AI Boom, Bust or Something Else? Citrini’s Future

The Luddites were nineteenth-century English textile workers who attacked the innovative new looms that threatened wage growth and labour protections. Even in relatively recent times, the appeal of machine-breaking has loomed large when ways of living, working and the social fabric itself seem at stake.

This fear of the disruptive goes a way toward explaining why a research note was able to dramatically move IBM's stock price after it was published on 22 February. The stock fell 13% in a single session, its worst day since 2000.

The note, from macro commentator Citrini Research, was written as a fictional memo from two years in the future. The scenario it envisaged was bold. Artificial intelligence (AI) capabilities have improved so rapidly that a graphics processing unit (GPU) cluster — a computing system that can execute thousands of small, repetitive operations in parallel, dramatically speeding up graphics rendering, AI model training, and scientific calculations — in North Dakota is suddenly able to generate the output of 10,000 Midtown Manhattan office workers. Productivity could explode. Margins could surge. And then the economy might fall apart; because those 10,000 displaced workers can no longer pay rent, buy groceries, or send their kids to summer camp. Citrini called it 'Ghost GDP': output that shows up in the national accounts but never circulates through the real economy. At the time of writing, the piece has attracted more than 28 million views on X.

The rebuttals came quickly. The Wall Street Journal's Greg Ip made the historical case bluntly: technology has never caused a job apocalypse in US history, the current data show no sign of one, and the more plausible near-term risk is a capex bust; the sums being poured into data centres already far exceed what those centres are earning, a dynamic that rhymes uncomfortably with 2001.

We’ve been here before

The historical record, however, points to a more uncomfortable precedent. Luddite projections were far from completely wrong. When the first industrial revolution mechanised production in early nineteenth-century England, output per worker rose sharply — but real wages stagnated for roughly six decades before workers began to share in the gains. The economic historian Robert Allen called this the 'Engels’ Pause'. Between 1780 and 1840 alone, output per worker rose roughly 50% while real wages rose just 12% — a divergence visible in Figure 1 that continued to widen until around 1870 when wages began to catch up. The question isn't only whether the economy adapts in the long run, but who bears the cost of getting there.

Figure 1. United Kingdom output per worker (solid) and real wages (dashed), 1760–1913, indexed to 100 at 1820. Seven-year centred moving averages

Problems loading this infographic? - Please click here

Problems loading this infographic? - Please click here

Source: Thomas & Dimsdale (2017), Bank of England Millennium Dataset.

Both sides of the debate are worth reading. But they share a premise that deserves examination: that AI is already transforming the macroeconomy in ways large enough to matter. That's the question we wanted to answer directly.

There's a test for this

In 2021, Nobel laureate William Nordhaus published his work on six empirical tests for whether the US economy is approaching a genuine singularity: a regime of explosive, self-sustaining growth driven by information technology. His framing was careful. He wasn't dismissing the idea; he was building an early warning system for it. He concluded at the time: 'The tentative conclusion is that the Singularity is not near, but we have developed tests that can give early warning signs of its occurrence.'

We ran those tests again with a decade of additional data. We included Fernald's quarterly total factor productivity (TFP) series, an indicator of US business sector productivity, adjusted for cyclical variations in labour intensity and capital utilisation, up to the third quarter of 2025. We also added BLS KLEMS¹ industry productivity data through 2024, extending Nordhaus's original sample by roughly 10 years.

The early warning system is still quiet.

What the data say

The most important test is productivity growth. If AI is driving a genuine transformation, total factor productivity, the economy's best measure of technological progress, stripped of cyclical noise, should be accelerating. It isn't. TFP averaged 0.88% per year over 2015–2025, a recovery from the post-crisis trough of 0.16%, but well below the 1.82% of the late-1990s tech boom, and with no trend acceleration in sight. A singularity doesn't look like a modest bounce from a trough.

The second decisive test is whether the information technology’s (IT's) benefits are spreading through the economy. If the technology were genuinely transforming the macroeconomy, the gains shouldn't be confined to firms building AI; they should show up in law firms, hospitals, logistics companies and retailers using it. We checked 57 non-IT industries using BLS data. Thirty-one of them — 54%, hardly a decisive majority — showed some productivity acceleration since 2000. The median gain was near zero. IT spillovers remain missing in action.

One test does pass: capital's share of income — profits, rents and interest relative to total income — has risen from roughly 31% in the early 1990s to about 40% today. That's consistent with singularity dynamics, but equally consistent with monopoly power, globalisation and three decades of declining union density. Passing a necessary condition isn't the same as making the case.

The remaining three tests sit in ambiguous territory. The most fundamental is the oldest question in economics about technology: as machines get cheaper relative to workers, does capital's share of income rise or fall? If it rises, machines and workers are substitutes: you can keep replacing one with the other, and self-sustaining growth becomes possible. If it falls, they are complements: machines need workers to be useful, and you eventually hit a wall. We tested this directly against 75 years of US data. The answer flips depending on whether capital's rising income share over the past three decades reflects machines displacing workers, or simply rising corporate market power and declining unions. The data cannot distinguish between the two.

The other two — whether capital is accumulating faster than output, and whether IT is becoming more central to production — show some movement in the expected direction but no decisive trend. The data is suggestive in places, but conclusive in none. Part of the shift from Nordhaus's original verdicts reflects measurement choice rather than new data: we use capital services rather than capital stock for the first, which weights short-lived IT assets more heavily; and income weights rather than stock shares for the second.

One pass. Three ambiguous. Two clear fails.

Problems loading this infographic? - Please click here

The six Nordhaus (2021) singularity diagnostic tests. Nordhaus column reflects data available at publication (ending 2014–2018 depending on the test).

The more interesting question

Here is where the Citrini piece deserves more credit than the market reaction it triggered; and where the rebuttals, while probably right in the near term, fall short.

The standard counter to AI doom is the recycling argument: production generates income; productivity gains get recycled into higher wages, lower prices, or new investment; the pie grows. History largely supports this.

But Citrini's scenario is asking a slightly different question, and it's one the standard rebuttal doesn't fully close. Capital's share of income has been rising for thirty years. If AI accelerates that trend — if the gains accrue predominantly to capital owners while labour income stagnates — then an economy that runs 70% on consumer spending faces a distributional problem the aggregate data won't show. The Engels' Pause, after all, was not a failure of aggregate growth. It was a failure of distribution. You can have record output and weak demand simultaneously.

This is distinct from the full doomsday scenario. It doesn't require mass unemployment, a mortgage crisis, or the 38% equity sell-off Citrini envisions. It just requires the productivity gains to be claimed faster than the economy can redistribute them — a slower-moving risk, and arguably a more credible one.

Ip's concern about capex deserves equal attention. The internet was genuinely transformative, and tech workers still lost their jobs in large numbers after 2001 — not because the technology failed, but because capital expenditure had raced ahead of demand. The sums being ploughed into data centres today far exceed what those centres are currently earning — suggesting the immediate risk, as Ip argues, isn't an economic singularity but a classic investment bust. That precedent is relevant and underappreciated in the current boom-or-doom binary.

What we're watching

Aggregate data are a lagging indicator, sometimes by decades. Electrification took 40 years to register in productivity statistics; the personal computer took around 20 years. Our own research has documented firm-level research and development (R&D) productivity at US companies declining at roughly 10% per year since 1975, even as R&D investment grows at 6% annually. If AI is reversing that trend, the signal will appear first in gross profit per R&D dollar, patent yields and revenue per employee — years before Fernald's quarterly TFP or BLS industry data pick it up. Our verdict reflects what the macro data show through the third quarter of 2025 and does not rule out an inflection already under way at the firm level.

The bottom line

Nordhaus built a diagnostic kit for detecting economic singularity. A decade of new data — spanning smartphones, cloud computing and the first wave of large language model deployment — has not changed the reading. Productivity is not accelerating. IT spillovers remain absent across most of the economy. Where the scorecard did shift since 2021, with tests four and five moving from clear verdicts to ambiguous, the change reflects better measurement, not a worsening signal.

The debate Citrini triggered is worth having. But the debate about what happens when AI transforms the macro economy is getting ahead of whether it already has. The most honest position right now sits between the doomsday scenario and the historical-precedent reassurance: an Engels' Pause (real productivity gains, slow to circulate, rewarding capital owners well before ordinary workers) is a more plausible outcome than either extreme. It is no less damaging for ordinary workers, and is the outcome current market pricing seems least prepared for. For investors, an Engels' Pause is most visible on the demand side: businesses whose customers are middle-income wage earners are exposed if productivity gains circulate slowly.

The warning system is quiet. But it is running.

1. Capital K, Labour L, Energy E, Materials M, and purchased services S. KLEMS provides industry-level measures of productivity by decomposing output growth into contributions from each of these input categories, covering a broad set of US industries.