A memo from 2028: If AI wins, what will we lose?

Written by: Citrini and Alap Shah

Compiled by: Deep Tide TechFlow

Introduction: If the AI ​​story continues to unfold according to the bull market script—productivity skyrockets, corporate profits hit record highs, and computing power stocks sweep everything—then who will be the losers?

This article doesn't give you the answer; it gives you a scenario: Suppose all our current optimistic predictions about AI are proven correct. Then what?

CitriniResearch's "June 2028 Macro Memo" is a deliberate thought experiment. Starting from today, 2026, it uses a reverse chronological order to reconstruct how a crisis that was never fully priced in by the market gradually escalated: white-collar unemployment → consumption collapse → private equity lending defaults → "ghost GDP" → mortgage market turmoil.

At the heart of the entire logical chain, there is only one sentence: AI has eliminated friction, but 70% of the US economy is built on human "laziness".

This is a "risk disclosure statement" written for those who still have time to re-examine their investment portfolios.

Foreword

What if our optimistic assessment of AI is correct... and that's exactly the bad news?

The following is a scenario simulation, not a prediction. This is not bear market pornography, nor is it fan fiction by AI doomsday theorists. The sole purpose of this article is to model a severely underestimated scenario. Our friend Alap Shah posed this question, and we brainstormed the answer together. This section was written by us; he also wrote two other articles, links to which are provided at the end.

Hopefully, after reading this, you will be better prepared for the potential left-tail risks that AI may bring to the increasingly bizarre state of the economy.

The following is a macro memo from CitriniResearch for June 2028, detailing the evolution and consequences of the "Global Intelligence Crisis".

Macro Memorandum

The Cost of AI Overabundance

This morning's unemployment data showed an increase of 10.2%, 0.3 percentage points higher than expected. The market fell 2% on the day, and the S&P 500 has now fallen 38% from its October 2026 high.

Traders have become desensitized. Six months ago, such data would have triggered a circuit breaker.

In just the last two years, the economy has transformed from "controllable" and "limited to a few industries" to an unrecognizable entity. This quarterly macroeconomic memo is a retrospective attempt to reconstruct this sequence—an analysis of the pre-crisis economy.

The excitement at the time was real.

In October 2026, the S&P 500 nearly reached 8,000 points, and the Nasdaq surpassed 30,000 points. At the beginning of 2026, the first wave of layoffs triggered by the replacement of humans began, and the effects of these layoffs were exactly what layoffs were supposed to have: profit margins expanded, earnings exceeded expectations, and stock prices rose. Record-breaking corporate profits were continuously poured back into AI computing power.

The numbers on the surface remain impressive. Nominal GDP has recorded mid-to-high single-digit annualized growth for several consecutive quarters. Productivity has exploded. Real output per hour is growing at its fastest pace since the 1950s—driven by AI agents that don't need sleep, don't take sick leave, and don't require health insurance.

Owners of computing power watched as labor costs evaporated, and their wealth skyrocketed. Meanwhile, real wage growth collapsed. Despite repeated government boasts of "record productivity," white-collar workers were losing their jobs and being forced into lower-paying positions.

As cracks began to appear in the consumer economy, economic commentators coined a new term: "ghost GDP"—output that appears in national accounts but never actually circulates in the real economy.

AI has exceeded expectations in every dimension; the market is AI. The only problem is… the economy isn't.

It should have been clear long ago: a GPU cluster in North Dakota generating all the output previously produced by 10,000 white-collar workers in Midtown Manhattan is less an economic boon and more an economic plague. The velocity of money is approaching zero. The human-centric consumer economy, which accounts for 70% of GDP, is beginning to shrink. If we had asked one question earlier, perhaps we could have discovered the problem sooner: How much money have machines spent on freely disposable goods? (Answer: Zero.)

AI capability enhancement → corporate downsizing → replacement workers cut spending → profit pressure forces companies to increase AI investment → AI capability enhancement again...

This is a negative feedback loop with no natural brakes. It's a spiral of human intelligence replacing human capabilities. The earning power (and rationally driven consumption) of white-collar workers is structurally damaged. Their income is the cornerstone of the $13 trillion mortgage market—forcing underwriters to re-examine: are prime mortgages still prime assets?

For seventeen years, there had been no real default cycle, and the private equity market was saturated with PE-backed software deals that valued ARR indefinitely. The first wave of defaults triggered by AI disruption in mid-2027 shattered this assumption.

If the damage had been confined to the software industry, it would have been manageable. But it wasn't. By the end of 2027, the threat had spread to all business models that relied on middleware monetization. Companies built on exploiting human friction collapsed en masse.

The entire system was originally a long chain of interconnected bets on white-collar productivity growth. The collapse in November 2027 merely accelerated all the already running negative feedback loops.

We've waited almost a year for the day when "bad news turned into good news." The government has begun discussing various bailout plans, but public confidence in the government's ability to launch any form of rescue has dried up. Policy lagging behind economic reality has always been the norm, but the lack of a comprehensive response is now threatening to accelerate the deflationary spiral.

starting point

By the end of 2025, the capabilities of AI programming tools will have seen a dramatic leap forward.

A skilled developer, working with Claude Code or Codex, can now replicate the core functionality of a mid-range SaaS product within weeks. It's not perfect, and not all edge cases are handled perfectly, but it's enough to prompt a CIO reviewing a $500,000 annual renewal contract to ask, "What if we built it ourselves?"

Fiscal years mostly align with calendar years, so enterprise procurement spending for 2026 was finalized in Q4 of 2025 when "AI agents" were still just a buzzword. The mid-year review was the first time procurement teams made decisions with a clear understanding of the true capabilities of these systems. Some witnessed their internal teams build prototypes within weeks that could replace six-figure SaaS contracts.

That summer, we spoke with a purchasing manager at a Fortune 500 company. He recounted a budget negotiation experience. The seller was still hoping for the old tricks: a 5% annual price increase, the standard "your team can't do without us" rhetoric. The purchasing manager told him he was already in talks with OpenAI, having their "frontline deployment engineers" use AI tools to directly replace the vendor. Ultimately, they renewed their contract with a 30% discount. "That's a good result," he said. Those so-called "SaaS long tails"—Monday.com, Zapier, Asana, and the like—are in a much worse situation.

Investors were prepared—and even somewhat expected—for the long tail to be hit hard. These long tails may account for a third of the spending in a typical enterprise technology stack, but they are clearly exposed. Core software at the "recording system" level, on the other hand, is considered to be secure.

The mechanism of this reflexivity only became clear in ServiceNow’s Q3 2026 financial report.

ServiceNow's net annual contract value growth rate plummeted from 23% to 14%; announced 15% layoffs and a "structural efficiency plan"; stock price fell 18% | Bloomberg, October 2026

SaaS is not dead. Self-built solutions still involve trade-offs in terms of operation and maintenance costs. However, self-built solutions have become an option, and this option influences price negotiations. More importantly, the competitive landscape has changed. AI has lowered the barrier to developing and releasing new features, causing differentiation to collapse. Established vendors are locked in fierce competition, while also facing price wars with a new wave of challengers who have risen to prominence through AI programming capabilities and have no historical baggage to protect.

This time, the interdependence between systems won't be clear until this financial report is released. ServiceNow makes money by selling licenses. When a Fortune 500 client cuts 15% of its staff, it automatically cancels 15% of its licenses. Those AI-driven layoffs that are pushing profit margins on client-side are mechanically destroying their own revenue base.

This company that sells workflow automation is being disrupted by better workflow automation, and its response is to lay off employees and invest the money saved in the technology that is disrupting it.

What else can it do? Just sit there and slowly wait to die?

Ironically, the companies most threatened by AI have become the most aggressive adopters of AI.

This statement seems obvious in hindsight, but it wasn't at the time (at least not to me). The historical pattern of disruption is: established companies resist new technologies, gradually lose market share to agile newcomers, and slowly die. This is the story of Kodak, Blockbuster, and BlackBerry. What happened in 2026 was different: established companies didn't resist because they couldn't afford to.

With stock prices halved and boards of directors demanding answers, companies threatened by AI have only one option: lay off employees, pour the savings into AI tools, and then use those tools to maintain their original output at a lower cost.

The individual rationality of each company has converged into a collective disaster. Every penny saved from wages flows into AI capabilities, and stronger AI capabilities make the next round of layoffs possible.

Software is just the opening act. While investors are still debating whether SaaS valuations have bottomed out, the reflexive cycle has already escaped the software industry. The same logic that justifies ServiceNow's layoffs applies to every company with a white-collar cost structure.

When friction reaches zero

By early 2027, the use of large language models will have become the default option. People will be using AI agents, but won't even know what an "AI agent" is—just as people don't understand what "cloud computing" is, yet still use streaming services. They'll treat it like autocomplete or spell check—which is what phones do now, automatically.

Tongyi Qianwen's open-source intelligent shopping agent has become a catalyst for AI to replace humans in making consumer decisions. Within weeks, all mainstream AI assistants have integrated some form of intelligent business functionality. Distillation models mean that these agents can run on mobile phones and laptops without relying on the cloud, significantly reducing the marginal cost of inference.

What should truly unsettle investors is that these intelligent agents don't wait for user commands. They run continuously in the background according to user preferences. Business activities are no longer a series of discrete human decisions, but a continuously running optimization process, operating 24/7, representing every connected consumer without interruption. In March 2027, the average daily token consumption by US users reached 400,000—ten times that at the end of 2026.

The next link in the chain has already begun to break.

Intermediate layer.

Over the past fifty years, the American economy has built a massive rent-collecting layer on top of human limitations: things take time, patience eventually runs out, brand familiarity replaces prudence, and most people are willing to accept a terrible price just to avoid the hassle of clicking a few more times. Trillions of dollars in corporate value depend on these constraints existing indefinitely.

It started simply: intelligent agents eliminate friction.

Those subscriptions and memberships you haven't used for months but are still being passively and automatically renewed. The entry-level pricing that secretly increases after the trial period. Each of these is being redefined as a negotiable "hijacking," with an AI agent negotiating on your behalf. The core metric upon which the entire subscription economy is built—customer lifetime value—is clearly beginning to decline.

Consumer agents are beginning to change the way almost all consumer transactions operate.

Humans simply don't have the time to compare prices on five different platforms before buying a box of protein bars. Machines do.

Travel booking platforms were the first victims because it was the easiest. By Q4 2026, our AI agent will be able to piece together a complete itinerary (flights, hotels, ground transportation, points optimization, budget control, refunds) faster and cheaper than any other platform.

Insurance renewal relies on the policyholder's inertia; the intelligent agent that helps you re-compare protection plans every year undermines the 15-20% of premiums that insurance companies earn from passive renewals.

Financial advisors. Tax filing. Routine legal matters. Any service provider whose value proposition ultimately boils down to "I'll help you with those complicated things you find tedious" has been completely disrupted because the intelligent agent finds nothing tedious.

Even those areas we thought were protected by the "value of interpersonal relationships" have proven incredibly fragile. Real estate—where buyers have tolerated 5-6% commission rates for decades due to information asymmetry between real estate agents and consumers—collapsed instantly once AI agents with MLS data access and decades of transaction data could instantly replicate this knowledge base. A March 2027 sell-side report called it "agent-on-agent violence." Median buyer commissions in major cities have shrunk from 2.5-3% to below 1%, and more and more transactions involve buyers without any human agents at all.

We overestimated the value of "interpersonal relationships." In reality, many of the "relationships" people talk about are nothing more than friction disguised as friendship.

This is just the beginning of the disruption of the middle layer. Successful companies have spent billions of dollars effectively exploiting the quirks of consumer behavior and human psychology—quirks that are now irrelevant.

Machines that optimize pricing and matching don't care about your favorite apps, the website you've been habitually opening for four years, or the allure of a well-designed checkout experience. They won't tirelessly choose the easiest option or default to "I always order from here."

This destroyed a particular type of moat: the habitual middle layer .

DoorDash has become the most typical representative of all this.

AI programming tools have eroded the barriers to entry for building food delivery apps. A skilled developer can deploy a functional competitor in weeks, and dozens can do just that. They pass 90-95% of delivery fees directly to riders, poaching riders from Domino's and Uber Eats. Multi-platform dashboards allow gig workers to track orders from 20-30 platforms simultaneously, completely eliminating the lock-in effect that established platforms relied on. The market fragmented overnight, and profit margins were squeezed to near zero.

Intelligent agents accelerate both sides of the disruption. They spawn competitors, only to then use them themselves. DoorDash's moat is essentially: "When you're hungry or lazy, this app is on your home screen." Intelligent agents don't have a home screen. They simultaneously check DoorDash, Uber Eats, restaurant websites, and twenty newly emerging competitors, always choosing the one with the lowest cost and fastest delivery.

For machines, the concept of app loyalty simply doesn't exist.

This is the only slightly ironic episode in the whole story: technology did a small favor for those white-collar workers facing unemployment. When they eventually turned to food delivery, at least half of their earnings no longer went to Uber and DoorDash. Of course, this small favor from technology didn't last long—self-driving vehicles soon followed.

Once the agents control the transactions, they begin looking for larger "paperclips".

Price comparison and aggregation have their limitations. For users, the biggest and most effective way to save money repeatedly (especially once agents start trading with each other) is to eliminate transaction fees. In machine-to-machine commercial transactions, the 2-3% credit card transaction fee becomes a clear target.

Smart agents began looking for faster and cheaper alternatives to credit cards. Most opted to use stablecoins via Solana or Ethereum L2, with settlements nearly instantaneous and transaction costs measured in cents.

Mastercard Q1 2027: Net revenue up 6% YoY; spending growth slows to 3.4% from 5.9% in the previous quarter; management cites "smart agent-driven price optimization" and "pressure on discretionary categories" | Bloomberg, April 29, 2027

Mastercard's Q1 2027 earnings report marked an irreversible turning point. Smart agent commerce transformed from a product story into an infrastructure story. Mastercard fell 9% the following day. Visa followed suit, but its losses narrowed after analysts pointed to its stronger position in stablecoin infrastructure.

Intelligent agent commerce bypasses credit card transaction fees, posing a far greater threat to banks and single card issuers that rely on credit cards—they collect the bulk of the 2-3% transaction fees and have built their entire business line around cashback programs that subsidize merchants.

American Express was hit hardest: on one hand, massive layoffs of white-collar workers eroded its customer base, and on the other hand, smart agents bypassing credit card fees destroyed its revenue model. Synchrony, Capital One, and Discover all fell by more than 10% in the following weeks.

Their moats are built on friction. And that friction is now reaching zero.

From industry risk to systemic risk

Throughout 2026, the market treated the negative impacts of AI as an industry-wide story. Software and consulting were battered, and payments and other "tollbooth" businesses teetered on the brink, but the overall macroeconomy looked relatively healthy. The labor market, while softening, didn't freefall. The prevailing view was that creative destruction is part of any technological innovation cycle. Localized pain would be intense, but the overall positive benefits of AI would outweigh any negative effects.

In our January 2027 macro memo, we pointed out that this is a flawed framework. The U.S. economy is a white-collar service sector-driven economy. White-collar workers make up 50% of total employment and drive approximately 75% of discretionary consumer spending. The businesses and jobs that AI is eroding are not on the periphery of the U.S. economy—they are the U.S. economy itself.

"Technological innovation will destroy jobs, but it will then create more." This was the most popular and persuasive rebuttal at the time. It was popular and persuasive because it had been true for the past two centuries. Even if we cannot imagine what future jobs will be, they will eventually arrive.

ATMs made bank branches cheaper, so banks opened more branches, and the number of teller employees continued to rise over the next two decades. The internet disrupted travel agencies, yellow pages, and brick-and-mortar retail, but it also created entirely new industries and created new jobs.

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

AI is now a form of general intelligence, and it continues to advance in precisely the tasks that humans would turn to do. Unemployed programmers cannot simply switch careers to "manage AI"—because AI can already do that.

Today, AI agents can handle research and development tasks that can last for weeks. The most advanced ones are far smarter than almost any human in almost everything. And they are becoming increasingly cheaper.

AI has indeed created new jobs: prompt word engineers, AI security researchers, infrastructure technicians. Humans are still involved, coordinating at the highest levels, or setting the tone and direction. But for every new job AI creates, dozens of old ones become redundant. The salaries for the new jobs are a fraction of those for the old ones.

JOLTS Report: Job openings fall below 5.5 million; the unemployment-to-open-jobs ratio climbs to about 1.7, the highest since August 2020 | Bloomberg, October 2026

Hiring intentions remained weak throughout the year, with the October 2026 JOLTS report providing decisive data. Job openings fell below 5.5 million, a 15% year-over-year decline.

Indeed: Hiring Plummets in Software, Finance, and Consulting Industries as "Productivity Improvement Plans" Spread | Indeed Recruitment Lab, November-December 2026

White-collar jobs are collapsing, while blue-collar jobs (construction, healthcare, skilled trades) are relatively stable. The job losses are concentrated in positions responsible for writing memos (which we still do to some extent), approving budgets, and lubricating the middle layer of the economy. However, real wage growth for both groups has been negative for more than six months and continues to decline.

The stock market's attention to JOLTS remains less than that of the news that "GE Winnow's gas turbine capacity is sold out until 2040," with the stock fluctuating sideways between negative macroeconomic data and positive AI infrastructure headlines.

The bond market (always smarter than the stock market, or at least less romantic) began to absorb the impact on the consumer side. The 10-year US Treasury yield gradually declined from 4.3% to 3.2% over the next four months. However, the overall unemployment rate did not rise significantly, and its structural nuances were still overlooked by some.

In a normal economic recession, the underlying causes eventually correct themselves. Over-construction leads to a slowdown in construction, which in turn triggers interest rate cuts, stimulating new construction. Excess inventory leads to destocking, followed by restocking. The seeds of self-repair are embedded within the cyclical mechanism.

The root cause of this round is not cyclical.

AI is getting better and cheaper. Companies lay off employees and use the savings to buy more AI capabilities, which in turn allow them to lay off even more people. The replaced workers reduce expenses. Companies that sell to consumers are selling less and declining, so they invest more in AI to maintain profit margins. AI is getting better and cheaper.

A negative feedback loop without natural braking.

Intuitively, one would expect that a decline in aggregate demand would slow down investment in AI infrastructure. However, this hasn't happened, because this isn't a massive, operator-style capital expenditure, but rather a substitution of operating costs. A company that previously spent $100 million annually on employees and $5 million on AI now spends $70 million on employees and $20 million on AI. AI investment has multiplied several times, but this is reflected in a decrease in overall operating costs. Every company's AI budget is increasing, but overall spending is shrinking.

This creates an irony: even as the economies AI is disrupting begin to deteriorate, the AI ​​infrastructure complex continues to perform strongly. Nvidia is still recording record revenue. TSMC is still operating at over 95% capacity utilization. Hyperscale cloud vendors are still spending $150 billion to $200 billion per quarter on data center capital expenditures. Economies with pure convexity to this trend—Taiwan and South Korea—are significantly outperforming.

India presents a mirror image. Its IT services sector, exporting over $200 billion annually, is the largest contributor to India's current account surplus and a hedge against its long-standing goods trade deficit. The entire model is built on a value proposition: Indian developers' costs are a fraction of their American counterparts. However, the marginal cost of AI programmable agents has collapsed to near the cost of electricity. Contract cancellations at Tata Consultancy, Infosys, and Wipro accelerated throughout 2027. As India's service sector surplus, which underpins its external account, evaporated, the rupee depreciated by 18% against the dollar in four months. In Q1 2028, the IMF and New Delhi began "preliminary consultations."

The engines driving disruption are getting better every quarter, and disruption is therefore accelerating every quarter. The labor market has no natural bottom.

In the United States, we've stopped asking when the AI ​​infrastructure bubble will burst. What we're asking is: what will happen to an economy that relies on consumer credit when consumers are replaced by machines?

Intelligent replacement spiral

In 2027, the macroeconomic story will finally be clear. The transmission mechanisms of the scattered but clearly negative developments over the past twelve months are now obvious. You don't need to delve into BLS data; attending a dinner with an old friend will suffice.

The white-collar workers who were replaced didn't sit idly by. They downgraded. Many moved into lower-paying service and gig economy jobs, which increased the labor supply in these segments and further depressed the wages of already struggling workers.

We have a friend who, in 2025, was a senior product manager at Salesforce. Position, health insurance, 401k, annual salary of $180,000. She was laid off in the third round of layoffs. After searching for six months, she started driving for Uber. Her income dropped to $45,000. The important thing isn't this person's story itself, but the second-order mathematics behind it: multiply this dynamic by the hundreds of thousands of workers in major metropolitan areas. The influx of a highly saturated workforce into the service sector and the gig economy further depressed the wages of existing workers who were already struggling. Industry-wide disruption spilled over into a nationwide wage crunch.

The remaining pool of human-centered jobs awaits another round of adjustments—even as we write this. Autonomous delivery and driverless vehicles are penetrating the gig economy, which absorbed the first wave of unemployed workers.

By February 2027, it will be clear that professionals still employed are consuming based on the assumption that "they might be next." They (mostly with the help of AI) are working desperately just to avoid being laid off, the idea of ​​promotion and salary increases has long been abandoned, savings rates are rising, and consumption is softening.

The most dangerous aspect is the lag effect. High-income groups used above-average savings to maintain a semblance of normalcy for two or three quarters. By the time hard data confirmed the problem, it was already old news in the real economy. Then, the figure that shattered the illusion appeared.

The number of Americans filing for unemployment benefits for the first time surged to 487,000, the highest level since April 2020; according to the U.S. Department of Labor, Q3 2027.

Initial jobless claims surged to 487,000, the highest level since April 2020. ADP and Equifax confirmed that the vast majority of new claimants were white-collar professionals.

The S&P 500 fell 6% in the following week. Negative macroeconomic factors began to gain the upper hand in the game.

In a normal recession, unemployment is widespread across all groups, with blue-collar and white-collar workers sharing the burden roughly according to their respective proportions of total employment. The impact on consumption is also widespread, and because low-income groups have a higher marginal propensity to consume, this impact is quickly reflected in the data.

This round of unemployment is concentrated in the top 10% of the income distribution. They represent a relatively small percentage of total employment, but their contribution to consumer spending far exceeds their proportion. The top 10% of income earners in the US contribute over 50% of all consumer spending; the top 20% contribute approximately 65%. It is this group that buys houses, cars, vacations, restaurant meals, private school tuition, and home renovations. They are the foundation of demand in the entire discretionary consumer economy.

When these workers lose their jobs, or their wages are halved and they move into their current positions, the impact on consumption is enormous relative to the number of unemployed. A 2% drop in white-collar employment roughly corresponds to a 3-4% drop in disposable consumption. Unlike blue-collar unemployment (where factory workers stop consuming the following week after being laid off), the impact of white-collar unemployment is delayed but deeper—because these workers have a savings buffer that can sustain their consumption for several months before their behavior changes.

In Q2 of 2027, the economy plunged into recession. The National Bureau of Economic Research won't officially determine the start date for several months (as it always has), but the data is undeniable—we've recorded negative real GDP growth for two consecutive quarters. But this wasn't a "financial crisis"...at least not at that time.

Dominoes with related bets

Private lending has ballooned from less than $1 trillion in 2015 to over $2.5 trillion in 2026. A significant portion of this capital has been invested in software and technology deals, many of which are leveraged buyouts of SaaS companies valued on the assumption of perpetual revenue growth of more than ten percent.

These assumptions died sometime between the first intelligent agent programming demonstration and the software crash in Q1 2026, but book valuations don't seem to have realized it yet.

As public market SaaS companies have fallen to 5-8 times EBITDA, PE-backed software companies are still sitting on fund books, valued at acquisition prices based on revenue multiples that no longer exist. Managers are gradually lowering their valuations: 100 points, 92 points, 85 points, while publicly comparable companies are already saying: 50 points.

Moody's downgraded the ratings of 14 issuers' $18 billion in PE-backed software debt, citing "long-term revenue headwinds from AI-driven competitive disruption"; this is the largest single-sector rating action since the energy sector in 2015 | Moody's Investors Service, April 2027

Everyone remembers what happened after the downgrade. Industry veterans had already seen the script for the 2015 energy downgrade.

Software-backed loans are expected to default starting in Q3 2027. Private equity portfolio companies in the information services and consulting sectors are following suit. Several well-known SaaS companies have entered restructuring proceedings following multi-billion dollar leveraged buyouts.

Zendesk is that smoking gun.

Zendesk defaults on debt covenants as AI-driven customer service automation erodes ARR; $5 billion in direct lending financing rated 58 points; largest software default in private lending history | Financial Times, September 2027

In 2022, Hellman & Friedman and Permira took Zendesk private for $10.2 billion. The debt package consisted of $5 billion in direct lending, the largest ARR-backed financing in Zendesk's history, led by Blackstone, with Apollo, Blue Owl, and HPS also in the syndicate. This loan was structurally based on the assumption that Zendesk's annual recurring revenue would continue to circulate. The leverage of approximately 25 times EBITDA only makes sense under this assumption.

By mid-2027, this assumption will no longer hold true.

AI agents have been autonomously handling customer service for nearly a year. The category defined by Zendesk (tickets, routing, managing human customer service interactions) has long been replaced by systems that resolve issues directly without generating tickets. The annual revolving revenue on which that loan was based is no longer revolving; it's just revenue that hasn't left yet.

The largest ARR-backed loan in history has become the largest private lending software default in history. Every lending table is simultaneously asking the same question: Who else is hiding structural headwinds beneath a cyclical facade?

But here's one thing, at least initially, the mainstream consensus was right: this shouldn't have been fatal.

Private lending is not the banking industry of 2008. The explicit design of the entire structure is to avoid forced sell-offs. It's a closed-end instrument; capital is locked up. Limited partners (LPs) have committed to seven to ten years. There's no depositor run, no repurchase agreements can be withdrawn. Managers can sit back on the damaged assets, slowly dispose of them, and wait for recovery. It's painful, but controllable. This system is meant to be flexible, not broken.

Executives at Blackstone, KKR, and Apollo cited software exposures of 7-13% of their assets. Controllable. Every sell-side report and Twitter finance account says the same thing: private lending has perpetual capital. It can absorb losses that would destroy leveraged banks.

Permanent capital. This phrase appears in every earnings call and investor letter, intended as reassurance. It has become a mantra. And like most mantras, nobody really pays attention to the details. What does it actually mean…?

Over the past decade, large alternative asset management firms have acquired life insurance companies, transforming them into financing vehicles. Apollo acquired Athene; Brookfield acquired U.S. Equity Life; and KKR acquired Global Atlantic. The logic is simple and elegant: annuity deposits provide a stable, long-term liability base; managers invest these deposits in their own private lending, charging fees from both ends—the insurance company earns the interest spread, and the asset management company collects management fees. It's a perpetual motion machine that, under certain conditions, generates ever-increasing fees.

Private lending must be to the good money.

The losses hit balance sheets that specifically hold non-current assets to cover long-term liabilities. So-called "permanent capital"—the portion of capital that should make the system resilient—is not some abstract, patient institutional money, or sophisticated investors bearing mature risk. It's American household savings, the money of "ordinary people," in the form of annuities, invested in private equity-backed software and technology bonds that are now defaulting. And the locked-up capital that can't be withdrawn is the money of life insurance policyholders, where the rules are slightly different.

Compared to banking regulation, insurance regulators have traditionally been mild—even indifferent—but this is a wake-up call. Already uneasy about the concentration of private credit in life insurance companies, regulators have begun to lower the risk capital treatment of these assets. This forces insurers to either raise capital or sell assets, neither of which can be done on favorable terms given the already frozen market.

New York and Iowa regulators tighten capital treatment for certain privately rated credit held by life insurers; NAIC guidance is expected to increase RBC ratios and trigger more SVO reviews | Reuters, November 2027

When Moody's placed Athene's financial strength rating on a negative outlook, Apollo's stock price plummeted 22% in two trading days. Brookfield, KKR, and other companies followed suit.

Things are only getting more complicated. These companies have not only created a perpetual motion machine for insurance companies, but also constructed a sophisticated offshore structure designed to maximize returns through regulatory arbitrage. US insurance companies issue annuities and then transfer the risk to Bermuda or Cayman reinsurance companies, also owned by them. These offshore entities are subject to less stringent regulations and require less capital to hold the same assets. The offshore subsidiary raises capital externally through an offshore special purpose vehicle also controlled by the parent company, introducing new counterparties alongside the insurance company to jointly invest in private lending initiated by the same parent company's asset management division.

Rating agencies—some of which are themselves private equity-owned—are far from exemplary in terms of transparency (which is no surprise). The spiderweb of interconnected companies and balance sheets is staggeringly opaque. When underlying loans default, it's impossible to determine in real time who truly bears the losses.

The crash in November 2027 marked a shift in market perception from "it might just be a normal cyclical downturn" to something extremely unsettling. Federal Reserve Chairman Kevin Warsh called it "a long chain of related bets on white-collar productivity growth" at the Fed's emergency meeting in November.

It's important to understand that a crisis is never caused by the loss itself, but by acknowledging the loss. And there's another, much larger and more important area of ​​finance where we increasingly worry about that moment of acknowledgment.

Questions about mortgage loans

Zillow Home Price Index: San Francisco down 11% year-over-year, Seattle down 9%, Austin down 8%; Fannie Mae marks "rising early default rates" in zip codes with over 40% tech/finance workforce | Zillow / Fannie Mae, June 2028

This month, the Zillow Home Price Index fell 11% year-over-year in San Francisco, 9% in Seattle, and 8% in Austin. This isn't the only worrying headline. Last month, Fannie Mae flagged an increase in early default rates in zip codes with high concentrations of large loans—areas home to borrowers with credit scores of 780+, historically considered a "flawless" group.

The U.S. residential mortgage market is worth approximately $13 trillion. Mortgage underwriting is based on a fundamental assumption: that borrowers will maintain roughly stable employment at their current income level for the duration of the loan term. For most 30-year mortgages, this is thirty years.

The white-collar employment crisis, with its persistent shifts in income expectations, threatens this assumption. We are now forced to confront a question that seemed absurd three years ago: Are prime mortgages still good assets?

Every mortgage crisis in U.S. history has stemmed from one of three factors: excessive speculation (lending to people who cannot afford to buy a house, such as in 2008), interest rate shocks (rising interest rates making adjustable-rate mortgages unpayable, such as in the early 1980s), or localized economic shocks (the collapse of a single industry in a single region, such as the Texas oil crisis in the 1980s or the Michigan auto crisis in 2009).

This time, none of the above three criteria apply. The borrowers in question were not subprime borrowers. Their credit scores were 780. They made a 20% down payment. They had clean credit records, stable employment histories, and verified and recorded income at the time the loans were issued. They were the cornerstone of all the risk models in the financial system used to establish credit quality.

In 2008, loans were bad from day one. In 2028, loans will be good from day one. It's the world... that changed after the loans were issued. The future people borrowed for is a future they can no longer believe in.

In 2027, we identified early signs of hidden stress: HELOC withdrawals, 401k early withdrawals, and a surge in credit card debt, while mortgage payments remained on schedule. As jobs were lost, hiring froze, and bonuses were cut, the debt-to-income ratio of these high-achieving households doubled.

They could still afford their mortgages, but at the cost of ceasing all discretionary spending, depleting their savings, and postponing any home repairs or renovations. Technically, they weren't in default yet, but just one more shock would have them in trouble—and the trajectory of AI capabilities suggests that shock was on its way. Subsequently, defaults began to climb in San Francisco, Seattle, Manhattan, and Austin, even though the national average remained within historically normal ranges.

We are now in the most acute phase. When marginal buyers are healthy, price declines are manageable. But now, marginal buyers face the same income damage.

Despite growing concerns, we are not yet in the midst of a full-blown mortgage crisis. Default rates are rising, but remain well below 2008 levels. The real threat is the trajectory.

The intelligent substitution spiral now possesses two financial catalysts that are accelerating the downturn in the real economy.

Labor substitution, mortgage lending concerns, and private equity market volatility are all reinforcing each other. Traditional policy tools (interest rate cuts, quantitative easing) can address the problems of the financial engine, but they cannot reach the engine of the real economy—because the driving force of the real economy is not tight financial conditions, but rather AI making human intelligence less scarce and less valuable. Lowering interest rates to zero, buying up all mortgage-backed securities and leveraged buyout debt from defaulted software…

None of this changes the fact that a Claude agent can do the work of a product manager earning $180,000 a year for a cost of $200 per month.

If these concerns materialize, the mortgage market will crack in the second half of this year. In that scenario, we expect the current stock market decline to eventually be comparable to the global financial crisis (a 57% drop at its peak). This would send the S&P 500 down to around 3500 points—the level we last saw before the ChatGPT moment in November 2022.

It is clear that the income assumptions underpinning $13 trillion in residential mortgages have been structurally damaged. What remains unclear is whether policy can intervene in time before the mortgage market fully absorbs what this means. We are hopeful, but we cannot deny the reasons that discourage optimism.

A battle with time

The first negative feedback loop operates in the real economy: AI capabilities improve, wages shrink, consumption softens, profit margins are under pressure, businesses buy more capabilities, and capabilities improve again. It then becomes a financial problem: income damage impacts mortgage lending, banks suffer losses and tighten credit, the wealth effect cracks, and the negative feedback loop accelerates. Both of these are exacerbated by a clearly inadequate policy response—from a government that, frankly, appears completely bewildered.

This system was never designed to handle a crisis like this. The federal government's revenue base is essentially a tax on human time. People work, businesses pay wages, and the government takes a cut. Personal income tax and payroll tax form the backbone of revenue in normal years.

In Q1 2028, federal revenue is projected to be 12% lower than the CBO baseline. Payroll taxes are declining due to reduced employment and lower wages. Income taxes are declining due to a structural decrease in earned income. Productivity is soaring, but the gains are flowing to capital and computing power rather than labor.

The share of labor in GDP fell from 64% in 1974 to 56% in 2024, a slow decline over four decades, driven by globalization, automation, and the continued erosion of workers' bargaining power. In the four years since AI began its exponential leap, this share has fallen to 46%, the most dramatic drop on record.

The output is still there. But it no longer flows from households to businesses—meaning it no longer passes through the tax authorities. The circular flow is breaking down, and the government is expected to step in to fix it.

As with every recession, spending rises precisely when taxes fall. The difference this time is that spending pressures are not cyclical. Automatic stabilizers are designed for temporary unemployment, not for structural replacement. The benefits this system pays assume workers will be reabsorbed. Many will not, at least not back to near their previous wage levels. During the pandemic, the government readily accepted a 15% fiscal deficit, but that was understood as temporary. Those who need government support today are not defeated by a pandemic from which they will recover. They are replaced by a continuously advancing technology.

The government needs to transfer more money to households right now, when tax revenue is decreasing.

The United States will not default. It spends its own money and repays its borrowers in the same currency. But this pressure is already manifesting elsewhere. Year-to-date performance of municipal bonds has shown a worrying divergence. States without income taxes are doing relatively well, but general obligation municipal bonds issued by states that rely on income taxes (mostly blue states) are beginning to price in a certain degree of default risk. Politicians quickly realized that the debate over who should receive bailouts had split along partisan lines.

This administration deserves credit for recognizing the structural nature of the crisis early on and for beginning discussions on the "Economic Transition Act," which has been proposed by both parties: a framework that would provide direct transfer payments to unemployed workers through a combination of fiscal deficit spending and a tax on AI inference computing power.

The most radical proposal on the table goes even further. The "AI Co-Prosperity Act" would establish a public claim on the returns to smart infrastructure itself—something somewhere between sovereign wealth funds and AI production taxes—to fund household transfer payments in the form of dividends. Private sector lobbyists have already flooded the media with slippery slope warnings.

The politics surrounding these discussions have frustratingly headed towards a predictable end, escalating amid political theatrics and hardline confrontations. The right wing equates transfer payments and redistribution with Marxism, warning that taxing computing power would hand over global leadership to China. The left warns that tax laws drafted with the participation of incumbent interest groups are merely another form of regulatory capture. Fiscal hawks point to unsustainable deficits. Fiscal doves cite premature austerity following the GFC as a cautionary tale. This division will only widen as this year's presidential election approaches.

While politicians are arguing incessantly, the social structure is disintegrating faster than the legislative process can keep up.

The "Occupy Silicon Valley" movement has become a microcosm of broader discontent. Last month, protesters blocked the entrances to Anthropic and OpenAI's San Francisco headquarters for three weeks. Their numbers are growing, and the media attention these demonstrations are attracting has surpassed the unemployment data that triggered them.

It's hard to imagine the public hating anyone more than bankers after the global financial crisis, but AI labs are making a splash. And from the perspective of the masses, they have every reason. The speed at which founders and early investors have accumulated wealth makes the Gilded Age seem shabby. The benefits of this productivity boom have almost entirely flowed to the owners of computing power and the shareholders of the labs built on that power, amplifying inequality in the United States to an unprecedented degree.

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

The evolution of AI capabilities is outpacing any institution's ability to adapt. Policy responses are operating according to ideological rhythms, not the rhythms of reality. If governments cannot quickly reach a consensus on what the problem is, this negative feedback loop will write the next chapter for them.

The dissolution of the premium for human intelligence

Throughout modern economic history, human intelligence has been the scarce factor of production. Capital is abundant (or at least replicable). Natural resources are limited but substitutable. Technological progress is slow enough for humans to adapt. Intelligence—the ability to analyze, decide, create, persuade, and coordinate—is something that cannot be replicated on a large scale.

The inherent premium of human intelligence stems from its scarcity. Every institution in our economy—from the labor market to the mortgage market to tax laws—is designed for a world where this assumption holds true.

We are now experiencing the dissipation of this premium. Machine intelligence has become a competitive and rapidly improving alternative to human intelligence, covering an ever-widening range of tasks. The financial system, optimized for scarce human intelligence for decades, is being repriced. This repricing is painful, disorderly, and far from over.

But repricing does not equate to collapse.

The economy can find a new equilibrium. Getting there is one of the few things that only humans can still do. We need to do it right.

For the first time in history, the most productive assets in the economy have resulted in fewer jobs, not more. No framework works because no framework was designed for a world where scarce factors of production become abundant. So we must build a new framework. Whether we can build it in time is the only important question.

But you're not reading this in June 2028. You're reading this in February 2026.

The S&P 500 is near its all-time high. The negative feedback loop hasn't even begun. We're certain that some of these scenarios won't materialize. We're also certain that machine intelligence will continue to accelerate. The premium for human intelligence will narrow.

As investors, we still have time to assess how much of our portfolio is built on the assumption that we won't survive this decade. As a society, we still have time to take proactive steps.

The canary in the mine is still alive.