The first piece in this arc argued that a productivity dividend goes to whoever has the pricing power to keep it.1 The second dropped the same mechanism to the level of a person.2 This one asks what happens to the system if the answer keeps coming back "capital" — and the uncomfortable conclusion is that the productivity boom and a classic credit crisis are not competing forecasts. They can be two sides of the same event, and the boom can be part of the mechanism that produces the collapse.
Start with what productivity does and does not measure.
It measures the efficiency of production. It says nothing about income distribution, debt service, aggregate demand, asset prices, or solvency. A firm that discovers 700 people can produce what 1,000 did will report spectacular productivity.
Whether the economy around it holds together depends entirely on what happens to the other 300 — and on what everyone's liabilities assumed about them.
The slow clock
The standard version of this story has a firing wave. That is not how it arrives. As the second piece argued, firms don't fire the exposed band — they stop backfilling it.2 Junior hiring slows, departures go unreplaced, functions consolidate. Nobody's cash flow breaks on a Tuesday.
That changes the shape of the shock without changing its size. The demand rot arrives through the hiring margin instead of the layoff line:
- The job never posted
- The household never formed
- The first mortgage never taken
- The starter home never bought
It is slower than 2008 and much harder to see, because every individual data point looks like softness rather than crisis.
But the people not hired were going to be more than employees. They were going to be customers, taxpayers, borrowers, renters, and contributors to retirement plans. The economy's liability structure — housing valuations, commercial leases, municipal budgets, consumer credit, equity multiples — capitalizes decades of assumed labor income. Technology can reprice the production function in a product cycle. Liabilities reprice over decades, through defaults, renegotiations, and write-downs.
That gap between the two repricing speeds is where the risk lives.
Closing the escape hatch
There is an apparent way out, and the first piece seems to supply it: if competition transmits the dividend to customers as lower prices, households get the gain back as purchasing power. Demand holds. Everyone is fine.
In a leveraged economy that path is not benign, and Irving Fisher explained why in 1933.3 Liabilities are nominal. The mortgage, the muni bond, and the leveraged loan are all fixed in dollars. Falling prices with fixed-dollar debts raise the real burden of every obligation in the system — the debt-deflation mechanism. Cheaper output makes existing leverage heavier.
Which closes the loop completely.
The capital path starves demand directly. The customer path inflates real debt. The only transmission channel that relieves balance sheets instead of stressing them is wages — and the second-quarter data shows that channel not operating.
Labor share at 52.9%, the lowest since 1947; real hourly compensation slightly negative over four quarters; unit nonlabor payments up 9%.4 The dividend is arriving. It is not arriving as income to the people who owe the money.
The fast clock
A classic crisis needs more than weak demand. It needs leverage, a maturity mismatch, and a forced seller. All three now exist, and they were built to finance the boom itself.
The BIS documents private credit loans to AI-related companies growing from near zero to more than $200 billion, with Morgan Stanley projecting an additional $800 billion of data center financing over the next two years.56 Data center securitization ran near $27 billion in 2025 and is projected to reach $30–40 billion annually in 2026 and 2027 — a meaningful slice of combined asset-backed and commercial-mortgage issuance.7 Total AI-related debt issuance passed $220 billion this year against the broader corporate calendar.812
The collateral is the weak joint.
The chips are depreciated over five to six years on the books while new generations now arrive annually, and the buildings around them carry 20–30 year economic lives with the debt maturing somewhere in between. The BIS's warning is explicit: data centers and GPUs financed today could become obsolete before their debt is repaid.5 And the timing is no longer hypothetical:
Warning
CoreWeave's $7.5 billion GPU-collateralized facility carried a variable rate near 11%, with repayments beginning in January 2026 — just as the collateral's market value was falling sharply.6
The forced seller is not the borrower. It is the vehicle.
A large share of this exposure sits in interval funds, non-traded BDCs, and tender-offer structures — the wealth channel's push into private credit — marketed as "infrastructure debt" or "asset-backed lending" without much clarity about how concentrated the AI exposure underneath actually is. The BIS itself now pairs the two risks in one sentence, flagging AI disruption to lenders' software-heavy books amid redemption pressures from private credit's push toward retail investors.9 Those structures offer periodic liquidity against loans that have none.
A semi-liquid fund holding illiquid paper against fast-depreciating collateral, facing redemption windows, is the structured investment vehicle of 2007 with GPUs where the mortgages used to be.
Contagion runs the familiar route. Banks' direct exposure to AI-adjacent industries averages just 0.8% of assets — but the Chicago Fed cautions that they carry additional exposure through lending to the funds themselves.6 Small direct, large intermediated. That is the 2008 geometry, and the BIS's 2026 Annual Report completes it, naming an AI capex bust and circular financing among its top financial-stability pressure points and warning that collateral in interlocking equity, debt, and supplier deals may be pledged multiple times.10
The slow clock sets off the fast one
Here is what makes this one argument rather than two. The fast clock's revenue assumptions depend on the dividend's incidence — the subject of the first piece.
The infrastructure debt is serviced by compute revenue. Compute revenue comes from firms buying AI capacity. Those firms will keep paying today's rents only if they can monetize the productivity gain — and the competition mechanism says that in most sectors they can't, because the gain gets competed through to their customers.1 Meanwhile, the household side of demand erodes through the hiring margin. If the dividend concentrates into profits and asset prices rather than broad income, the spending that was supposed to validate $3 trillion or more of projected infrastructure investment doesn't materialize on schedule.
The incidence problem and the credit problem are not parallel risks. The first causes the second. The boom financed a bet that its own dividend would be broadly monetized, which makes the boom the fuse.
Why the statistics will look good
The strangest feature of this scenario is that the early readings flatter it. GDP holds. Productivity accelerates. Margins expand. Inflation falls, because production got cheaper. Some equities soar. Underneath, employment breadth narrows, bargaining power erodes, and income concentrates.
This has a name and a decade: the 1920s. Strong productivity growth, expanding margins, a falling labor share, stable prices, and demand increasingly sustained by credit instead of income. The combination read as prosperity the entire way up.
The modern version adds a twist the 1920s lacked at this scale: consumption itself has concentrated. Moody's Analytics estimates the top tenth of households by income at 49.2% of consumer spending — a record in data back to 1989, up from roughly 36% three decades ago.11 The methodology is contested, and the true share may be lower; the direction is not seriously in dispute. Zandi draws the AI connection himself: the boom lifted the stock prices of AI companies, and those companies are owned by precisely the households doing the spending, with the wealth effect contributing roughly a quarter of GDP growth in a recent year.11
Caution
Concentration doesn't just weaken demand slowly — it makes demand hostage to the AI trade. The wealth effect and the credit structure crack together, because they are the same portfolio.
Policy is poorly shaped for this. It would present as a technological boom and a balance-sheet recession at the same time, and the Fed's instrument amplifies the disease: falling inflation gives it maximal room to cut, but cuts subsidize capital deepening, accelerate the automation, and inflate the assets of the dividend's current recipients.
Easing into an incidence problem treats the symptom with the cause. The tool that matches the problem is fiscal, and fiscal is slow. That is the precise sense in which the institutions responsible for distributing the dividend adapt more slowly than the technology creating it.
What to watch
Two clocks, two dashboards.
-
The slow clock is labor: the labor share, the hiring rate, quits, and the gap between wage growth and productivity growth. Those tell you whether the wage channel — the only benign transmission path — turns on.
-
The fast clock is credit: redemption activity in semi-liquid private credit vehicles, GPU lease rates and secondary prices, data center securitization spreads, and covenant amendments in the neocloud complex. Those tell you whether the collateral is being marked before the marketing catches up.
The great economic risk of AI is not that it fails to produce a productivity boom. It is that the boom arrives faster than the institutions that distribute its dividend can adapt — that we build enormous productive abundance on top of deteriorating household solvency and call it growth until the redemption notices go out.
A sufficiently large productivity shock can be the most bullish supply-side development in generations and the catalyst for a classic demand and credit crisis. Not as competing outcomes. As one event, reported by two different dashboards, one of which updates faster than the other.