaging output as fixed-fee, subscription, or outcome-priced work — which means renegotiating every client relationship the firm has while its competitors do the same thing.

Hourly billing converts a productivity gain into a revenue loss. That is a mechanical result, not a risk.

The professional services firms that survive this will be the ones that decouple price from time first. The ones that wait will discover their productivity improvement showing up in their competitors' bids.

The bond market is already arguing about this

There is a version of this economy that looks strange on paper: real growth accelerating, inflation contained, the Fed eventually easier, and long yields structurally elevated anyway.

The logic holds. Higher expected trend productivity raises the return on capital, which raises the equilibrium real rate. JPMorgan Private Bank's U.S. head of investment strategy, Jacob Manoukian, made exactly this argument in late August: the rise in long yields may partly reflect the bond market pricing a better productivity cycle rather than only fiscal and inflation anxiety.²

The complication arrives in the same interview. AI-related debt issuance has passed $220 billion this year, double last year's total, against $1.68 trillion in total U.S. corporate issuance, up roughly 27% year over year. Manoukian's own estimate is that AI-related issuance could reach half of U.S. Treasury coupon issuance by year-end.² The 30-year has traded above 5%.³

So the long end is absorbing an enormous supply shock at the same moment it is supposedly pricing a productivity boom. Both stories predict higher yields.

Only one of them is bullish for equities.

Output per decision

The first-order story is someone doing today's job faster. That is the boring part.

The disruptive part is the shift from one person doing one task, to one person supervising three agents, to one person supervising six — with each agent operating recursively across its own subtasks. The relevant economic quantity stops being output per hour and becomes something closer to output per human decision.

That is not a linear improvement. It is a change in what a headcount number means.

And the diffusion is wildly uneven, which is where the dispersion comes from. Census BTOS data put firm-level AI use at 18% during the winter supplement period — but 32% on an employment-weighted basis, because large firms adopted faster.⁵ By May 2026 the national rate was 19.8%, with Information at 39.7% and Finance and Insurance at 33.9%, against roughly 14% in retail trade.⁶ Fed research using a different survey frame estimates that 78% of the labor force works at a firm that has adopted AI, and 54% at a firm using LLMs.⁷

Those three numbers describe the same economy. The gap between them is the whole trade.

Important

An aggregate productivity gain of half a point to a point can coexist with individual firms running 30% or 50% differences in effective labor productivity. The macro series will look benign while the cross-section tears itself apart.

What to watch

Not AI revenue. That measures the sellers.

Watch revenue per employee, SG&A as a percentage of revenue, headcount growth against revenue growth, R&D output per dollar, gross margin trend in sectors where adoption is highest, and — the one most people are skipping — the spread between unit labor costs and unit nonlabor payments. That spread is where you see whether the dividend is being retained or competed away.

At the aggregate level, the labor share is the honest scoreboard. It has been falling to a record low while productivity rises.

• If it keeps falling, capital is keeping the gain.
• If it stabilizes while productivity keeps rising, the dividend is going to workers.
• If productivity rises and sector price indices fall together, it went to customers and nobody on the producing side got anything.

Three different outcomes, three different trades, one set of statistics that will tell you which is happening about two quarters before the earnings calls do.

The market narrative right now is that companies are spending enormous amounts on AI. The narrative that replaces it is that AI changed the production function. That transition matters more for valuations than another generation of better models — and the first place it becomes visible is not in a productivity release.

It is in which companies stop hiring and keep growing.