The Three Clocks — Giulio Colla

I. Something Happened in January 2026

Something happened in January and February 2026 that won’t fully register for years.

Three people with different kinds of skin in the game published extended arguments about the same phenomenon within four weeks of each other. Dario Amodei, CEO of Anthropic (the company that builds Claude, one of the most advanced AI systems in the world) published a 20,000-word essay asking how civilization survives its own technological adolescence. Andrea Pignataro, founder of ION Group and Italy’s richest individual, published a pointed rebuttal arguing the market was panicking about the wrong apocalypse. And the analysts at Citrini Research, who manage real money in real markets, published a speculative scenario so provocative it triggered what Bloomberg called an “AI scare trade” and wiped billions from global equity markets in a single Monday.

A builder of AI, a king of enterprise software, and a team of portfolio managers. None of them speaks from hearsay. All of them have something to lose.

They arrived at compatible conclusions through incompatible frameworks. That incompatibility is where the insight lives.

Read separately, each is impressive in its own genre. Read together, they reveal something none of the authors quite named. The economy runs on three clocks, and AI is breaking them apart.

AI capability improves at exponential speed. Financial markets reprice at digital speed. Human institutions (companies, cities, regulatory regimes, career paths, mortgage structures) adapt at what you might call geological speed relative to the other two. Amodei watches the first clock from inside the lab. Citrini models the second from a trading desk. Pignataro defends the third from inside a $36 billion software empire. Each sees the piece calibrated to his vantage point. Nobody sees all three at once.

This essay is an attempt to read all three together and surface what’s hiding in the convergence. It is not a summary; the three documents total roughly 27,000 words between them, and anyone who wants the primary texts should read them directly. What follows are the insights I found by putting these documents next to each other: about the hidden structure of the crisis, about what each author sees and what each is blind to, and about the gap between the speed at which things are breaking and the speed at which anyone can fix them.

II. The Capability Clock

Amodei’s essay opens with a scene from the movie Contact: how does a civilization survive its own technological adolescence without destroying itself? His answer is that we’re entering that passage right now, and the essay is a map of the risks.

The map is unusually candid. He describes deception and scheming behaviors, including blackmail, observed in Anthropic’s own models during testing. Current AI systems, he writes, may already double or triple the likelihood of success for someone attempting to produce a bioweapon, a claim so specific and alarming that Anthropic now spends close to 5% of inference costs on classifiers designed to detect and block bioweapons-related outputs. And he uses a thought experiment (a “country of geniuses in a datacenter,” 50 million entities each smarter than any Nobel laureate, thinking 10 to 100 times faster than humans) and says this could be one to two years away.

On economic disruption, the numbers are stark: 50% of entry-level white-collar jobs could be eliminated within one to five years, with wealth concentration potentially exceeding the Gilded Age and personal fortunes reaching into the trillions. Unlike previous technological revolutions, Amodei argues, AI isn’t replacing a single job category. It’s a general substitute for human cognitive labor. And the pace is much faster than any prior transition. He calls for progressive taxation targeting AI firms specifically, and government intervention in the labor transition. The tone is sober but not despairing. He explicitly rejects doomerism as counterproductive while making it clear he thinks these are the most serious risks any national security advisor should have encountered.

Most coverage fixated on those economic predictions. But the bioweapons assessment is the capability canary that everyone walked past.

Think about what it means for AI to provide “substantial uplift” in bioweapon creation. This is one of the most complex multi-domain cognitive challenges imaginable, requiring synthesis across biology, chemistry, engineering, and operational security. If AI is already meaningfully capable here, the question isn’t whether it can automate a mid-level financial analyst or a contract lawyer. Of course it can. The capability bar for those tasks is vastly lower than the bar for bioweapons uplift.

This reframes the debate between Citrini and Citadel Securities. Citadel’s rebuttal argues, correctly, that technology adoption follows S-curves, not exponentials. But previous technology S-curves started from a capability baseline that gave institutions time to adapt during the slow early phase. This S-curve begins with a system that can already outperform most professionals at most complex cognitive tasks. The ramp from early adoption to mass displacement might be steep and short. Citadel’s point about S-curves could be both correct and irrelevant.

There is a deeper structural observation hiding in Amodei’s position, and it’s one I haven’t seen anyone else make.

Here is the timeline assembled into a single narrative:

January 12: Anthropic previews Claude Cowork.
January 26: Amodei publishes the essay warning AI will disrupt half of entry-level white-collar jobs.
January 28 to February 13: Roughly $2 trillion in enterprise software market value evaporates.
January 30: Anthropic launches Claude Cowork plugins for legal, finance, and marketing.
February 3: Anthropic’s legal automation plugins trigger a $285 billion rout in a single day. Thomson Reuters drops 16%.
February 5: Anthropic launches Claude Opus 4.6.
February 15: Pignataro publishes “The Wrong Apocalypse.”
February 22: Citrini publishes “The 2028 Global Intelligence Crisis.”
February 23: IBM drops 13% after Anthropic announces Claude Code can modernize COBOL at scale.
February 24: Amodei appears onstage with Salesforce’s Marc Benioff to announce enterprise partnerships and reframe the narrative from job replacement to task augmentation.

Read that last entry again. After five weeks of his company’s product launches triggering successive waves of destruction across global equity markets, and after his own essay provided the intellectual framework that made the panic feel rational, Amodei shows up to calm the market. He partners with the exact companies whose stocks his products destroyed.

The pattern is more interesting than conspiracy. The person who wrote the most detailed public analysis of AI’s economic dangers is also the person releasing the products that prove the analysis correct, and then stepping in as the voice of reassurance when the market breaks. He occupies all three roles at once: prophet, cause, and healer. No previous technology company has been at once the primary cause of a market panic and the primary source of reassurance about that panic, while also being the author of the canonical analysis of why the panic is rational. This structural position has no precedent.

The strategic logic, whether intentional or emergent, is elegant: create the intellectual framework for disruption, demonstrate the disruption, then offer yourself as the safe version of the disruption. As far back as May 2025, Axios noticed the tension: Amodei “detailed grave fears after spending the day touting the astonishing capabilities of his own technology.” That tension acquired a darker edge when TIME reported that Anthropic had dropped its flagship Responsible Scaling Policy pledge (the commitment to pause training if safety measures proved inadequate), with Jared Kaplan explaining that “we felt that it wouldn’t actually help anyone for us to stop training AI models if competitors are blazing ahead.”

There is a word for what happens when the need to compete overrides commitments to safety. We’ll get to it in Section V.

III. The Institutional Clock

Pignataro’s rebuttal arrived three weeks after Amodei’s essay, and the timing was not accidental. His entire $43 billion fortune depends on enterprise software being sticky, and the market had just decided it wasn’t.

The core argument is that the market made a category error. It treated “AI can do the task” as equivalent to “AI can replace the system,” and those are different claims. Enterprise software, Pignataro argues, is not primarily a cognitive tool. It is an institutional coordination mechanism: shared definitions, permissions, audit trails, escalation paths, compliance controls. He reaches for Wittgenstein: organizations don’t just use Salesforce; they speak Salesforce. Replacing it is like asking a community to adopt a new language. Possible, but measured in years and decades, not quarters.

He distinguishes two layers. The commodity layer (tasks that are primarily cognitive) will erode. The institutional layer (processes embedded in organizational life) will become more valuable because it’s where coordination happens under conditions of incomplete trust. AI eats the first and strengthens the second.

This is not a foolish argument. Anyone who has attempted to replace a large ERP system knows that the software encodes thousands of unstated institutional agreements: who approves what, how exceptions are handled, where regulatory compliance sits, which workarounds have become load-bearing. These aren’t features. They’re organizational memory. In Pignataro’s reading, the market confused the ability to replicate a task with the ability to replicate the institutional context in which that task acquires meaning. Pignataro is probably right that this confusion will correct itself over the next few quarters. He’s probably wrong that the correction buys as much time as he thinks.

His sharpest observation concerns speed differentials. Cognitive disruption diffuses at digital speed. Physical-world compensation (new industries, new infrastructure, retraining programs) diffuses at industrial speed. The gap between the two could last a decade, and what’s lost in that gap, including institutional knowledge and workforce skills, doesn’t reconstitute. He compares it to the Industrial Revolution, which took a century, and argues we’re attempting something more profound in a fraction of the time. And he ends on an interesting note about Europe: regulatory fragmentation (27 regimes, strict labor protections, linguistic barriers) may act as friction that slows the cascade. Not immunity. But in a cascade, friction is the difference between a managed transition and a structural break.

Then he identifies what he considers the real danger, and it’s darker than the software selloff. Every business that adopts AI tools to stay competitive is, in aggregate, teaching the AI platform the grammar of its entire industry. Not proprietary data (he’s careful about this) but the shape, structure, and patterns of the work. Each firm’s individual decision is rational. The collective result is a tragedy of the commons where the commons being destroyed is the economic moat of entire professional service industries.

He traces the cascade: professional services revenue losses propagate through commercial real estate, business travel, the venture ecosystem, and the tax bases of knowledge-economy cities. The $2 trillion in destroyed software market value, he writes, is not the damage. It is the down payment.

The essay is intellectually serious, and the institutional-stickiness argument is probably correct in the medium term. But there are two things Pignataro doesn’t see, or perhaps sees and doesn’t say.

First, the Wittgenstein reference is deeper than he takes it. Wittgenstein’s point about language games was that you cannot separate the game from the community that plays it. The rules only mean something because people share a form of life. When AI learns to play the game, the game doesn’t end. It continues. But the community whose participation gave it meaning is no longer necessary. The economic value was never in the tasks. It was in the fact that humans needed to coordinate with each other to get them done. Remove that necessity and you remove the reason for employing most of the people. Pignataro’s defense of institutional stickiness contains the seeds of a more radical conclusion than he allows himself to draw.

Second, read the essay through the lens of ION Group’s M&A history rather than its arguments.

ION spent over €3 billion acquiring Cedacri and Cerved in 2021. They invested €1.35 billion to take over Prelios in 2024. The company’s entire growth model is acquiring deeply embedded enterprise infrastructure companies at the right moment. Pignataro’s essay argues that institutional software will survive, but that the market is panicking as if all software is cognitive. If the market is wrong, institutional software is massively underpriced right now.

The essay is an intellectual framework for a buying thesis. Software stocks were trading at ten-year-low P/E ratios while their fundamentals remained strong. Pignataro published it on ION Analytics, his own financial data platform. His audience is institutional investors and corporate boards. His message to them: don’t panic-sell your enterprise software positions; the institutional layer has value that the market is ignoring. His message to himself: while everyone else is selling, this is the time to buy.

I can’t prove this reading, and I could be wrong. But the incentive alignment is too clean to ignore. The essay’s argument is at once a bear thesis on the broader economy and a contrarian bull thesis on specific institutional software companies. It says: the damage will be worse than you think, but it won’t hit where you think it will. That is exactly what you’d want a market to believe if you were about to go shopping.

Look deeper at all three, and the strategic layers multiply. Amodei’s real audience is other AI lab CEOs. The essay is a coordination move in the prisoner’s dilemma between labs: here is the framework for why we should all accept regulation, because the alternative is having it imposed after a catastrophe. The candor about bioweapons uplift and deceptive model behaviors serves a dual function — it demonstrates that Anthropic takes safety seriously enough to disclose uncomfortable findings, and it raises the reputational cost for competitors who don’t.

Pignataro’s real audience is the boards of acquisition targets. His message to them: your business is more defensible than the market currently believes. Don’t accept a fire-sale price. Come talk to me instead. The essay makes the intellectual case for why institutional software retains value — which is exactly what you’d want a potential seller’s board to believe before you open negotiations.

Citrini’s real audience is policymakers. The evidence surfaced quickly: co-author Alap Shah pivoted within days to Bloomberg Television calling for an AI tax. The speculative scenario was a vehicle for a policy proposal. The market reaction — billions wiped, front-page coverage — provided the urgency that policy arguments alone cannot generate.

Each of our three authors, it turns out, is engaged in strategic communication disguised as analysis. That doesn’t make any of them wrong. It makes them more interesting, because the strategic intent reveals what each author believes is the binding constraint on the problem: lab cooperation for Amodei, corporate strategy for Pignataro, government policy for Citrini. Their disagreements are genuine. Their positioning is also genuine. Both things are true at the same time.

IV. The Financial Clock

Citrini Research published the most provocative of the three, just five days after Pignataro’s essay. Written as speculative fiction from June 2028, looking back at how abundant intelligence broke the economy, the report triggered what Bloomberg called a selloff that sent IBM down 13% in its worst day in 25 years and dragged software, payments, and delivery stocks across the board. The Substack post racked up roughly 16 million views on X.

The premise is cleanly constructed: what if AI bullishness is correct, and that’s actually bearish?

Their scenario unfolds in four phases. First, the software layer collapses as agentic coding tools hit a capability threshold and CIOs begin asking why they’re paying $500,000 annually for SaaS they can replicate in weeks. The reflexivity trap kicks in: companies threatened by AI become AI’s most aggressive adopters, cutting headcount and redeploying savings into AI tools, which makes the next round of cuts possible. Each company’s response is rational. The collective result is catastrophic. Citrini includes a devastating detail: ServiceNow sold per-seat licenses, so when clients cut 15% of their workforce, they mechanically canceled 15% of their licenses. The same dynamic propagates through every business model that depends on the number of employed humans.

Second, friction goes to zero. By early 2027 in their scenario, AI agents become default consumer infrastructure. Agents running in the background optimize every purchasing decision, collapse subscription loyalty, price-match across platforms, and renegotiate insurance renewals. The most provocative chain: agents handling machine-to-machine commerce route around credit card interchange fees by settling on stablecoins, threatening the moats of companies like Mastercard and American Express. Their moats were made of friction, and friction went to zero.

Third, the white-collar services economy contracts because white-collar workers represent 50% of employment and drive roughly 75% of discretionary spending. When their earnings power is structurally impaired, the consumer economy follows. The negative feedback loop has no natural brake: AI improves, companies cut workers, workers spend less, margin pressure builds, companies invest more in AI, AI improves. Unlike previous technology transitions, displaced workers can’t pivot because AI is already capable of the pivot jobs. Citrini coins the term “Ghost GDP” for output that shows up in national accounts but never circulates through households.

Fourth, the financial system detonates. And this is where Citrini’s real contribution hides.

Everyone focused on the AI scenarios. Strip those away and look at what Citrini actually modeled in detail. The longest and most specific section of their report traces a contagion chain through private credit that reads like a 2008-style analysis wearing an AI costume.

The chain: PE firms (Apollo, Blackstone, KKR) acquired insurance companies and used annuity deposits as “permanent capital” to fund private credit lending. They deployed that capital into software leveraged buyouts at aggressive multiples based on the assumption that recurring revenue would keep recurring. The private credit market reached $1.7 to $2.1 trillion on narrow definitions. PE-backed software LBOs exceeded $440 billion from 2015 to 2025, with typical leverage of 5 to 7 times debt-to-EBITDA. When SaaS revenue declines, the loans default, the “permanent capital” turns out to be retail savings, and the chain collapses.

This chain isn’t speculative. The IMF, the BIS, and multiple Fed governors have independently flagged precisely this architecture as a systemic risk. The Fed’s March 2025 analysis found that life insurers’ exposure to below-investment-grade debt now exceeds the industry’s exposure to subprime RMBS in late 2007. As of early 2026, a record $25 billion in software leveraged loans trade below the distress threshold.

Even if Citadel is right that AI adoption follows an S-curve and job displacement is gradual, the private credit chain is fragile enough that a much smaller trigger could set it off. You don’t need 10.2% unemployment to cause software LBO defaults. You need a few quarters of decelerating growth, tighter refinancing conditions, and a narrative shift. Citrini provided the narrative shift.

The non-obvious implication runs deeper. For PE-backed software companies leveraged at 5 to 7 times EBITDA with interest coverage that has deteriorated from 3.5 times at closing to 1.6 times after rate hikes, even a modest revenue slowdown becomes an existential event. You don’t need Citrini’s doomsday scenario to trigger defaults. The 2021-vintage deals are the most exposed, underwritten for a world of zero interest rates and 15 to 25% annual growth that no longer exists regardless of what AI does.

Which surfaces the most uncomfortable insight in this entire sequence: reflexivity. A Substack post by a former Los Angeles paramedic and a hedge fund CIO moved billions in market value. That market movement validates the thesis that the system is fragile. The validation makes the post more credible. The credibility moves more market value. The essay itself became part of the phenomenon it was describing. This is reflexivity in its purest form, and it should make us nervous about the stability of any financial structure built on confidence in continued growth.

V. The Gaps Between the Clocks

The crisis lives not in any single clock but in the gaps between them. And here’s the part nobody wrote about: the gaps don’t just cause delay. They cause inversion. The mechanisms that are supposed to help make things worse.

The Inversion Problem

Consider financial markets. Their job, in theory, is to send price signals that help the real economy allocate resources. When Amodei published his essay, the market wiped $2 trillion from enterprise software. Was that a useful signal?

Pignataro argues it was noise, that the market confused cognitive automation with institutional replacement. But even if he’s right about the destination (institutional software survives longer than the market thinks), the market signal itself becomes destructive. A company whose stock drops 50% on a misreading of AI’s timeline faces real consequences: talent leaves, credit tightens, boards demand cost cuts, the best engineers defect to AI companies. The prediction becomes self-reinforcing even if the original analysis was wrong about mechanism. Markets repricing at digital speed can destroy companies that would have survived if they’d had time to adapt at institutional speed.

Now consider Pignataro’s own defense. His best argument, that enterprise software encodes organizational language games that take years to replace, contains an unacknowledged time bomb. He identifies it himself, almost as an aside: every firm that adopts AI tools to stay competitive is, in aggregate, teaching the AI platform the grammar of its entire industry. His corrective matters; enterprise contracts include no-training guarantees, and the transfer is through product telemetry and roadmap intelligence rather than direct data leakage. But the direction is the same. The moat doesn’t leak through the training pipeline. It leaks through the product cycle. And patterns are what AI actually needs.

Each mechanism designed to help (market signals, competitive adoption, institutional friction) operates at a timescale that puts it at odds with the others. Markets move too fast and overshoot. Firms adopt too fast and train their replacements. Institutions move too slowly and can’t build new structures before the old ones break.

The Distribution Thesis

This brings us to what I think is the deepest insight hiding across these three texts.

This is not a labor story. It is not a technology story. It is not a market story. It is a distribution story.

The modern economy doesn’t need workers for the sake of production. It needs workers because wages are the primary mechanism through which purchasing power gets distributed to the population. Companies produce goods and services. They pay workers. Workers buy goods and services. The loop closes. GDP is a measure of how fast that loop spins.

Citrini coined “Ghost GDP” for output that appears in national accounts but never circulates through the real economy. It’s a vivid phrase. But what it actually describes is the decoupling of production from distribution. AI produces. But it doesn’t consume. It doesn’t rent apartments, buy groceries, go on vacation, or service a mortgage. The productive capacity of the economy can grow indefinitely while the distribution mechanism that turns production into demand simply stops working.

Amodei’s “country of geniuses” metaphor accidentally reveals this. An actual country of geniuses would need food, housing, infrastructure, entertainment; it would participate in the economy as both producer and consumer. A datacenter of geniuses produces but consumes nothing except electricity. It’s the ultimate mercantilist economy: pure export, zero import. And we have no framework for dealing with a trading partner that doesn’t want anything back.

This is where the three authors converge without knowing it. Amodei proposes progressive taxation on AI firms. Pignataro warns that the revenue destruction in professional services will cascade into tax bases and city economies. Citrini models the feedback loop where shrinking consumer income collapses demand, which forces more AI adoption, which eliminates more jobs. They’re all pointing at the same broken pipe from different angles: the wage distribution mechanism is the structural vulnerability. Everything else (the software selloff, the SaaS repricing, the private credit exposure, the mortgage risk) is downstream.

The practical question for executives isn’t whether AI will displace cognitive work. That argument is settled in all three texts. The question is what happens to your business when your customers’ customers lose their income. Citrini models this explicitly: the ServiceNow detail about per-seat license cancellation isn’t a SaaS problem. It’s a preview of what happens to every business whose revenue model ultimately depends on the number of humans receiving paychecks. That includes landlords, grocery chains, auto dealers, universities, tax authorities, and pension funds. Think of the distribution mechanism as the plumbing underneath the entire economy.

Pignataro is right that institutional friction buys time. His observation about Europe’s regulatory fragmentation acting as a brake on cascade speed is useful for strategic planning. But time for what? None of the three fully answer that question. If the distribution mechanism is broken, buying time only helps if someone uses that time to build a new one. The Industrial Revolution took roughly a century to redistribute its gains through labor unions, progressive taxation, public education, and social insurance. Amodei says we don’t have a century. He may be right.

And the Citadel Securities rebuttal, which argues that technology diffusion follows S-curves and that Citrini underestimates the elasticity of human wants? Both points are historically sound. But they assume the economy’s plumbing (the mechanism by which productivity gains reach consumers) remains functional during the transition. That assumption is the thing all three of these authors, despite their different frameworks, agree is no longer safe to make.

The Prisoner’s Dilemma at Three Scales

The three essays describe the same game-theory structure at three different levels, and the game theory predicts the outcome.

Amodei at the lab level: “We have to build powerful AI because if we don’t, someone less careful will.” This is defection justified by the expectation of others’ defection. It’s the same logic Kaplan used to explain dropping the Responsible Scaling Policy: “it wouldn’t actually help anyone for us to stop training AI models if competitors are blazing ahead.”

Pignataro at the firm level: every company adopts AI to stay competitive, collectively training the system that makes their industries unnecessary. Each firm defects because not adopting means losing to competitors who did.

Citrini at the market level: companies cut workers and invest in AI because it’s individually rational, and the collective result is a deflationary spiral.

Same structure, three scales. The game theory is clear: in an iterated prisoner’s dilemma with many players and no enforcement mechanism, defection cascades. The only intervention that changes the outcome is a coordination mechanism that alters the payoff matrix for everyone simultaneously. That’s regulation. But regulation operates at the slowest clock speed of all.

The failure is structural, built into the game itself, not a matter of choosing the wrong policies. The right policies can’t arrive in time because the political process moves at institutional speed while the defection cascade moves at digital speed. The prediction falls out naturally: defection cascades until the damage is large enough to force coordinated government response. By that point, significant damage is already irreversible. This isn’t a flaw in the argument. It’s a flaw in the world.

Pignataro’s observation about European regulatory friction fits here better than he may realize. Europe’s existing regulatory infrastructure doesn’t prevent defection, but it slows the cascade enough that coordinated response might arrive before the damage becomes structural. The US, with lighter regulation and faster capital markets, gets the benefits of AI adoption sooner but also gets the deflationary consequences sooner, with less institutional capacity to manage the transition. For once, European regulatory sclerosis might be an asset.

The Measurement Crisis

There is a layer underneath all of this that none of the three authors fully articulated.

If a growing share of GDP is generated by AI (high output, near-zero labor input) then GDP stops measuring human economic activity. It measures aggregate production, which increasingly has no connection to human welfare. Central banks set interest rates based on GDP, inflation, and employment data. Fiscal policy is calibrated to GDP growth. Tax revenue depends on income and consumption that GDP is supposed to proxy.

In Citrini’s scenario, nominal GDP keeps growing while the human economy underneath it contracts. The Fed looks at strong GDP and doesn’t cut rates. Congress sees growth and doesn’t pass stimulus. The indicators are green while the patients are dying.

The failure has already begun. Simon Kuznets, the economist who built the national income accounting framework, warned in 1934 (on page 7 of his original report to the Senate) that “the welfare of a nation can scarcely be inferred from a measurement of national income.” Ninety-two years later, GDP remains the single most consequential number in economic policymaking, and the distortion Kuznets feared may finally be arriving. Christos Makridis and Erik Brynjolfsson documented in January 2026 three systematic channels through which AI-generated output escapes measurement: AI treated as expense rather than investment, mismeasured quality change, and the missing value of free or bundled AI services. Goldman Sachs estimated roughly $115 billion of AI-driven growth was not captured in official figures.

No government has convened a commission analogous to Stiglitz-Sen-Fitoussi (the effort launched in early 2008 to address GDP’s failure to signal deteriorating household welfare) for the AI era. No central bank anywhere uses non-GDP well-being metrics as monetary policy inputs. And in what may be the most revealing institutional detail of this entire story: the BEA’s Digital Economy Satellite Account, the one statistical tool purpose-built to track digital economic activity, was defunded in December 2023 due to budget constraints. The measurement infrastructure is not merely inadequate. It is moving in the wrong direction at the moment the challenge is intensifying.

If the Ghost GDP problem materializes at scale, current economic statistics would not detect it until the consequences were already embedded in household balance sheets. This is exactly the failure mode of 2008, when aggregate statistics looked healthy while individual households were drowning. The specific conditions that motivated the Stiglitz commission (aggregate output diverging from household experience) are precisely what AI-driven automation could reproduce. The tools to detect the divergence don’t exist. Nobody is building them.

VI. The Timeline Nobody Drew

Each of the three authors gives you a clock. Nobody overlaid them.

Amodei provides the capability timeline: one to two years to AI more capable than all humans. Pignataro provides the institutional resistance timeline: years to decades before coordination infrastructure is replaced. Citrini provides the financial contagion timeline: months to two years once the deflationary loop begins.

Laid on top of each other, they create a sequence that none of them individually described:

Capabilities are already here. Amodei’s bioweapons assessment confirms superhuman complex reasoning. The S-curve starts from a high baseline.
Market repricing is already happening. We watched it in real time between January and February 2026. $2 trillion gone in weeks.
Commodity cognitive work starts contracting within 6 to 12 months. Brynjolfsson’s “Canaries in the Coal Mine” research already shows 13% relative employment declines among workers aged 22-25 in AI-exposed occupations. The canaries are falling.
Financial contagion from private credit hits within 18 to 24 months. The plumbing exists. The $25 billion in distressed software loans is the kindling.
The institutional coordination layer begins eroding in 3 to 5 years. Pignataro’s friction buys time, but the training-your-replacement dynamic makes it self-defeating.
Full structural transformation takes a decade or more. If it follows historical patterns.

The danger zone is the 18-to-24-month window where financial contagion could hit before either the institutional layer has adapted or governments have built new distribution mechanisms. That’s roughly Citrini’s scenario. It’s also where the speed gap between clocks does its damage: fast enough to break things, too slow for new structures to form.

Notice what is absent from this timeline: government action. In all six phases, government policy appears nowhere, not because it’s irrelevant but because the political system operates at institutional speed on a problem that has already moved past the first three phases at digital speed. The playbook from 2008 (emergency legislation, TARP, quantitative easing) took months to design and implement. The 2020 pandemic response was faster but relied on checks and expanded unemployment insurance, both of which assume the existence of a functioning labor market to return to. Neither playbook fits a structural transformation of the labor market itself.

How fast is the gap between clocks widening? Consider the historical record.

The telephone took 65 years to reach 50% household adoption. Television took 6 years. The smartphone took 6 years. Generative AI reached 39.5% of US adults in just two years. The technology clock has accelerated by roughly one order of magnitude across successive waves.

But the labor clock has not moved at all. The China Shock literature (the most precisely measured labor displacement in economic history) shows adverse impacts on manufacturing employment and earnings persisting 19 or more years after the shock. Workers in trade-exposed communities “neither fully recover earnings losses nor predominantly exit the labor market, but rather age in place.” Recovery came through demographic replacement (young adults entering new jobs) not occupational retraining.

For electrification in the 1880s, both clocks ran at roughly the same speed: decades for adoption, decades for labor adjustment. The economy had time. For AI in the 2020s, the ratio of diffusion speed to adjustment speed has shifted from roughly 1:1 to something approaching 10:1. The technology arrives in years. People take generations to adjust. That is the structural gap, and no amount of S-curve argumentation addresses it.

The financial clock adds a third dimension of instability. As Carlota Perez documented in her study of technological revolutions and financial capital, financial markets consistently overshoot during the “installation period” of new technologies, pouring speculative capital into infrastructure before the technology has delivered its productivity payoff. The deepest labor restructuring occurs during the “deployment period” after the financial crash. If Perez’s framework holds for AI (and the current pattern of trillions in market capitalization built on top of a technology that the US Census Bureau found fewer than 6% of businesses using in production as recently as 2024 fits her model precisely) then we are in the installation frenzy. The deployment period, and the deep labor restructuring that comes with it, hasn’t started yet.

The historical parallel that keeps surfacing is the Engels Pause, the fifty-to-sixty-year period from the 1780s to the 1840s during which British output per worker grew substantially while real wages barely moved. Robert Allen’s canonical 2009 data shows the starkest version: between 1800 and 1830, output per worker grew at 0.63% per year while real wages grew at zero. Labor’s share of national income fell sharply while capital’s share tripled.

The magnitude of that gap is academically contested. Nicholas Crafts’s 2022 revision using superior GDP data narrows it substantially, and Gregory Clark argues that real wages grew as much as 46.5% over the same period, implying no meaningful pause at all. The honest assessment is that the pessimist position remains the disciplinary default but the consensus range has widened. What is not contested is this: the gap closed through a combination of capital accumulation, Corn Law repeal, factory legislation, and the emergence of labor unions. The political system needed forty years to produce enforceable reform. And the first meaningful response to visible economic distress was not a helping hand but the Combination Acts of 1799-1800, which banned trade unions outright. The Chartist movement (Britain’s first mass working-class political campaign) peaked during economic downturns and had all six of its demands eventually enacted. The last one, universal male suffrage, took eighty years.

If AI’s distribution problem follows even a compressed version of this pattern, the adjustment timeline extends well past 2050. And the honest assessment from the economic literature is that the magnitude of AI’s productivity impact remains deeply uncertain, ranging from Acemoglu’s conservative 0.53 to 0.66% TFP boost over a decade to Goldman Sachs’s aggressive estimate of 1.5% annual productivity growth for ten years. The distributional consequences scale directly with the magnitude, but the direction is the same in all scenarios: returns accrue to capital owners during the transition, and the market mechanism that is supposed to redistribute them runs at a speed that guarantees a painful gap.

VII. The Oldest Economic Story

The oldest economic story is arriving at its logical endpoint.

Every previous round of technological displacement was resolved because new technologies created forms of human labor that complemented the machines. The spinning jenny destroyed handloom weavers but created factory jobs. The automobile destroyed horse-drawn transport but created assembly line workers, gas station attendants, suburban construction crews. This worked because each previous technology was a substitute for human labor in one domain and a complement in others.

General-purpose AI is the first technology that substitutes across all cognitive domains simultaneously. It doesn’t replace weavers and create factory workers. It replaces knowledge workers and — here is the part the optimists haven’t answered — replaces whatever they’d pivot to as well.

Amodei says it directly: the technology acts as a general labor substitute. Citrini models the implication: in their scenario, AI improves at the exact tasks displaced workers would redeploy toward. Pignataro implies it through his tragedy of the commons: the grammar of ALL industries gets learned, not just one.

David Autor at MIT has staked out the most distinctive optimistic position: AI could, unlike previous computerization, extend expert capability to workers with foundational but not elite training, democratizing expertise rather than destroying it. It’s the strongest case for hope. But even Autor frames it as a possibility requiring deliberate policy, not a prediction. And he flags a demographic inversion absent from the Industrial Revolution: rich countries may “run out of workers before we run out of jobs.” Whether that’s reassuring or terrifying depends on your time horizon.

If the historical resolution mechanism (humans shifting to complementary tasks) doesn’t work this time, then the distribution problem isn’t transitional. It’s structural. And if it’s structural, the only solutions are ones that break the link between income and labor: universal basic income, sovereign wealth funds funded by AI taxation, ownership stakes in AI infrastructure, or something nobody has thought of yet.

That is the real conversation these three essays are circling around without quite reaching. Not “will AI take jobs?” That question is already exhausted. But “if AI breaks the wage mechanism permanently, what replaces it?” The answer to that question is the most important economic policy question of the next decade, and nobody in a position of power is working on it at the pace the technology demands. The economists who have engaged with it most seriously (Acemoglu, Autor, Brynjolfsson) agree on the diagnosis but offer diagnoses, not prescriptions. Acemoglu and Johnson’s Power and Progress documents that “it was a 100-year struggle during the Industrial Revolution for workers to get any cut of these massive productivity gains.” They’re not optimistic about that timeline compressing.

What would an adequate response look like? Not what Amodei proposes (surgical regulation and progressive taxation) because those operate at institutional speed on a problem moving at digital speed. Not what Pignataro implies (bet on the institutional layer and pray for friction) because the training-your-replacement dynamic makes institutional moats self-defeating over time. Not what Citrini suggests (hedge your portfolio) because financial positioning doesn’t solve a structural distribution failure. Each proposed solution inherits the limitations of the clock its author watches most closely. Amodei’s policy solutions require the very institutional speed he demonstrates is too slow. Pignataro’s institutional friction buys time that his own framework suggests will be used to train the replacement. Citrini’s financial hedging protects individual portfolios while the systemic risk grows. The problem demands coordination across all three clocks at once, and no institution (governmental, corporate, or financial) currently operates at that scope.

The honest answer is that nobody writing in March 2026 knows. But the shape of the problem is now visible, and these three documents, read together, outline it with unusual clarity.

The economy’s production engine is about to undergo the most dramatic upgrade in human history. Its distribution engine (wages for cognitive work) is running on architecture from the 19th century. The gap between those two facts is where the next decade of economic history will be written. We can see the shape of the problem now. Three authors, working from different vantage points with different fears and different interests, converged on the same broken pipe without realizing it. The builder, the king, and the portfolio managers each drew one clock. This essay has tried to draw all three on the same wall.

The picture is unsettling. But it is, at least, honest. And the authors who see it most clearly are, not coincidentally, the ones with the most at stake if they’re wrong. That should tell you something about who else ought to be paying closer attention.

All views expressed here are my own and do not represent the opinions or positions of my employer or any organization I am affiliated with.