2028 Global Intelligent Crisis: A Thought Experiment in Financial History from the Future

Author: CitriniResearch

Introduction

If our bullish expectations for AI continue to materialize… but what if this is actually a bearish signal?

The following describes a hypothetical scenario rather than an accurate prediction. This is not intended to incite panic, nor is it a fan fiction of AI doomsayers. The sole purpose of this article is to model a relatively underexplored scenario. Our friend Alap Shah posed the question, and we collaboratively conceived the answer. We wrote this section, and he wrote the other two parts.

We hope that after reading this article, you will be better prepared for potential left-tail risks as AI makes the economy increasingly bizarre.

The following is a macro memo published by CitriniResearch in June 2028, detailing the evolution and aftermath of the "Global Intelligence Crisis."

Macro Memo: The Consequences of Abundant Intelligence

CitriniResearch February 22, 2026 ~ June 30, 2028

The unemployment rate announced this morning reached 10.2%, exceeding expectations by 0.3%. The market sold off 2% in response to this data, leading to a cumulative drawdown of 38% from the S&P 500 index's peak in October 2026.

Traders have become numb. If this data had come out six months ago, it would have absolutely triggered a circuit breaker.

Just two years. That is the time it took for the economy to evolve from "risk manageable," "contained to specific sectors" to a form that none of us can recognize. This quarter's macro memo is our attempt to reconstruct this series of events—a post-mortem report on the pre-crisis economy.

Once upon a time, the market's euphoric sentiment was within reach. By October 2026, the S&P 500 index was nearing 8000 points, and the Nasdaq index had surpassed the 30,000 mark. The first wave of layoffs triggered by human obsolescence began in early 2026, perfectly achieving their intended goals: margin expansion, earnings exceeding expectations, and a soaring stock market. Record corporate profits were directly reinvested into AI computing power.

At that time, macro indicators remained bright. Nominal GDP repeatedly recorded mid-to-high single-digit annual growth. Productivity was thriving. Driven by AI agents that do not sleep, take sick leave, or require health insurance, the growth rate of real output per hour reached its highest level since the 1950s.

With the disappearance of labor costs, the wealth of computing power owners exploded. Meanwhile, real wage growth collapsed. Despite the government repeatedly boasting about record productivity, white-collar workers lost jobs to machines and were forced to shift to low-paying positions.

As cracks began to appear in the consumer economy, economic commentators popularized a term—"Ghost GDP": output reflected in national accounts that never circulated in the real economy.

In every aspect, AI exceeded expectations, and the market was AI. The only question was… the economy was not.

We should have understood long ago that a GPU cluster in North Dakota could produce the output of 10,000 white-collar workers in Midtown Manhattan; rather than being an economic panacea, it was more like an economic plague. The velocity of money stagnated. The human-centric consumer economy, which once accounted for 70% of GDP, withered. If we had asked earlier, "How much will machines spend on discretionary goods?" we might have figured this out long ago. (Hint: the answer is zero.)

AI capabilities improved, companies needed fewer employees, white-collar layoffs increased, laid-off employees consumed less, margin pressure forced companies to invest more in AI, and AI capabilities further improved…

This is a negative feedback loop without natural brakes. This is the human intelligence substitution spiral. White-collar workers watched their profitability (and, naturally, their consumption capacity) suffer structural damage. Their income was once the cornerstone of a $13 trillion mortgage market—which forced underwriters to reassess whether prime mortgages were still safe.

Seventeen years without a real default cycle left the private market flooded with private equity (PE)-backed software deals, predicated on the assumption that annual recurring revenue (ARR) would remain "recurring." By mid-2027, the first wave of defaults triggered by AI disruption challenged this assumption.

If this disruption had been limited to the software industry, the situation might have been manageable, but it was not. By the end of 2027, it threatened every business model built on intermediary services. Large numbers of companies that monetized through human friction collapsed.

It turned out that the entire system was a massive correlated bet built on expectations of white-collar productivity growth. The crash in November 2027 merely accelerated all the existing negative feedback loops.

We have been waiting for the logic of "bad news is good news" to work for nearly a year. The government began to consider proposing legislation, but public confidence in the government's ability to implement any form of rescue had waned. Policy responses are always lagging behind economic realities, but the lack of a comprehensive plan now threatens to accelerate the deflationary spiral.

How the Crisis Began

At the end of 2025, the capabilities of agentic coding tools experienced a leap.

A skilled developer using Claude Code or Codex can now replicate the core functionalities of a mid-market SaaS product in a matter of weeks. While not perfect and unable to handle all edge cases, it was enough to make a Chief Information Officer (CIO) reviewing a $500,000 annual renewal contract think, "What if we built one ourselves?"

Fiscal years typically align with calendar years, so corporate spending for 2026 was set in the fourth quarter of 2025, when "agentic AI" was still just a buzzword. The mid-year review was the first time procurement teams made decisions with a clear understanding of what these systems could actually do. Some watched their internal teams launch prototypes in weeks, replicating six-figure SaaS contracts.

That summer, we interviewed a procurement manager from a Fortune 500 company. He told us about a budget negotiation he had. The salesperson expected to play the same game as last year: a 5% price increase annually, the standard "your team can't live without us" sales pitch. The procurement manager told him he had been discussing with OpenAI to have their "frontline deployment engineers" use AI tools to completely replace this vendor. In the end, they renewed at a 30% discount. He said it was a good outcome. Those "long-tail SaaS" companies, like Monday.com, Zapier, and Asana, were in much worse shape.

Investors were prepared—even expecting—that the long-tail market would be severely impacted. They might account for a third of typical enterprise tech stack spending, but they were clearly exposed to risk. However, "record systems" were supposed to be immune to disruption.

It wasn't until ServiceNow released its Q3 2026 earnings report that the mechanism of reflexivity became clear.

SERVICENOW's net new ACV growth slowed from 23% to 14%; announced layoffs of 15% and initiated a "structural efficiency plan"; stock price plummeted 18% | Bloomberg, October 2026

SaaS was not "dead." Running and supporting internally built systems still required cost-benefit analysis. But internal building became an option and was factored into pricing negotiations. Perhaps more importantly, the competitive landscape had changed. AI made it easier to develop and release new features, thus collapsing differentiation. Legacy companies competed on pricing—both in a close-quarters battle with each other and in a struggle against the constantly emerging newcomers. Inspired by the leap in agentic programming capabilities and unburdened by legacy cost structures, these newcomers aggressively seized market share.

It was not until this earnings report was released that people fully realized the interconnected nature of these systems. ServiceNow charged per seat. When Fortune 500 clients laid off 15%, they canceled 15% of their licenses. Those AI-driven layoffs that improved client profit margins were mechanically destroying their own revenue base.

A company selling workflow automation was being disrupted by better workflow automation, and its response was to lay off employees and use the saved funds to fund the very technology that was disrupting it.

What else could they do? Sit back and die a slower death? The companies most threatened by AI became the most aggressive adopters of AI.

This seems obvious in hindsight, but it was not at the time (at least not to me). Historical disruption models suggest that legacy companies resist new technologies, lose market share to agile entrants, and then slowly die. This is what happened to Kodak, Blockbuster, and Blackberry. But what happened in 2026 was different; legacy companies did not resist because they could not afford to.

Faced with stock prices dropping 40-60% and pressure from boards to explain, companies threatened by AI did the only thing they could do: lay off employees, redeploy the saved funds into AI tools, and use those tools to maintain output at a lower cost.

Each company's individual response was rational. But the collective result was catastrophic. Every dollar saved from layoffs flowed into AI capabilities, making the next round of layoffs possible.

The software industry was just the opening act. While investors were still debating whether SaaS valuation multiples had bottomed out, they missed that this reflexive cycle had jumped out of the software industry. The same logic that justified ServiceNow's layoffs applied to every company with a white-collar cost structure.

When Friction Approaches Zero

By early 2027, using large language models (LLMs) had become the default behavior. People were using AI agents that didn’t even know what "AI agents" were, just as those who had never learned what "cloud computing" was used streaming services. They viewed it like they viewed autocomplete or spell check—just something their phones could now do.

The open-source agent shopping bot Qwen (Tongyi Qianwen) was a catalyst for AI taking over consumer decision-making. Within weeks, every mainstream AI assistant integrated some form of agent commerce functionality. Model distillation meant these agents could run not only on cloud instances but also on phones and laptops, significantly lowering the marginal cost of inference.

What should have made investors uneasy (but didn’t) was that these agents did not need to wait to be awakened. They operated in the background based on user preferences. Commerce was no longer a series of discrete human decisions but a continuous optimization process, running 24/7 on behalf of every connected consumer. By March 2027, the median American was consuming 400,000 tokens daily—ten times what it was at the end of 2026.

The next link in the chain began to break.

Intermediation.

For the past fifty years, the U.S. economy had built a massive "rent-seeking layer" on top of human limitations: doing things took time, patience was limited, brand familiarity replaced due diligence, and most people were willing to accept a bad price to avoid clicking a few more times. Trillions of dollars in enterprise value depended on the continued existence of these constraints.

At first, it was simple. Agents eliminated friction.

Those subscriptions and memberships that default to auto-renew even if not used for months. Those entry prices that secretly double after the trial period. All of these were redefined as "hostage crises" that agents could negotiate. As a metric foundational to the entire subscription economy—average customer lifetime value—declined significantly.

Consumer agents began to change the way nearly all consumer transactions operated.

Humans do not have the time to compare prices across five competing platforms before buying a box of protein bars. But machines do.

Travel booking platforms were among the earliest victims because they were the simplest. By Q4 2026, our agents could assemble a complete itinerary (flights, hotels, ground transportation, loyalty optimization, budget constraints, refunds) faster and cheaper than any platform.

The insurance renewal model was completely reshaped, as its entire model relied on the inertia of policyholders. Agents that re-quoted coverage for you every year dismantled the 15-20% premium that insurance companies earned from passive renewals.

Financial advice. Tax preparation. Routine legal work. Any category of service provider whose value proposition boiled down to "I will handle the tedious complexities you find boring" was disrupted because agents do not find anything boring.

Even those areas we thought were protected by "interpersonal value" proved to be vulnerable. In real estate, buyers had endured 5-6% commissions for decades due to the information asymmetry between agents and consumers. However, once AI agents equipped with MLS access and decades of transaction data could instantly replicate that knowledge base, the industry collapsed. A seller's research report from March 2027 referred to it as "agent-on-agent violence." The median buyer commission in major metropolitan areas compressed from 2.5-3% to less than 1%, and increasingly, buyers in more transactions had no human agents involved at all.

We overestimated the value of "interpersonal relationships." It turns out that what people call relationships is largely just "friction with a friendly mask."

This was just the beginning of the destruction of the intermediation layer. Successful companies had spent billions effectively exploiting the weaknesses of consumer behavior and human psychology, but now none of that mattered.

Machines optimized for price and fit did not care about your favorite app, did not care about the websites you habitually opened for the past four years, and did not feel the allure of a carefully designed checkout experience. They did not get tired and accept the simplest option, nor did they default to "I always order here."

This destroyed a specific moat: habitual intermediation.

DoorDash (DASH US) is a typical example.

Programming agents dismantled the entry barriers to launching delivery apps. A capable developer could deploy a fully functional competitor in weeks, and many did. They attracted drivers away from DoorDash and Uber Eats by passing 90-95% of delivery fees directly to them. Multi-app dashboards allowed gig workers to track orders from twenty or thirty platforms simultaneously, eliminating the lock-in effects existing businesses relied on. The market fragmented overnight, and margins were compressed to nearly zero.

Agents accelerated the destruction on both the supply and demand sides. They empowered competitors and then used them. DoorDash's moat literally meant "you're hungry, you're lazy, this is the app on your home screen." Agents had no home screen. They would check DoorDash, Uber Eats, the restaurant's own website, and twenty newly written alternatives to choose the lowest fee and fastest delivery every time.

Machines did not have habitual app loyalty, which was the foundation of the entire business model.

This has a strange poetic quality; it may be the only instance in this entire event where agents helped the white-collar workers about to be replaced. When they eventually became delivery drivers, at least half of their income did not go to Uber and DoorDash. Of course, with the proliferation of self-driving cars, this boon from technology did not last long.

Once agents controlled transactions, they began to look for bigger prey.

The potential for price comparison and aggregation is limited. The most significant way to repeatedly save users money (especially as agents began trading with each other) is to eliminate fees. In machine-to-machine commerce, the 2-3% card interchange fee became an obvious target.

Agents began to look for faster and cheaper options than credit cards. Most chose to use stablecoins via Solana or Ethereum L2, where settlements were nearly instantaneous, and transaction costs were measured in fractions of a cent.

Mastercard Q1 2027: Net revenue grew 6% year-over-year; purchase volume growth slowed from +5.9% in the previous quarter to +3.4%; management noted "agent-led price optimization" and "pressure in discretionary spending categories" | Bloomberg, April 29, 2027

Mastercard's Q1 2027 earnings report was a point of no return. Agent commerce shifted from a product story to an infrastructure pipeline story. Mastercard (MA) plummeted 9% the next day. Visa also dropped, but its decline was mitigated after analysts pointed out its stronger positioning in stablecoin infrastructure.

Agent commerce bypassed interchange fees, posing a far greater risk to banks centered on card issuance and single-business card issuers. These institutions collected the vast majority of that 2-3% fee and built entire business departments around rewards programs funded by merchant subsidies.

American Express (AXP US) was hit hardest; it faced dual pressures: white-collar layoffs hollowed out its customer base, while agents bypassed interchange fees hollowed out its revenue model. Synchrony (SYF US), Capital One (COF US), and Discover (DFS US) also fell more than 10% in the following weeks.

Their moat was made of friction. And friction was approaching zero.

From Industry Risk to Systemic Risk

Throughout 2026, the market viewed the negative impacts of AI as an "industry-level" story. The software and consulting industries were being hit hard, payment and other "toll booth" sectors were teetering, but the broader economy seemed unscathed. The labor market, while softening, was not in free fall. The prevailing consensus was that creative destruction was part of any technological innovation cycle. It would be painful in some areas, but the overall net benefits brought by AI would outweigh any negative impacts.

In our January 2027 macro memo, we pointed out that this was a flawed mental model. The U.S. economy is a white-collar service economy. White-collar workers account for 50% of employment and drive about 75% of discretionary consumer spending. The businesses and jobs being consumed by AI are not on the margins of the U.S. economy; they are the U.S. economy itself.

"Technological innovation creates more jobs even as it eliminates jobs." This was the most popular and persuasive counterargument at the time. It was popular and persuasive because it had been true for the past two centuries. Even if we could not envision what future jobs would look like, they would surely arrive.

ATMs reduced the operating costs of bank branches, so banks opened more branches, and the number of tellers actually increased over the next twenty years. The internet disrupted travel agencies, yellow pages, and brick-and-mortar retail, but it created entirely new industries and birthed new jobs.

However, every new job requires a human to perform it.

AI is now a general intelligence, and it is continuously improving on those tasks that humans would have been redeployed to. Displaced programmers cannot simply turn to "AI management" because AI already possesses management capabilities.

Today, AI agents are responsible for handling R&D tasks that last for weeks. While business school professors try every year to fit data into a new S-curve, exponential growth has crushed our understanding of possibilities.

They wrote nearly all the code. The best-performing agents are far smarter than nearly all humans in almost everything. And they are getting cheaper.

AI does create new jobs. Prompt engineers. AI safety researchers. Infrastructure technicians. Humans remain in the loop, coordinating at the highest levels or curating taste. But for every new role created by AI, it has eliminated dozens of old roles. And the salaries for new roles are just a fraction of the old roles.

U.S. JOLTS: Job openings fall below 5.5 million; the ratio of unemployed persons to job openings rises to about 1.7, the highest level since August 2020 | Bloomberg, October 2026

Hiring rates remained sluggish throughout the year, but the JOLTS data from October 2026 provided some decisive evidence. Job openings fell below 5.5 million, down 15% year-over-year.

INDEED: Job postings in software, finance, and consulting sectors plummeted as the "productivity initiative" spread | Indeed Hiring Lab, November-December 2026

White-collar job openings were collapsing, while blue-collar openings (construction, healthcare, technical workers) remained relatively stable. Attrition was concentrated in roles that wrote memos (somehow, we are still in business), approved budgets, and kept the economy's middle class running. However, real wage growth for both groups remained negative for most of the year and continued to decline.

The stock market remained largely indifferent to the JOLTS data; they were more concerned with the news that General Electric's Vernova had sold all its turbine capacity through 2040, oscillating between negative macro news and positive AI infrastructure news.

However, the bond market (always smarter than the stock market, or at least less romantic) began to price in the consumer hit. The yield on the 10-year Treasury bond began a decline from 4.3% to 3.2% over the next four months. Even so, the overall unemployment rate did not explode, and the structural nuances were still overlooked by some.

In a normal economic recession, the causes would eventually self-correct. Overbuilding leads to a slowdown in construction, which in turn leads to lower interest rates, stimulating new construction. Inventory excess leads to destocking, triggering restocking. Cyclical mechanisms contain the seeds of their own recovery.

The cause of this cycle is not cyclical.

AI is getting better and cheaper. Companies lay off employees and then use the saved money to buy more AI capabilities, allowing them to lay off more employees. Laid-off employees consume less. Companies selling goods to consumers sell less, weakening their strength, so they invest more in AI to protect margins. AI becomes better and cheaper.

A negative feedback loop without natural brakes.

The intuitive expectation is that the decline in aggregate demand will slow the pace of AI construction. But that has not happened because this is not capital expenditure (CapEx) in the style of hyperscale computing facilities. This is an operational expenditure (OpEx) substitution. A company that previously spent $100 million annually on employees and $5 million on AI is now spending $70 million on employees and $20 million on AI. AI investment is growing exponentially, but it is happening as part of a reduction in total operating costs. Every company's AI budget is growing while its overall spending is shrinking.

Ironically, even as the economies being disrupted by AI begin to deteriorate, the AI infrastructure complex continues to perform excellently. Nvidia (NVDA) is still reporting record revenues. TSMC (TSM) is still operating at over 95% utilization. Hyperscale cloud providers are still spending $150-200 billion each quarter on data center capital expenditures. Economies that fully highlight this trend, such as Taiwan and South Korea, are performing significantly better than the broader market.

India, on the other hand, is the opposite. The country's IT services sector exports over $200 billion annually, making it the largest contributor to India's current account surplus and a source of funds used to offset its ongoing goods trade deficit. The entire model is built on a value proposition: the cost of Indian developers is only a fraction of their American counterparts. However, the marginal cost of AI programming agents has collapsed to essentially just the cost of electricity. In 2027, Tata Consultancy Services (TCS), Infosys, and Wipro saw a surge in contract cancellations. With the service sector surplus anchoring India's external accounts evaporating, the rupee fell 18% against the dollar in four months. By Q1 2028, the International Monetary Fund (IMF) had begun "preliminary discussions" with New Delhi.

The engine of disruption is getting stronger each quarter, meaning the pace of disruption is accelerating each quarter. The labor market has no natural bottom.

In the U.S., we are no longer asking how the AI infrastructure bubble will burst. We are beginning to ask, what happens to a consumer credit-based economy when consumers are replaced by machines.

The Spiral of Intelligent Substitution

2027 was a year when the macroeconomic narrative became no longer subtle. The transmission mechanisms of those disjointed but clearly negative trends over the past 12 months became evident. You don’t need to look at the U.S. Bureau of Labor Statistics (BLS) data; just attending a friend's dinner party would tell you.

Displaced white-collar workers were not idle. They "downgraded." Many took lower-paying jobs in the service sector and gig economy, increasing the labor supply in those areas and driving down wages there.

One of our friends was a senior product manager at Salesforce in 2025. With a title, health insurance, a 401k pension, and an annual salary of $180,000, she lost her job in the third round of layoffs. After six months of searching, she began driving for Uber. Her income dropped to $45,000. The focus is not on the individual story but on the second-order mathematical effects. Multiply this dynamic by hundreds of thousands of workers in every major metropolitan area. An overqualified labor force flooded into the service sector and gig economy, driving down wages for existing workers who were already struggling. Industry-specific disruptions worsened into wage compression across the entire economy.

The remaining human-centric labor pool had yet to experience a correction, and it was happening at the moment we were writing these words. This is because automated delivery and self-driving cars were sweeping through the gig economy that absorbed the first wave of displaced workers.

By February 2027, it was clear that the consumption patterns of professionals still employed resembled those who might be next to be laid off. They doubled down on working hard (mainly with the help of AI) just to avoid being fired, and hopes for promotions or raises had evaporated. Savings rates slightly increased, and spending remained weak.

The most dangerous part was the lag. High-income earners used their above-average savings to maintain a normal appearance for two to three quarters. Hard data did not confirm the existence of the problem until it had already become old news in the real economy. Then, data that broke this illusion was published.

U.S. initial jobless claims surged to 487,000, the highest level since April 2020 | Department of Labor, Q3 2027

Initial jobless claims soared to 487,000, setting a record high since April 2020. ADP and Equifax confirmed that the vast majority of new applicants were white-collar professionals.

The S&P 500 index dropped 6% in the following week. Negative macro factors began to gain the upper hand in the tug-of-war.

In a typical economic recession, unemployment is broadly distributed. The pain borne by blue-collar and white-collar workers is roughly proportional to each group's share of total employment. The consumer hit is also broadly distributed and will soon manifest in the data, as low-income workers have a higher marginal propensity to consume.

In this cycle, unemployment was concentrated in the highest income deciles. They represented a relatively small share of total employment, but they drove a disproportionate share of consumer spending. The top 10% of earners accounted for more than half of all consumer spending in the U.S. The top 20% accounted for about 65%. These individuals buy homes, cars, take vacations, dine out, pay private school tuition, and renovate homes. They are the demand foundation for the entire non-essential consumer economy.

When these workers lose their jobs or accept a 50% pay cut to take available positions, the consumer hit is enormous relative to the number of lost positions. A 2% decline in white-collar employment translates to an impact on discretionary consumer spending of about 3-4%. Unlike blue-collar unemployment, which typically has an immediate impact (you get laid off from the factory, and you stop consuming next week), the effects of white-collar unemployment have a lag but are deeper because these workers have a savings buffer that allows them to maintain spending for months before a fundamental behavioral shift occurs.

By Q2 2027, the economy was in recession. The National Bureau of Economic Research (NBER) would not officially determine the start date of the recession for several months (they always do), but the data was clear—we had experienced two consecutive quarters of negative real GDP growth. But at this point, it was not yet a "financial crisis"… temporarily.

The Chain of Correlated Bets

Private credit grew from less than $1 trillion in 2015 to over $2.5 trillion in 2026. A significant portion of that capital was deployed in software and technology deals, many of which were leveraged buyouts (LBOs) based on the assumption that revenue would forever maintain mid-to-high double-digit growth.

These assumptions had died between the first agentic coding demonstrations and the software stock crash in Q1 2026, but the mark-to-market values of these assets seemed unaware that they were dead.

When many publicly traded SaaS companies' trading prices fell to 5-8 times EBITDA, the book values of PE-backed software companies on balance sheets still reflected acquisition valuations based on revenue multiples that no longer existed. Management gradually lowered book values from $1.00 to $0.92, $0.85, while publicly traded comparables were valued at $0.50.

Moody's downgraded the ratings of 14 issuers of up to $18 billion in private equity-backed software debt, citing "long-term revenue headwinds from AI-driven competitive disruption"; this was the largest single-industry downgrade action since the 2015 energy crisis | Moody's Investor Service, April 2027

Everyone remembers what happened after the downgrades. Industry veterans had seen this playbook after the 2015 energy downgrades.

Software-backed loans began to default in Q3 2027. PE portfolio companies in the information services and consulting sectors followed suit. Several billion-dollar leveraged buyouts involving well-known SaaS companies entered restructuring.

Zendesk is the definitive evidence.

ZENDESK failed to meet debt covenants due to AI-driven customer service automation eroding ARR; $5 billion in direct loan facilities were marked at 58 cents; setting the record for the largest private credit software default in history | Financial Times, September 2027

In 2022, Hellman & Friedman and Permira took Zendesk private for $10.2 billion. The debt portfolio included $5 billion in direct loans, the largest ARR-backed credit facility in history, led by Blackstone, with Apollo, Blue Owl, and HPS also in the lending syndicate. The explicit structure of this loan was predicated on the assumption that Zendesk's annual recurring revenue (ARR) would remain stable. At a leverage ratio of about 25 times EBITDA, leverage only made sense under that premise.

By mid-2027, that premise no longer existed.

For over half a year, AI agents had begun autonomously handling customer service. The category defined by Zendesk (tickets, routing, managing human customer interactions) had been replaced by systems that could resolve issues without generating tickets. The "annual recurring revenue" that served as the basis for loan underwriting was no longer recurring; it was merely revenue that had not yet left.

The largest ARR-backed loan in history turned into the largest private credit software default in history. Every credit trading desk was simultaneously asking the same question: who else had disguised long-term headwinds as cyclical headwinds?

But this was where the initial consensus was correct (at least at the start): this should have been survivable.

Private credit is not banking from 2008. The entire structure was explicitly designed to avoid forced liquidation. These are closed-end vehicles that lock up capital. Limited partners (LPs) committed for seven to ten years. No depositors would run, no repurchase agreements would be unwound. Managers could hold impaired assets, gradually resolve them over time, and wait for recovery. Painful, yes, but manageable. The system was designed to bend but not break.

Executives at Blackstone, KKR, and Apollo cited that software risk exposure accounted for only 7-13% of assets. The risk was manageable. Every sell-side report and Twitter (fintwit) credit influencer was saying the same thing: private credit had permanent capital. They could absorb losses large enough to blow up leveraged banks.

Permanent capital. This term appeared in every earnings call and investor letter aimed at reassuring. It became a mantra. Like most mantras, no one paid attention to the finer details. Here is what it really meant…

Over the past decade, large alternative asset management firms acquired life insurance companies and turned them into funding vehicles. Apollo acquired Athene. Brookfield acquired American Equity. KKR acquired Global Atlantic. The logic was elegant: annuity deposits provided a stable, long-term liability base. Fund managers invested those deposits into the private credit they originated and earned double compensation—earning a spread on the insurance and management fees on the assets. A "fee-on-fee" perpetual motion machine that worked well under one condition.

The premise was that private credit had to guarantee principal safety.

Losses hit the balance sheets holding illiquid assets to counter long-term liabilities. What should have made the system resilient—"permanent capital"—was not some abstract, patient institutional money that could bear complex risks. It was the savings of American households, "Main Street," structured as annuities, invested in the same PE-backed software and technology notes that were now defaulting. Those locked-up capital that could not be withdrawn were the money of life insurance policyholders, and in that realm, the rules are a bit different.

Compared to the banking system, insurance regulators had always been docile—even somewhat complacent—but this was the moment to sound the alarm. Regulators already uneasy about the highly concentrated private credit in life insurance companies began to tighten the risk-based capital (RBC) treatment of these assets. This forced insurance companies to either raise capital or sell assets, but neither could be achieved on attractive terms in an already frozen market.

New York and Iowa regulators took action to tighten capital treatment for certain privately rated credits held by life insurers; NAIC (National Association of Insurance Commissioners) guidance is expected to increase RBC factors and trigger additional scrutiny | Reuters, November 2027

When Moody's downgraded Athene's financial strength rating to a negative outlook, Apollo's stock plummeted 22% in two trading days. Brookfield, KKR, and others followed suit.

The situation became increasingly complex. These companies not only created their insurance perpetual motion machines, but they also built a complex offshore structure designed to maximize returns through regulatory arbitrage. U.S. insurance companies wrote annuities and then transferred the risk to their Bermuda or Cayman subsidiary reinsurers—these companies were established to take advantage of more flexible regulations that allowed them to hold less capital on the same assets. The subsidiary raised external capital through offshore special purpose vehicles (SPVs), and this layer of new counterparties invested alongside the insurance companies in private credit initiated by the same parent company's asset management division.

Rating agencies, some of which were themselves owned by PE, were not models of transparency (no one was surprised by this). The spiderweb of different companies connecting different balance sheets was astonishingly opaque. When the underlying loans defaulted, who exactly bore the losses was a question that was genuinely impossible to answer in real-time.

The crash in November 2027 marked a shift in perception from a potential ordinary cyclical pullback to something more unsettling. Federal Reserve Chairman Kevin Warsh referred to it at the FOMC's emergency meeting in November as: "a massive correlated bet chain built on expectations of white-collar productivity growth."

You see, the crisis was never about the losses themselves. It was about acknowledging (recognizing) those losses. In finance, there is a much larger, far more important area that we are increasingly afraid to acknowledge.

The Mortgage Problem

ZILLOW Home Value Index shows San Francisco down 11% year-over-year, Seattle down 9%, Austin down 8%; Fannie Mae indicates "early delinquency rates are rising" in postal code areas with over 40% tech/finance employment | Zillow / Fannie Mae, June 2028

This month, the Zillow Home Value Index showed an 11% year-over-year decline in San Francisco, a 9% decline in Seattle, and an 8% decline in Austin. This is not the only concerning headline. Last month, Fannie Mae pointed out that "jumbo-heavy" postal code areas had high early delinquency rates—these areas are inhabited by borrowers with credit scores over 780, typically considered "bulletproof" (extremely safe).

The U.S. residential mortgage market is approximately $13 trillion in size. Mortgage underwriting is built on a fundamental assumption: borrowers will maintain roughly their current income levels throughout the life of the loan. In most mortgages, this means thirty years.

The white-collar employment crisis threatens this assumption through a continuous change in income expectations. We now have to ask a question that seemed absurd three years ago—Are prime mortgages still safe?

Every previous mortgage crisis in U.S. history was driven by one of three causes: excessive speculation (lending money to people who cannot afford homes, as in 2008), interest rate shocks (rising rates making adjustable-rate mortgages unaffordable, as in the early 1980s), or localized economic shocks (a single industry collapsing in a single region, such as the Texas oil industry in the 1980s or the Michigan auto industry in 2009).

None of these apply to the current situation. The borrowers in question are not subprime. They have 780 FICO credit scores. They made a 20% down payment. They have good credit histories, stable work records, and their income was verified and documented at the time of loan origination. They are borrowers that every risk model in the financial system views as the cornerstone of credit quality.

In 2008, loans were bad from day one. In 2028, loans are good from day one. It’s just that in this world… something changed after the loans were written. People borrowed money betting on a future they can no longer afford.

In 2027, we marked the early signs of invisible pressure: surging home equity line of credit (HELOC) withdrawals, 401(k) withdrawals, and skyrocketing credit card debt, while mortgage repayments remained current. Due to unemployment, hiring freezes, and bonus cuts, these prime households watched their debt-to-income ratios double.

They could still repay their mortgages, but only if they stopped all discretionary spending, drained their savings, and postponed any home maintenance or improvements. Technically, their mortgages were still current, but they were just one step away from distress, and the trajectory of AI capabilities suggested that this shock was imminent. Subsequently, we saw delinquency rates in San Francisco, Seattle, Manhattan, and Austin begin to soar, even though the national average remained within historical normal ranges.

We are now in the most critical phase. If marginal buyers (i.e., those who might take over) are healthy, then the decline in home prices is manageable. And here, marginal buyers are facing the same income loss issues.

Despite growing concerns, we have not yet fallen into a full-blown mortgage crisis. While delinquency rates have risen, they remain far below the levels seen in 2008. The real threat lies in its trajectory.

Today, the spiral of intelligent substitution has two financial boosters accelerating the decline of the real economy.

Labor substitution, mortgage concerns, private market turmoil. Each reinforces the other. And traditional policy toolkits (rate cuts, quantitative easing QE) can address the issues of the financial engine but cannot solve the problems of the real economy, because the real economy is not driven by tightening financial conditions. It is driven by the decline in human intelligence due to AI leading to scarcity and devaluation. You can lower interest rates to zero and buy up all the mortgage-backed securities (MBS) and defaulted software leveraged buyout debt…

But that does not change the fact that one Claude agent can do the work of a $180,000 product manager for the cost of $200 a month.

If these fears become reality, the mortgage market will collapse in the second half of this year. In that case, we expect the current drawdown in the stock market to ultimately rival that of the Global Financial Crisis (GFC) (a 57% drop from peak to trough). This would pull the S&P 500 index down to about 3,500 points—levels we have not seen since a month before the ChatGPT moment in November 2022.

It is clear that the income assumptions supporting $13 trillion in residential mortgages have suffered structural damage. What is unclear is whether policy can intervene before the mortgage market fully digests this implication. We are hopeful, but we cannot deny the reasons that make us feel pessimistic.

The Race Against Time

The first negative feedback loop occurred in the real economy: AI capabilities improved, total wages shrank, consumption weakened, margins tightened, companies purchased more AI capabilities, and AI capabilities further improved. Then, the crisis spread to the financial sector: income damage hit mortgages, banks lost money and tightened credit, the wealth effect shattered, and the feedback loop accelerated. Both were exacerbated by the government's confusion and inadequate response in the face of the crisis.

Our system was not designed for this crisis. The federal government's revenue base is essentially a tax on human time. People work, companies pay them, and the government takes a cut. In normal years, personal income taxes and payroll taxes are the pillars of government revenue.

Until the first quarter of this year, federal tax revenues were 12% lower than the Congressional Budget Office (CBO) baseline forecast. With fewer people able to command those high salaries, payroll tax revenues are declining. As people's absolute income structurally declines, income tax revenues are also falling. Productivity surged, but the gains flowed to capital and computing power, not labor.

The share of labor in GDP has fallen from 64% in 1974 to 56% in 2024, driven by a slow decline over forty years due to globalization, automation, and steadily weakening worker bargaining power. But in the past four years, as AI began to advance exponentially, that share has plummeted to 46%. This is the largest drop on record.

Output still exists. But it no longer circulates back to businesses through households, meaning it no longer flows through the IRS either. The circular flow is breaking down, and people expect the government to intervene to fix it.

As with every economic recession, spending increases while income declines. But this time is different; the spending pressure is not cyclical. Automatic stabilizers are designed to respond to temporary unemployment, not structural substitution. The premise of the system paying out benefits is that workers will be reabsorbed into the job market. But many will not, at least not reemployed at close to their previous wage levels. During the COVID pandemic, the government readily accepted a 15% deficit because it was widely recognized as a "temporary" situation. However, the groups needing government support today are not those who experienced a recoverable pandemic; they are those permanently replaced by an evolving technology.

Just as the government is collecting less tax revenue from households, it needs to transfer more funds to households.

The U.S. will not default. It prints the money it uses to consume, the same money it uses to repay borrowers. But this pressure has already manifested elsewhere. Municipal bonds have shown troubling signs of divergence in performance year-to-date. States without income taxes are faring reasonably well, but states reliant on income taxes (mostly blue states) are beginning to price in default risk for general obligation municipal bonds. Politicians quickly took notice, and the debate over "who should be rescued" rapidly devolved into a partisan struggle.

It is certain that the government recognized the structural nature of the crisis early on and began deliberating on a bipartisan proposal they called the Transition Economy Act: a framework for directly transferring payments to unemployed workers through a combination of deficit spending and taxing AI inference computing.

The most radical proposal on the table goes further. The Shared AI Prosperity Act proposes establishing public claims on the profits of intelligent infrastructure, which would fall somewhere between a sovereign wealth fund and royalties on AI-generated output, using those proceeds to fund household transfers. Unsurprisingly, private sector lobbyists flooded the media warning of the catastrophic slippery slope this approach would bring.

The political games behind the policy discussions are extremely cliché, filled with grandstanding and marginal policies. The right denounces wealth transfers and redistribution as Marxist and warns that taxing computing power is tantamount to ceding leadership to China. The left warns that tax laws drafted with the help of vested interests will only be a form of regulatory capture. Fiscal hawks point to unsustainable deficits. Doves point to the premature tightening policies implemented after the Global Financial Crisis (GFC) as a cautionary tale. As the presidential election approaches this year, these divisions will only be amplified.

While politicians bicker, the pace of societal fractures far outstrips the legislative process.

The "Occupy Silicon Valley" movement became a microcosm of widespread public discontent. Last month, protesters blocked the entrances to Anthropic and OpenAI's San Francisco offices for three consecutive weeks. Their numbers grew, and the media coverage attracted by this protest even surpassed the unemployment data that triggered it.

It is hard to imagine that the public could hate anyone more than they hated bankers in the aftermath of the Global Financial Crisis, but AI labs are trying to break that record. And from the public's perspective, this hatred is understandable. The wealth accumulation of AI lab founders and early investors makes the Gilded Age look modest. The gains from the productivity boom have almost entirely accrued to the owners of computing power and the lab shareholders operating on it, amplifying inequality in America to unprecedented levels.

Each side has its own villain in mind, but the real villain is time.

The pace of evolution of AI capabilities is faster than the pace of institutional adaptation. Policy responses are advancing at the speed of ideology, not the speed of reality. If the government does not quickly reach a consensus on the core of the problem, the feedback loops mentioned above will write the next chapter in their stead.

The End of the Intelligence Premium

Throughout modern economic history, human intelligence has been the most scarce input factor. Capital is abundant (or at least replicable). Natural resources are limited but can be substituted. The pace of technological advancement has been slow enough that humans have had time to adapt. Only intelligence—the ability to analyze, decide, create, persuade, and coordinate—cannot be replicated at scale.

The intrinsic premium of human intelligence stems from its scarcity. Every institution in our economic system, from the labor market to the mortgage market to tax law, is designed for a world where this assumption holds.

Now, we are experiencing the unwinding of this premium. In an increasing number of tasks, machine intelligence is becoming a competent and rapidly advancing substitute for human intelligence. The financial system—having optimized for decades to adapt to a world of scarce human minds—is being repriced. This repricing is painful, chaotic, and far from over.

But repricing does not equate to collapse.

The economic system can find a new equilibrium. How to get there is one of the few tasks that can only be accomplished by humans. We must get this right.

For the first time in history, the most productive assets in the economy are producing fewer, not more, jobs. No one's analytical framework applies because none were designed for a world where "scarce factors become extremely abundant." Therefore, we must build new frameworks. Whether we can establish these frameworks in time is the only question that matters.

However, you are not reading this article in June 2028. You are reading it in February 2026.

The S&P 500 index is near historical highs. The negative feedback loop has not yet begun. We are confident that some of these scenarios will not materialize. We are equally confident that machine intelligence will continue to accelerate. The premium for human intelligence will shrink.

As investors, we still have time to assess how much of our portfolios are built on assumptions that will not survive this decade. As a society, we still have time to prepare for the worst.

The canary in the coal mine is still alive.