There Is No Late Stage

TL;DR

“Late-stage stagnation” is a frame, not a finding. Underneath it sit three separately measured things: a productivity slowdown from roughly 3% a year in 1920–70¹ to about 1.5% from 2004 to 2022,² a hollowing of middle-skill employment from 75% to 67% of US jobs between 1980 and 2009,³ and a decades-long fall in the startup rate that halved the share of employment at young firms.⁴ All three are real. But since 2023 productivity has run at about 2.7%,² new business applications hit 5.4 million — the most since the series began in 2004,² and post-pandemic wage compression reversed something between a quarter and a third of forty years of accumulated wage inequality.⁵⁶ A cycle theory requires the parts to move together. They stopped.

A derelict factory — the image every stagnation argument reaches for, which is part of why the argument is so rarely checked. Photo: Ввласенко, CC BY-SA 3.0.⁷

The Frame Arrives Before the Evidence

Ask what “late-stage” economic stagnation is and you will find a very confident answer waiting, usually with numbered phases. An economy builds real wealth through industry and infrastructure. It financialises. It extracts rent from what it built. It stagnates and decays. The stages are named, the current position is identified — always somewhere in the last third — and the conclusion follows automatically.

The thing to notice is that this structure is doing work before any data arrives. Thomas Cole painted it in 1836, in five canvases, with the same rock visible in every frame while the civilisation around it rises and falls.⁸ The narrative is old, satisfying, and independent of whichever economy it’s currently applied to.

That doesn’t make it wrong. It does mean the honest procedure is to take the frame apart and check whether the pieces behave the way a stage theory needs them to. There are three real pieces, and they are worth taking seriously, because each is genuinely well evidenced.

Piece One: The Productivity Slowdown

Robert Gordon’s case is the strongest version of the pessimistic argument, and it deserves the respect it gets. His claim is that 1870–1970 was a “special century” — a period unprecedented in its breadth of change and unlikely to repeat.¹ Output per hour grew about 1.5% annually before 1920, nearly doubled to roughly 3% between 1920 and 1970, then fell back to 1.62% after 1970.¹

His explanation is not decadence. It is arithmetic. Electricity, the internal combustion engine, indoor plumbing and the elevator each transformed daily life once, and cannot do it again — factories are already electrified, houses already plumbed. On his account, “Nothing that happened after 1940” matched them.¹ He adds a structural point that gets less attention: information technology is only about 7% of GDP, which caps how much of the economy it can transform.¹

The subsequent numbers held up. US productivity grew 2.9% a year from 1994 to 2004, then 1.5% from 2004 to 2022 — the long slowdown, right on schedule.²

Piece Two: The Hollowed Middle

The second piece is the one people actually feel, and the shape is not in dispute. Between 1980 and 2009, mid-skill jobs fell from 75% to 67% of US employment. Machine operators went from 10% of the workforce to 4%. Administrative support went from 18% to 14%. Over the same period high-skill work rose from 12% to 15% and low-skill from 13% to 17%.³

That is the diamond becoming an hourglass, measured. And David Autor’s later work sharpens why it hurt in a specific way: the transitions were not symmetric. College-educated workers moved in both directions across the wage distribution; non-college workers moved essentially one way, out of specialised middle-skill occupations and into low-wage work requiring only generic skills.⁹ The urban wage premium that once rewarded a non-college worker for moving to a big city was substantially attenuated.⁹ Cities became more skill-intensive while the people in them without degrees did less skilled work than before.

Piece Three: The One That Kept Going Longest

The third piece is the least discussed and, until recently, the most persistent. The rate at which new firms are founded in the US declined steadily from 1981 to 2013, and the share of employment at young firms fell from 20% to 10%.⁴

Why that matters is not sentimental. A large share of productivity growth comes from reallocation — resources moving from less productive firms to more productive ones. And the link between a firm’s productivity and its employment growth has been weakening over time, while the gap between the most and least productive firms has widened, particularly in the information sector.⁴ Productive firms are less reliably absorbing the economy’s labour. That is a mechanism, not a mood, and it plausibly holds down measured productivity growth.

Here is where the decline story usually reaches for its villain: monopoly. Big incumbents accumulate market power, block entry, and the economy calcifies. It’s a tidy account and it may yet be right in parts, but the cleanest test of it fails. Federal Reserve researchers compared industries against each other and found that industries with larger markup increases saw smaller declines in dynamism — the opposite of the prediction. Their conclusion is blunt: rising markups “are not a major explanatory factor behind declining dynamism.”¹⁰

So the most popular mechanism for the most persistent symptom doesn’t survive a cross-industry check. Hold that thought.

Then Three Things Happened

If stagnation were a stage, the pieces should deepen together. Instead:

Productivity accelerated. From 2023 the US ran at roughly 2.7% annual productivity growth — back in the range of the 1990s boom and nearly double the preceding two decades.²

New business formation surged. Business applications reached 5.4 million in 2023, more than any year since the series began in 2004, with particularly strong entry in high-tech sectors.² The Aspen Institute’s analysts identify that surge as a major driver of the productivity rebound² — which is to say the third piece, the supposedly structural decline in dynamism, reversed hard and appears to have pulled the first piece with it.

Wages compressed from the bottom. Autor, Dube and McGrew found that post-pandemic labour-market tightness produced rapid relative wage growth at the bottom of the distribution, reducing the college wage premium and counteracting a substantial share of forty years of accumulated inequality. Their own figure moves between versions of the paper — around one-third in the NBER abstract, approximately one-quarter in the summary published by MIT’s Blueprint Labs — so call it a quarter to a third, and note the honest imprecision.⁵⁶ The mechanism they identify is competition: the pandemic raised the elasticity of labour supply to firms in the low-wage market, reducing employer market power, with gains concentrated among young non-college workers who changed employers.⁵⁶

Read that last one against Piece Two. The group the hollowing-out story identified as the losers — young workers without degrees — are the group that gained most, through the mechanism the story said had been foreclosed.

What the Stage Frame Gets Wrong

Not the symptoms. The grammar.

A stage theory makes a strong claim: that these phenomena are manifestations of one underlying process at one point in its arc. That claim generates a testable prediction — the indicators should move together, and in one direction. Over 2023–2025 they did not. One pillar reversed sharply, another pulled up behind it, and the distributional damage the theory treats as terminal partially unwound.

The alternative reading is less dramatic and fits better: these are three separable problems with separable causes, which happened to point the same way for about three decades. The productivity slowdown is substantially about which technologies were available to deploy. The hollowed middle is about automation and trade acting on specific routine task bundles. The dynamism decline is about firm entry and reallocation, and we demonstrably do not know its main cause, since the leading candidate just failed its test.

Three problems with three causes can improve independently. A civilisation in its late stage cannot. The frames are not equivalent, and only one of them survives 2023.

There is also a quieter geographic point. The stage narrative is indexed almost entirely to countries that completed industrialisation and then watched their manufacturing middle contract — it is a rich-country story about a rich-country sequence. Most of the world’s population is at a different point in that transition entirely, which should at minimum make one wary of a theory of the economy derived from the last fifty years of a handful of them.

Where I’d Hold This Loosely

Three real caveats, and the first is substantial.

Two or three good years are not a trend, and the people producing these numbers say so first. The Aspen analysis is explicit that it is unclear whether the recent stretch signals a longer-term shift, and notes that generative AI cannot yet be the explanation, since firm adoption was around 4% at the time of writing.² The pandemic was an enormous, unrepeatable shock to labour markets; a reversal driven by it may reverse again. If productivity settles back to 1.5% by 2030, the pessimists were right and I was reading noise.

Second, “no stage theory” is not “no problem.” Everything in Piece Two happened, and the person who lost a middle-skill job in 1998 is not consoled by wage compression in 2022. Levels and trends are different questions — a stable or improving trend can still describe a bad level, which is a distinction this site keeps running into.

Third, I could not access the analyses arguing that occupational polarisation specifically stopped in the 2000s, so I have not leaned on them; the reversal evidence here is about wages, which I could read, not about occupational shares, which I could not. If polarisation did continue quietly through the 2010s, the picture is muddier than I’ve drawn it.

What survives all three caveats is narrow and, I think, worth holding onto. The confident numbered-stage account of where an economy sits in its life cycle is not a finding about economies. It is a shape imported from somewhere older — Cole had it in 1836, and he wasn’t doing economics either. The measured components are real, separable, and have already demonstrated they can move apart. Whatever is wrong with these economies, it isn’t a stage, because a stage is something you cannot reverse without going backwards through it. In 2023 the dynamism indicator went up. Nothing in the theory permits that.

---

Footnotes

1. https://knowledge.wharton.upenn.edu/article/dazzling-yet-disappointing-why-u-s-growth-productivity-will-underperform-the-past/ ↩ ↩² ↩³ ↩⁴ ↩⁵

2. https://www.economicstrategygroup.org/publication/in-brief-us-labor-productivity/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://libertystreeteconomics.newyorkfed.org/2011/11/job-polarization-in-the-united-states-a-widening-gap-and-shrinking-middle/ ↩ ↩²

4. https://www.brookings.edu/articles/declining-business-dynamism-implications-for-productivity/ ↩ ↩² ↩³

5. https://www.nber.org/papers/w31010 ↩ ↩² ↩³

6. https://blueprintlabs.mit.edu/research/the-unexpected-compression-competition-at-work-in-the-low-wage-labor-market/ ↩ ↩² ↩³

7. https://commons.wikimedia.org/wiki/File:Abandoned_factory._Kiev%2C_Ukraine._2008.jpg ↩

8. https://en.wikipedia.org/wiki/The_Course_of_Empire_(paintings) ↩

9. https://www.nber.org/papers/w25588 ↩ ↩²

10. https://www.federalreserve.gov/econres/notes/feds-notes/rising-markups-and-declining-business-dynamism-evidence-from-the-industry-cross-section-20240308.html ↩