an awkronos model · published apr 8, 2026

When output outruns labor.

American unemployment is rising. Corporate profit margins are at a record. The math that makes both true at once.

0%unemployment rateBLS, Mar 2026
0%S&P 500 profit marginFactSet, Q4 2025
0%new-grad unemploymentNY Fed, Feb 2026
$700BAI capex, five firms (2026)MUFG, Feb 2026
Fig. 01
S&P 500 (teal, left axis) and unemployment rate (grey, right axis), quarterly 2020 through March 2026. Both climbing at once is new.

For most of the last century, the stock market and the job market moved together. Since 2023 they don’t. This piece is the math that explains why, the dashboard that lets you test it, and the limits of what it can tell you.

A note on what the proofs prove. Every number on this page traces to a theorem in an open Lean 4 repository with zero unchecked axioms. Those theorems verify the algebra inside the model. They do notverify that the calibration is right, that the AI-to-labor-share mapping is the right one, or that this is the right macro structure for the current moment. The receipts prove the internal logic. They don’t prove the inputs. This is a small structural model with transparent assumptions, not a forecast.
scenarios at a glance · +36 months

Three guesses about what AI does next.

metric at end of 2028if AI plateaus todayif AI grows linearlyif AI keeps doubling
AI task horizon what a job takes a human vs what AI can finish4 hrs~1 day~3 months
GDP, cumulative change model output at t=36mo from Cobb-Douglas+13.0%+22.8%+30.9%
Labor share of income fraction of output going to workers60% → 58%60% → 43%60% → 30%
Median consumption (welfare) consumption-equivalent change for a median worker+9.2%−5.8%−21.0%
…with non-substitutable labor floor (25% of jobs) GDP gain capped by L_non factor (Economy/JobSwapping.lean)+9.6%+16.6%+22.4%
…with average recession risk GDP minus expected Poisson drag (Economy/RecessionShock.lean)+11.7%+21.4%+29.5%
…with capex bubble pop at t=18 months gK collapses to maintenance (Economy/CapexBubble.lean)+6.1%+15.3%+22.9%
Model output at +36 months, under three growth shapes. Active column reflects METR’s current ~3-month doubling pace (±1.5× uncertainty). Drag the sliders in the dashboard below to test your own numbers.
scene one · the formula

One equation runs this whole story.

Economists have a recipe for how a country makes stuff. It fits on one line, it’s almost a hundred years old, and once you see it the paradox collapses into arithmetic.

Y = A · K0.4 · L0.6

Y is output. A is productivity (how good we are at our jobs). K is capital (the machines and server farms). L is labor (workers). The 0.6 next to L means workers contribute about 60 percent of what the country produces. The 0.4 says capital and productivity do the other 40.

The sneaky part: the exponents are less than 1, so the formula is concave in each input. Halve the workforce and output only drops by about a third. Cut it to a tenth and output is still a quarter of what it was. Try the knobs below.

scene two · drag the knobs

Feel the shape of the economy.

Pull labor down slowly. Watch what output does. This is the same formula above, live. Cut workers in half and output barely budges. Cut them to a tenth and output is still a quarter of where it started. The line is curved, not straight.

1.00
1.00
1.00
output Y
1.00
Y = 1.00 · 1.00^0.4 · 1.00^0.6 = 1.00
how Y depends on labor

That curve is the whole story in miniature. Half the workers, two thirds the output. A tenth of the workers, a quarter of the output. The economy holds itself together, mostly, as long as some workers are still in it. Then, one day, the last of them walks off, and the whole thing goes at once. Hold that shape in your head. You’re going to need it in a minute.

scene three · ghost gdp

Plug in 2026.

+0.8%productivity (A)BLS private nonfarm TFP, 2025 annual
+5%capital (K)BEA fixed assets, hyperscaler capex driven
flatlabor (L)BLS total nonfarm payrolls, 12-mo change
Y_growth = 0.8% + 0.4 · 5% + 0.6 · 0% = 2.8%

GDP up about 2.8 percent with flat hiring. Finance calls this Ghost GDP: growth on paper without anyone on the ground feeling richer. No magic, just the formula.

scene four · the capital

Where the K is coming from.

Fig. 02
Contribution to Q4 2025 real GDP growth, in percentage points. Personal consumption expenditures are still the dominant driver. AI capex is inside the investment column — visible, but not larger than consumer spending. Sources: BEA NIPA Table 1.1.2 (advance estimate), MUFG Hyperscaler Capex Tracker (Feb 2026).

Five companies (Microsoft, Google, Amazon, Meta, Oracle) are pouring over $700 billion into servers, GPUs, and datacenters this year. That’s 2.5 percent of US GDP, and it’s a larger share of capital formation than any single industry has claimed in the post-war data.

scene five · the productivity

Where the A is coming from.

20197 seconds
202315 minutes
2026~12 hours
Fig. 03
Task horizon: how long a human job needs to be before the best AI can’t finish it. Log scale. Grey dashed = plateau at today. Grey solid = linear growth. Teal = exponential doubling. Drag the slider to bend the teal line.
METR’s measured value: 4 months. Source: METR TH1.1, Mar 2026.

METR measures one honest number: how long a task has to be before the best AI chokes. In 2019 that was seven seconds. This year it’s about twelve hours, with about 1.5x uncertainty either direction from modeling choices. The spread between the three curves above is the honest width of what we don’t know about the next three years.

scene six · the dashboard

Run the model yourself.

Six sliders. Three charts. Three preset buttons. Drag anything. The charts redraw live. Teal dots are proved values; lines between are the formula interpolating. Presets snap to a canonical calibration. The growth-shape toggle picks whether AI plateaus, grows linearly, or keeps doubling.

growth shape
labor’s share of national income today (the α in Cobb-Douglas)
how long the simplest tasks AI can do today take a human
task length above which AI exposure stops growing
marginal propensity to consume
capital concentration among the top decile
METR doubling time. 4 months is the measured value today.
?
This shows AI task-horizon over time on a log scale. The teal line is the active scenario, controlled by the AI doubling-time slider. The grey lines are pinned baseline and pessimistic scenarios for comparison. Solid teal dots are values that can be checked exactly against the underlying math when the doubling time matches a kernel-exact value.
?
This shows the share of output going to workers over time. It declines as AI exposure rises and capital takes a larger fraction. Driven by the worker-share slider and the doubling-speed slider together.
?
This compares aggregate GDP growth to consumption-equivalent welfare for a representative worker. The gap is the model’s central observation: GDP can rise while welfare falls if the labor share collapses fast enough.

Every dot on these charts is computed from a formula in an open repository. Hover any number in the post above to see the file and line. press ? for keyboard shortcuts

Keyboard shortcuts

  • 1 Today’s data
  • 2 If AI slows down
  • 3 If labor share drops more
  • r reset
  • ? toggle this panel
  • Esc close this panel
scene eight · entry-level work

What the math says about entry-level work.

Here’s the uncomfortable part. In the same AI-exposed jobs, the younger workers are losing their footing and the older ones are gaining it. That isn’t a story about AI replacing labor in general. It’s a story about AI replacing the bottom rung of the ladder. The internships. The first jobs. The “we’ll train you up” roles that used to turn twenty-two-year-olds into thirty-year-olds with a career.

Run the machine forward three and a half years and here’s what you get. By the end of 2029:

  • The best AI can reliably do tasks that would take a human five hundred and twelve hours. About three months of a full-time job, done in an afternoon. Today’s number is four hours.
  • The share of jobs AI has any opinion about stops growing. It maxes out somewhere around 2028, not because AI gets worse, but because it has already reached every corner it can reach.
  • The slice of the pie going to workers keeps shrinking, because capital and productivity keep growing faster than wages can catch up.
  • Profit margins keep going up, because the same output takes fewer people.
today, early 2026
~12hours
→
end of 2029, METR pace
~18weeks
which is
~4months
of full-time human work, with ±1.5x modeling uncertainty either way
Task horizon at 36 months under METR’s measured 4-month doubling. Kernel: pipeline_metr_horizon_36mo proves intelligenceLevel 36 (log 2 / metrFastSlope) = 512 exactly.

This isn’t a prediction of an unemployment number. It’s a prediction of a shape. The first-job market doesn’t recover by itself, on this trajectory, unless something in the formula changes. The next section is the list of things that could.

scene nine · the breaks

What could break the model.

The model is simple on purpose. Simple things are easier to audit, and easier to break. The five cracks below aren’t hand-waves: three of them are kernel-verified theorems with concrete numbers, two are honest unknowns. Any one would turn most of this page into wallpaper. Click to open each one.

1. AI slows downThe four-month pace is recent. It might not stick.

Nothing doubles forever. If the pace slips from four months back to seven, the whole model slows down with it, and we get a few more years to figure things out. You can watch this happen: drag the doubling-time slider to the right and watch the teal line in the dashboard bend downward. That version of the future is slower and kinder. Or it stops doubling at all, and grows linearly. Or it plateaus today. Try the growth-shape toggle in the dashboard above to see what each of those looks like.

2. The government does somethingA basic income, an AI tax, rules about what workers get paid.

None of that is in the model. If Congress passes a basic income, or taxes AI-driven work, or sets a floor on how much of the economy has to go to workers, the lines move. A labor-share floor in particular would directly break the trajectory chart above. That’s the whole point of one: to stop that line from falling.

3. Capex bubble popsSeven hundred billion dollars isn’t all paid for yet.

The five companies spending all that money aren’t quite earning enough to cover it. They’re borrowing roughly $400B against $1.7T in market cap. The kernel says: if debt service ever exceeds cash-flow growth, the capital growth rate drops from about 6 percent a year to about 1.5 percent — maintenance only. Plug that into the model and the 36-month GDP gain falls from +30.9% to +22.9%, an eight-point haircut. The theorem is CapexBubble.bubble_pop_ghost_gdp_loss: the loss equals (1−α)·(g_hyper − g_maint), no multiplier or feedback assumed.

4. Some jobs don’t swapChildcare. Plumbing. Nursing. A good boss.

The formula quietly assumes a worker and a machine are swappable. They aren’t. Childcare is 1.5% of US payroll, nursing is 2.0%, skilled trades 4.6%, personal care 2.8% — about 25% of jobs sit in a bucket no GPU can substitute for. The kernel models this as a Cobb-Douglas nest over (capital + AI) and the non-substitutable bucket: output bottoms out at L_non^(1−α), which AI literally cannot grow. With α = 0.6 and a 25% floor, the 36-month GDP gain falls from +30.9% to +22.4%. The theorem is JobSwapping.jobSwap_cobbDouglas_zero_Lnon_limit: as L_non goes to zero, output goes to zero, regardless of how much capital or AI you stack.

5. Recession hitsSahm indicator is at 0.20 and rising. Not triggered yet.

The smooth-growth assumption is a useful fiction most of the time and a terrible one during a downturn. NBER says the post-1950 US has had a recession about once every five and a half years, lasting roughly ten months on average, with a peak-to-trough GDP hit around 2.5%. The kernel encodes this as a Poisson hazard of 0.015 per month against a per-month shock of 0.0025. Expected drag over 36 months: 1.35% off cumulative GDP, dropping the headline +30.9% to +29.5%. Sahm’s rule (3-month-average unemployment up 0.5pp from its trailing 12-month minimum) is the early-warning bell. As of March 2026 the FRED real-time reading is 0.20, up from near zero a year ago but still below the 0.50 trigger. It hasn’t rung. It’s rising. Worth watching. The theorem is RecessionShock.recession_expected_loss.

So here’s the thing.

The stock market and the job market did something they’re not supposed to do. A ninety-year-old formula explains it. The formula has three ingredients: productivity, capital, and labor. Two of them are growing fast. The one called labor isn’t. The math says output keeps climbing. It doesn’t say anything about who gets paid.

Every number on this page traces to a theorem in an open repository. 202 theorems, zero sorries, zero unchecked axioms. The model isn’t a forecast. It’s a lens. It shows you why the thing you’re seeing isn’t a contradiction, what the three most honest guesses about the next three years look like, and where the cracks are when those guesses go wrong.

You can argue with the calibration. You can change the sliders. You can toggle the growth shape. What you can’t do is look at 4.3% unemployment, 13.3% record profit margins, and $700 billion in AI capex from five firms at the same time and call it a paradox. It isn’t one. It’s the formula, doing exactly what the formula does.

The math fits the world you’re living in. Believe your eyes.

A thing this absurd has a number. A thing you’ve always wanted to make probably has a number too.