Citrini Research published a piece this week called The 2028 Global Intelligence Crisis. It’s written as a fictional macro memo from June 2028, looking back at how AI capabilities exceeded every expectation - and how that was bearish, not bullish. The mechanism is simple: AI replaces knowledge workers, margins expand, stocks rally, but displaced workers stop spending, consumer demand collapses, companies automate harder to protect margins, and the loop tightens until the economy discovers that machines don’t buy groceries.¹ S&P drops 38%. Unemployment goes up to 10.2%. The piece calls this the “intelligence displacement spiral.”

Citrini is upfront that it’s a scenario, not a prediction. But the directional logic is sound, and it maps to something I’ve been thinking about since early 2025: the technology-driven deflation story, taken to its logical extreme.

How Quickly Can We Get There?

Dario Amodei, CEO of Anthropic, recently told Dwarkesh Patel that he believes we could have “models that are a country of geniuses in the data center in one to two years.”² If that’s even roughly right, the Citrini scenario becomes less hypothetical and more a question of sequencing and speed.

But economies have friction that technology doesn’t. Enterprise software contracts are multi-year. Procurement cycles are slow. Regulatory frameworks take years to adapt. Distribution networks - the relationships, trust, and institutional plumbing that connect products to customers - don’t get replaced overnight, even when the product is demonstrably better.³

I work in fintech and I can see the “build it ourselves” dynamic that the piece describes starting to play out in real time. Engineers at my company who wouldn’t have considered replacing software vendors two years ago are now seriously evaluating it, specifically because AI coding tools make the project manageable for the first time. But “seriously evaluating” is not “doing it next quarter.” There’s a gap between technical feasibility and institutional adoption, and that gap is measured in years, not months.⁴

So the direction is right. The speed is probably 3-5x slower than the piece suggests. Which still means dramatic change within a decade. And the speed doesn’t change the investment logic - it just changes how urgently you need to act on it.

The Deflationary Force

Strip away the narrative and the Citrini scenario is essentially just a deflation story.

AI makes the marginal cost of cognitive work approach zero. Knowledge workers lose jobs or take pay cuts. They spend less. Companies automate harder to preserve margins. Prices fall. Wages fall. The loop compounds. Labour’s share of GDP - already declining from roughly 64% in the early 1970s to around 57% today⁵ - accelerates downward.

This isn’t new. Technology-driven deflation has happened before. But previous waves of automation hit specific sectors - agriculture, then manufacturing - and displaced workers eventually moved into services and knowledge work. The historical pattern was painful but absorptive: the economy destroyed jobs in one category and created them in another. And the timelines were often longer.

The uncomfortable question the Citrini piece raises is: what if AI is good enough at the new categories too? If displaced product managers retrain as “AI coordinators” and AI can already do that job, the escape valve doesn’t work.⁶ I’m not certain this is true today. I’m less certain it won’t be true in five years.

Central Banks Cannot Allow Deflation

Here’s the piece’s biggest gap: it models the crisis in detail but barely touches the policy response.

Sustained deflation with a large debt overhang is an extinction-level event for the financial system. Every mortgage, every corporate bond, every government obligation becomes harder to service as incomes and revenues fall - even if interest rates go to zero, because the nominal debt stays fixed while the income to service it shrinks. Irving Fisher described this mechanism in 1933⁷ and it remains the nightmare scenario for every central banker alive.

The Fed knows this. They have demonstrated, repeatedly and with increasing aggression, that they will do whatever it takes to prevent a deflationary spiral. In 2008, the Fed’s balance sheet went from roughly $900 billion to $4.5 trillion over several years.⁸ In 2020, they expanded by over $4.6 trillion - more than all three previous QE programmes combined - with purchases in April 2020 alone exceeding $1 trillion.⁹

The response function is clear and it has been escalating. My expectation is that the scenario described in the Citrini piece would trigger fiscal and monetary intervention on a scale that makes 2020 look like a dress rehearsal. Direct transfers to displaced workers. Deficit spending at levels that would have been unthinkable five years earlier. Possibly a tax on AI inference compute.¹⁰ The political pressure of visible, widespread unemployment among the professional class - the people who vote, donate, and write opinion pieces - would be overwhelming and bipartisan.

The Bifurcated Economy

This is where it gets strange. I don’t think we get clean deflation or clean inflation. I think we get both, simultaneously, in different parts of the economy.

AI-exposed goods and services deflate. Software subscriptions, knowledge services, financial advice, legal work - anything where the value proposition was “I will navigate complexity that you find tedious” gets dramatically cheaper. The Citrini piece is right about this part.

But the stimulus response inflates everything else. Housing, food, energy, healthcare - things that are supply-constrained in the physical world and that absorb newly-printed money. You could end up in a world where your streaming subscription costs $2/month and your groceries are up 40%.¹¹

Adam Fergusson documented something structurally similar in When Money Dies - the Weimar hyperinflation didn’t hit all prices equally. Some goods became astronomically expensive while others remained relatively cheap for extended periods, creating bizarre distortions where a month’s rent cost less than a single meal.¹² The mechanism would be different here, but the principle - that monetary expansion doesn’t distribute evenly across the economy - is the same.

This bifurcation breaks traditional portfolio construction. Assets that are good inflation hedges (property, commodities) might behave differently from assets that benefit from deflation (long bonds, cash). You’re facing both forces at once, in different sectors, and the balance between them shifts depending on the policy response.

What This Means for Assets

I’m not a financial advisor.¹³ But I’ve thought about my own positioning through this lens, and the logic might be useful to someone else thinking through the same questions.

No debt. This is the single clearest implication. Fixed nominal obligations in a world of falling incomes is the worst possible position. If your salary is tied to knowledge work - and mine is - you don’t want a large mortgage or leveraged investments that require servicing regardless of what happens to your earning power. Being debt-free gives you the most valuable thing in a volatile environment: time to figure out what’s happening before you need to respond.¹⁴

Hard assets. Physical property in a desirable location holds value in real terms across most scenarios. There’s a particular version of this that I like: gold.¹⁵

Gold is insurance against the “central banks overcorrect” scenario, which I believe is the highest-probability response to AI-driven deflation. If the Fed prints aggressively enough to stop the deflationary spiral - and they will - gold benefits from both the monetary expansion and the erosion of institutional credibility. I own enough gold that I’m comfortable with either outcome - gold going nowhere while the rest of my portfolio does fine, or gold going through the roof while the rest of my portfolio burns.

Equity exposure to the AI transformation. In any scenario where AI continues to advance, the companies building it capture an enormous share of value. The displacement spiral is bad for the economy but good for the firms at its centre. The revenue growth rates at the frontier AI companies are unlike anything in the history of enterprise software - Anthropic went from $100 million to $14 billion in annualised revenue in three years, growing roughly 10x annually.¹⁶ Broad index exposure through something like MSCI World ETFs gives you upside participation without concentrating in single names, though if you can get direct exposure to the AI labs, the asymmetry is hard to ignore.

Optionality over optimisation. This is the Taleb influence.¹⁷ When the range of outcomes is genuinely wider than any model can capture, you don’t want a portfolio optimised for one scenario. The theoretical ideal is a barbell: very safe on one end (treasuries, inflation-linked bonds, cash), convex upside on the other (direct exposure to AI winners, maybe put options on the broader market), and nothing in the middle. How to determine “very safe” for this scenario is surprisingly hard, so my portfolio is not a traditional barbell, but the direction is right: as little as possible debt, hard assets as a floor, equity exposure to the transformation, and no leveraged bets on any single outcome. I’d rather be roughly right about the shape than precisely wrong about the specific scenario.

The combination amounts to: participate in the upside of the AI transformation through equity exposure, while holding insurance against the monetary chaos that transformation might trigger. It’s a portfolio that doesn’t require you to be right about the specific outcome - only about the direction of travel and the width of the distribution.

A Note from Outside the US

Most of the commentary on the Citrini piece - and it’s generated a lot - comes from an American perspective.¹⁸ I live in South Africa, and the dynamics here are different in ways that matter.

South Africa’s tax base is already narrow. The professional class that funds government transfers - SASSA grants, public services, infrastructure - is small relative to the population.¹⁹ If AI compresses knowledge-worker incomes globally, that tax base erodes at precisely the moment more people need support.²⁰

The counterweight is that South Africa is a gold producer, and most of the scenarios I’ve described are very good for gold prices. Whether mining revenue can offset the erosion of the professional tax base is probably the most important macro question for the country over the next decade. I don’t know the answer.

The Canary

The Citrini piece ends with the observation that you’re reading it in February 2026, not June 2028. That the negative feedback loops haven’t started. That the canary is still alive.

I’m not so sure. In 2024, colleagues laughed when I said AI would write better than most people. In 2025, others were hesitant to accept that AI could automate coding - not engineering, but the mechanical act of writing code. Given the rate of improvement we’re seeing now, and better tooling to give models the context they need,²¹ proper software engineering might be next. The feedback loops may not have reached crisis proportions, but the direction of travel is already visible to anyone paying attention. The canary is alive, but did I just hear a cough?

The policy response, when it comes, will be messy, late, and probably too aggressive rather than too cautious - because that’s how governments respond to crises that threaten the professional class.

The actionable implication isn’t to panic. It’s to make an honest assessment of the weaknesses in your financial position and your work - your income, your portfolio, your debt, your skills - and ask how much of it is built on the assumption that human intelligence remains scarce and valuable. For most knowledge workers, the answer is almost all of it.

If you don’t do that assessment in the next one to three years, the market will do it for you.

---

Footnotes

1. The Citrini piece coined the term “Ghost GDP” for this: output that shows up in the national accounts but never circulates through the real economy. ↩

2. Amodei said this in a February 2026 interview with Dwarkesh Patel. The full quote: “I really do believe that we could have models that are a country of geniuses in the data center in one to two years.” He immediately followed with a caveat about the lag between technical capability and economic returns - noting that even being wrong about revenue timing by a single year could be “ruinous.” It’s a remarkable combination of confidence in the technology and uncertainty about its economic consequences, which is basically the thesis of this entire post. See: “Anthropic CEO Dario Amodei explains his spending caution”, Fortune, February 14, 2026. ↩

3. If you want proof that distribution beats product, consider that Microsoft Teams exists. It is - and I will die on this hill - one of the worst pieces of software ever made. It is slow, unintuitive, and somehow manages to suck my Mac dry of all available RAM and storage space. And yet it has over 300 million monthly active users, because it comes free with the Microsoft 365 bundle that every enterprise already pays for. Teams won because it was there, not because it is any good. ↩

4. This is first-hand evidence of the mechanism the Citrini piece describes - the CIO reviewing a renewal and wondering “what if we just built this ourselves?” I’m watching it happen at my own company with our enterprise software. The engineers are not yet building the replacement. But they’ve stopped assuming the vendor is irreplaceable. That shift in assumption is the leading indicator. If you are a vendor, and you’re wondering whether I’m talking about you - yes. ↩

5. The labour share of US GDP has been declining for decades, though the exact magnitude depends on how you measure it. The BLS nonfarm business sector measure shows a decline from roughly 64% in the early 1970s to around 57% in recent years. See: “Estimating the U.S. labor share”, Monthly Labor Review, Bureau of Labor Statistics, 2017. For a more nuanced analysis showing that netting out depreciation and production taxes changes the picture, see: Karabarbounis, L., “Perspectives on the Labor Share”, Journal of Economic Perspectives, 2024. The Citrini piece projects it falling to 46% by 2028 due to AI. Whether or not you believe that specific number, the direction is hard to argue with. ↩

6. This is the weakest point in the Citrini piece but also the hardest to refute. The argument that “AI creates new jobs that AI can already do” sounds compelling but is essentially unfalsifiable - it assumes AI capability generalises perfectly to any new task category. Today that’s not true. But I’d argue the gap is narrower than most people think, and the constraint might be less about raw intelligence than about context. I wrote recently about how most people get generic, mediocre outputs from AI - not because the models are dumb, but because they provide no context. The same principle applies to replacing knowledge workers: AI fails at ambiguous product decisions or organisational politics not because it can’t reason about them, but because it doesn’t have the context. The moment someone builds systems that feed it the right context - org charts, meeting histories, preferences - the gap closes fast. The skill isn’t intelligence. It’s architecture. ↩

7. Fisher, I. “The Debt-Deflation Theory of Great Depressions,” Econometrica 1(4), 1933. Fisher’s insight was that deflation increases the real burden of debt, which forces debtors to sell assets, which depresses prices further, which increases the real burden of debt further. The loop is self-reinforcing and can be devastating. Fisher knew this from personal experience - he had been spectacularly wrong about the stock market in 1929 and lost most of his personal fortune. The debt-deflation theory was, in a sense, his post-mortem on his own financial ruin. ↩

8. Through three rounds of quantitative easing (QE1, QE2, QE3) from 2008 to 2014, the Fed’s balance sheet expanded nearly five-fold. From roughly $870 billion before the crisis to around $4.5 trillion. See: Richmond Fed, “The Fed Is Shrinking Its Balance Sheet”, 2022. ↩

9. The COVID-era expansion was staggering by any historical standard. Cumulative purchases exceeded $4.6 trillion, and in a single month - April 2020 - purchases exceeded $1 trillion. The Fed’s total balance sheet peaked at $8.97 trillion by early 2022. For context, the entire pre-2008 balance sheet was smaller than a single month of pandemic-era purchases. See: Kansas City Fed, “The Evolving Role of the Fed’s Balance Sheet”, Economic Review, 2022. ↩

10. The Citrini piece itself mentions a proposed “Transition Economy Act” with an AI compute tax. Governments have never met a crisis they couldn’t respond to with a tax on whatever caused it. Though, in this case, I don’t know of a better solution. ↩

11. 70% of US GDP is consumer spending. If the top 10% of earners - who account for nearly half of all consumer spending - lose their jobs or take substantial pay cuts, the consumption hit is enormous. But they’d still need to eat. ↩

12. Fergusson, A. When Money Dies: The Nightmare of the Weimar Inflation, 1975 (highly recommended reading). The book documents, in painful detail, how the German middle class was destroyed while certain asset holders and industrialists prospered. The chapter on how a piano could be purchased for the price of a week’s food - while a loaf of bread cost billions of marks - is particularly poignant. The lesson is that monetary distortions create non-linear, non-uniform effects that nobody predicts correctly in advance. ↩

13. This should go without saying, but the internet has taught me that nothing goes without saying. None of this is financial advice. I am a product manager who reads too much and has opinions that get me in trouble sometimes. If you make financial decisions based on a blog post written by a guy who recently built a todo app over Christmas, you deserve whatever happens to you. ↩

14. The Citrini piece makes the same point differently: “People borrowed against a future they can no longer afford to believe in.” In their scenario, the borrowers were technically current on their mortgages but had stopped all discretionary spending and were draining savings. One more shock away from distress. If your income depends on knowledge work and you have significant debt, you are structurally long the assumption that human intelligence remains valuable. That might be a fine bet. It might not be. I’d rather not find out the hard way. ↩

15. The standard objection to gold is that it “doesn’t produce anything.” This is correct. That’s the point. In a crisis of confidence in institutions and financial systems, you want something whose value doesn’t depend on any institution or system functioning correctly. Physical gold doesn’t need a functioning stock exchange, a solvent bank, or a stable government to hold its value. In times of turmoil, gold’s properties can come in very handy. ↩

16. Anthropic’s revenue trajectory: $100 million in 2023, $1 billion in 2024, ~$10 billion in 2025, $14 billion annualised as of February 2026. Dario Amodei confirmed these numbers at Davos in January 2026 and noted, correctly, that “obviously that curve can’t go on forever.” OpenAI’s revenue has been growing at roughly 3-4x per year, reaching about $20 billion in 2025. These are growth rates that simply don’t exist at this scale in the history of enterprise software. See: Epoch AI, “Anthropic could surpass OpenAI in annualized revenue by mid-2026”, February 2026. There’s a tension here worth acknowledging: the bull case for my equity portfolio - AI boosts margins, companies become more efficient - is the same mechanism that creates the bear case for my income and potentially my employer’s business model. The AI transformation is simultaneously the best thing to be invested in and the biggest threat to the income that funds the investment. ↩

17. Taleb’s barbell strategy, as described in Antifragile, involves combining extreme risk aversion on one end with small, speculative, convex bets on the other - and avoiding the middle. The middle is where you get “moderate” returns with hidden tail risk. ↩

18. As of writing, the Citrini piece has spawned at least one full-length bull-case companion piece, several response essays, and extensive commentary on Wall Street Oasis, Hacker News, and elsewhere. All from a developed-world, primarily American perspective. Which is understandable - the US economy is the subject - but the second and third-order effects on emerging markets with narrow tax bases are worth someone’s attention. ↩

19. South Africa has roughly 28 million SASSA grant recipients, disproportionately funded by a narrow personal income tax base (~5 million people above the tax-free threshold). That’s why the narrow tax base problem isn’t abstract here. It’s arithmetic. ↩

20. South Africa’s official unemployment rate exceeds 30%, with expanded unemployment closer to 42%. The informal economy and manual labour sectors are the last to be automated. But those sectors depend on a functioning state, which depends on tax revenue, which depends on the professional class. The chain is only as strong as its weakest link, and AI might be coming for the link that funds everything else. ↩

21. I wrote about this in January. The gap between “AI that produces generic slop” and “AI that does genuinely useful work” isn’t purely model quality - it’s context. The models are already good enough in many domains. What’s missing is the architecture that feeds them the right information about the specific task, the specific organisation, the specific constraints. That architecture is being built right now, by thousands of companies, and every improvement makes the next wave of displacement more feasible. ↩

More posts

4,500 Physicians Agree Feb 28, 2026 On breakfast, engagement rings, and regime change. essay psychology Why Killing Epstein Was a Catastrophic Engineering Failure Feb 19, 2026 In physics, a metastable system looks perfectly stable - until it isn't. The Epstein network held for thirty years. Then someone tried to stabilise it. essay systems Why Vampires Live Forever Feb 9, 2026 The modern longevity movement's obsession with blood transfusions makes perfect sense — if you accept that Peter Thiel and Bryan Johnson are vampires. essay longevity