Author：区块律动

Editor's Note: As machine intelligence begins to become a large-scale substitute for human intelligence, we are no longer facing just a technological leap, but a systemic reassessment surrounding value distribution, asset pricing, and the premises of institutional design. For a long time, the operation of the modern economy has been built on a core assumption—that human intelligence is scarce. The labor market, the housing mortgage system, the expansion of private credit, and even the structure of the fiscal tax base all revolve around this premium. However, as AI capabilities continue to evolve, this premise is loosening. The premium of human intelligence is gradually being diluted, and its impact has spread from corporate profit models to financial balance sheets, and further extends to public finance and social structures. When the most productive assets begin to reduce employment rather than create it, an economic system centered on labor income is bound to face adjustment.

The author of this article, Alap Shah, is based in New York and has served as CEO of Littlebird since September 2024. A graduate of Harvard University with a degree in economics, he spent two years after graduation as an analyst at the top hedge fund Citadel LLC.

The following is the original text:

If our long-term bullish view on AI remains valid, could it actually mean that it's not good news for the overall economy?

The following content is not a prediction, but a scenario simulation. It is neither a deliberately panic-inducing bear market narrative nor a doomsday fantasy about AI. The sole purpose of this article is to attempt a systematic model of a possible path that has not been adequately discussed previously. This problem was first raised by our friend Alap Shah, and we jointly explored this idea through discussions. This article is written by us, while the other two were written by him and can be found separately.

Hopefully, this article can help readers better prepare for potential left-tail risks before AI gradually changes how the economy operates and even makes the structure itself increasingly counterintuitive.

The following is a macro memo written by CitriniResearch in June 2028, which attempts to trace and review the formation process of the global intelligence crisis and its chain reaction.

Macro Memo: The Economic Consequences of Excess Intelligence

CitriniResearch

(February 22, 2026) June 30, 2028

The unemployment rate released this morning was 10.2%, 0.3 percentage points higher than market expectations. As a result, the market fell by 2%, and the S&P 500 index has now retreated 38% from its October 2026 high.

Traders have barely reacted to this; just six months ago, such a high unemployment rate would have triggered a circuit breaker.

In just two years, the economy has evolved from one where "risks were manageable and the impact was limited to specific industries" into a system that no longer aligns with any of our past experiences. This quarterly macroeconomic memo attempts to reconstruct this evolutionary process and conduct a post-hoc systematic analysis of the economic structure before the crisis truly arrived.

Once upon a time, market sentiment was high. In October 2026, the S&P 500 approached 8,000 points, and the Nasdaq broke through 30,000 points. The first round of layoffs surrounding the replacement of human labor began in early 2026, and it did indeed achieve the effects expected by the capital market: improved profit margins, better-than-expected performance, and rising stock prices.

The company’s record profits were quickly reinvested in expanding AI computing power.

Macroeconomic data remains outwardly positive, with nominal GDP repeatedly recording mid-to-high single-digit annualized growth, productivity significantly improved, and real output per unit hour reaching its highest growth rate since the 1950s. All of this stems from AI agents that don't rest, don't take sick leave, and require no welfare benefits.

The wealth of those who own computing power has ballooned rapidly, while real wage growth has weakened significantly. Despite repeated official pronouncements of record productivity, an increasing number of white-collar jobs are being replaced by machines, forcing workers to seek lower-paying positions.

When signs of easing on the consumption side began to emerge, commentators put forward a new concept, "ghost GDP": outputs that are reflected in statistical reports but have not actually entered the real economic cycle.

AI is exceeding expectations across almost all technical metrics; the capital markets narrative revolves almost entirely around AI. The only discrepancy is that the economic structure itself has not benefited in tandem.

In hindsight, this logic is actually quite simple. If the output of a GPU cluster located in North Dakota is equivalent to the economic contribution of 10,000 white-collar workers in Midtown Manhattan, then its impact is more like an economic pandemic than a cure.

The velocity of money stagnated. The consumer-driven economy, centered on people and accounting for about 70% of GDP, contracted rapidly. Perhaps we could have realized this much earlier by simply asking the question: how much do machines spend on discretionary consumer goods?

The answer is obvious: zero.

Subsequently, the negative feedback loop begins to reinforce itself: AI capabilities improve → companies need fewer employees → white-collar layoffs increase → those being replaced cut spending → profit pressure forces companies to further increase their AI investment → AI capabilities continue to improve...

This is a cycle lacking a self-braking mechanism, a spiral process in which human intelligence is systematically replaced.

The income capacity and resulting willingness to consume among white-collar workers are being eroded structurally, and this income is the very foundation upon which the $13 trillion mortgage market is built. Underwriters are forced to reassess a long-taken-for-granted question: do so-called prime mortgages still offer a sufficient margin of safety?

Meanwhile, the absence of a true default cycle for 17 consecutive years has led to a large accumulation of private equity-backed software asset transactions in the private market. These transactions are almost all based on the same assumption: that ARR (Annual Recurring Revenue) will grow steadily and continuously in the long term and possess compounding properties.

However, the first wave of defaults triggered by AI disruption in mid-2027 directly undermined this premise.

If the impact had been limited to the software industry, the situation might still be manageable, but that is not the case in reality.

By the end of 2027, almost all business models built on the role of intermediaries began to come under pressure, and companies that profited from providing frictional intermediary services to humans experienced widespread collapse.

At a deeper level, the entire economic system is essentially a chain of bets highly correlated with the continuous improvement of white-collar productivity. The market crash in November 2027 was not the starting point of the shock, but merely the full acceleration of various existing negative feedback mechanisms.

The market has been waiting for nearly a year for a turning point where bad news turns into good news. While the government has begun discussing solutions, public confidence in its ability to implement effective relief is rapidly waning. Policy responses have always lagged behind economic realities, and at this stage, the lack of a systemic solution is itself pushing the deflationary spiral deeper.

How it all started

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

An experienced developer, using Claude Code or Codex, can replicate the core functionality of a mid-sized SaaS product within weeks. While it's difficult to fully cover all edge scenarios, its maturity is enough to make a CIO reviewing a $500,000 annual renewal contract seriously consider the question – "Why don't we do it ourselves?"

Since most companies' fiscal years coincide with the calendar year, their IT spending budgets for 2026 were finalized as early as the fourth quarter of 2025. At that time, "agent-based AI" was still just a concept.

Therefore, the mid-year review became the first real stress test, and for the first time, the procurement team re-evaluated existing spending decisions based on a full understanding of the true capabilities of these systems.

That summer, we interviewed a purchasing manager from a Fortune 500 company. He recalled a crucial budget negotiation: the sales team originally planned to use the same negotiation template as in previous years, namely a 5% annual price increase, plus a standard line like "Your team can't do without us anymore." But this purchasing manager frankly stated that he had already contacted OpenAI, considering having its frontline deployment engineers use AI tools to directly replace existing suppliers.

Ultimately, the contract was renewed at a 30% discount. In his view, this was already a relatively ideal outcome. Long-tail SaaS companies like Monday.com, Zapier, and Asana face a much more difficult situation.

Investors had already anticipated that the long tail of SaaS would be the first to be impacted. After all, they account for about one-third of enterprise technology stack spending, making them the most vulnerable.

The real blind spot is that the core software that is considered a system-level recording system was originally thought to be secure enough.

The reflexivity mechanism only truly came to light when ServiceNow released its Q3 2026 financial report:

ServiceNow’s new ACV growth slowed from 23% to 14%; it announced a 15% reduction in staff and launched a structural efficiency plan; its stock price fell 18%.

SaaS is not dead, and self-built systems still involve trade-offs between operational costs and complexity. However, the very fact that self-built systems have become a viable option has fundamentally changed the starting point for pricing negotiations.

More importantly, the competitive landscape has undergone a structural change. AI has significantly lowered the barriers to feature development and product iteration, causing differentiation to collapse rapidly. Established manufacturers are forced into price wars, fighting each other while also facing a new wave of challengers who have no historical cost burdens and are directly empowered by proxy programming capabilities.

It was only at this moment that the market truly realized the high degree of interconnectivity between these systems.

ServiceNow charges by the number of seats, so when one of its Fortune 500 clients lays off 15% of its staff, it means that 15% of its licenses will be cancelled simultaneously.

The same AI-driven layoff logic that drives improved customer profit margins is also eroding its own revenue base in an almost mechanical way. This company that sells workflow automation is ultimately disrupted by even more efficient workflow automation; and its only response is to lay off employees and reinvest the savings into the very technology that is disrupting it.

What else can we do? Stay put and slowly wait to die?

Thus, a most direct and ironic result emerged: the companies most threatened by AI became the most radical adopters of AI.

In hindsight, this seems logical; but at the time (at least to me), it wasn't. The traditional model of technological disruption usually involves established giants resisting new technologies, allowing more agile newcomers to erode their market share, and ultimately leading to a slow decline. Kodak, Blockbuster, and BlackBerry are all examples of this.

But 2026 will be different. Existing companies will not be choosing to resist, but rather they will be utterly powerless to resist. When stock prices fall by 40% to 60% and boards of directors demand clear countermeasures from management, these companies at the heart of the AI impact will essentially have only one path left: lay off employees, invest the savings in AI, and then use AI to maintain output at an even lower cost.

From the perspective of an individual company, such a decision is entirely rational; however, at the overall level, it has disastrous consequences. Every dollar saved in labor costs is converted into investment in strengthening AI capabilities, which in turn paves the way for the next round of layoffs.

Software, however, is merely the beginning.

While investors are still debating whether SaaS valuations have bottomed out, a more crucial change has occurred: this reflexive logic has already spread beyond the software industry. The same logic that underpinned ServiceNow's layoffs applies equally to all companies whose core business revolves around white-collar costs.

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 unconsciously, without even being aware of the concept, just as most people didn't understand what cloud computing was back then, yet were already accustomed to watching videos via streaming media. In the eyes of ordinary users, it's more like a low-level function such as auto-completion or spell checking, a capability that a device should naturally possess.

Qwen's open-source agent-based shopping assistant has become a key catalyst for AI taking over consumer decision-making. Within just a few weeks, almost all mainstream AI assistants have embedded different forms of agent-based e-commerce functionality. The maturity of the distillation model allows these agents to run directly on terminal devices such as mobile phones and laptops, no longer relying entirely on cloud computing, significantly reducing the marginal cost of inference.

What should truly alarm investors is that these agents don't wait for explicit user instructions; they run continuously in the background according to preset preferences. Consumption is no longer a series of discrete choices made one by one by humans, but has transformed into a 24/7 automatically optimized process, running continuously for every connected consumer. By March 2027, the average individual in the United States was consuming 400,000 tokens per day, a tenfold increase from the end of 2026.

And the next link in this chain has begun to loosen.

Intermediation layer

Over the past fifty years, the US economy has developed a massive rent-seeking structure on top of human limitations. Decision-making takes time, patience is limited, brand familiarity often replaces detailed comparison, and most people are willing to accept less-than-ideal prices to avoid clicking through many pages. Trillions of dollars in corporate value are built on the premise that these behavioral frictions have existed for a long time.

The initial changes seemed insignificant; agents began to smooth over friction. Subscription services that automatically renewed after months of inactivity, and pricing models that quietly increased after the trial period, were all redefined as negotiable terms. Customer Lifetime Value (LTV), the core metric underpinning the subscription economy, began to show a substantial decline.

Consumer agents have gradually rewritten the operating logic of almost all consumer transactions. Before buying a box of protein bars, it's difficult for an individual to have the time and energy to compare prices across five different platforms, but a machine can.

Travel booking platforms were the first to be impacted because their business logic is highly standardized. By the fourth quarter of 2026, AI agents were able to create complete itinerary plans faster and at a lower cost, covering airfares, hotels, ground transportation, points optimization, budget constraints, and refund and cancellation policies, with overall efficiency surpassing that of traditional platforms.

Insurance renewals were no exception. The business model that relied on policyholders' inertia to maintain profits was quickly undermined by agents who automatically compared prices annually: the 15%–20% premium from passive renewals disappeared almost in a short period of time.

Financial advisors, tax services, routine legal matters… any industry whose value proposition is based on handling complex and tedious tasks for clients is impacted. Because for agents, the concept of "tedious" doesn't exist.

Even those areas considered to be protected by interpersonal value were not spared.

The real estate industry has long relied on information asymmetry between buyers and sellers to maintain a commission structure of 5%–6%. When AI agents access MLS data and can instantly access decades of transaction records, this knowledge advantage is rapidly replicated.

A sell-side report in March 2027 described this phenomenon as a war between agents. The median buyer commission in major cities has been squeezed from 2.5%–3% to below 1%, and more and more transactions are being completed without any human brokers involved on the buyer's side.

We overestimate the value of interpersonal relationships. Many so-called relationships are nothing more than friction disguised as friendship.

This is just the beginning of the collapse of the intermediary layer. Successful companies have invested billions of dollars to build moats by leveraging consumer behavioral biases and psychological inertia, but these mechanisms quickly fail in the face of machines.

The machine only optimizes price and match. It doesn't care about your favorite app, won't be attracted by a fancy checkout page, won't choose the most convenient option because of fatigue, and won't repeatedly place orders on the same platform out of habit.

What was destroyed was a special kind of moat: the habitual intermediary.

DoorDash serves as a prime example. Proxy programming significantly lowered the barrier to entry for food delivery platforms, allowing a skilled developer to deploy a fully functional competitor within weeks. A plethora of new platforms emerged, rapidly attracting drivers by directly allocating 90-95% of delivery fees to them. Multi-platform management panels enabled gig workers to connect to 20-30 platforms simultaneously, virtually eliminating the previous lock-in effect. The market fragmented rapidly, squeezing profit margins to near zero.

The proxy simultaneously accelerated the collapse of both ends: it both fostered competitors and prioritized their use. DoorDash's moat is essentially built on a simple premise: "You're hungry, you're too lazy to compare prices, this app is right on the homepage."

The agent doesn't have a "homepage"; it searches DoorDash, Uber Eats, restaurant websites, and dozens of new platforms simultaneously, always choosing the option with the lowest cost and fastest delivery.

Habitual loyalty does not exist for machines.

Ironically, in this chain reaction, this may be the only time that the agents helped the white-collar workers who were about to be replaced. When they became delivery drivers, at least half of their income was no longer taken by the platform. However, this kindness brought by technology did not last long; with the popularization of autonomous driving, the situation quickly reversed.

When agents begin to control the transactions themselves, they continue to look for even greater opportunities for optimization.

Simple price comparison and aggregation have their limits. To continuously reduce costs for users, especially after agents begin trading with each other, the most direct way is to eliminate transaction fees. In machine-to-machine transaction scenarios, the 2%–3% bank card exchange fee naturally becomes the most prominent target.

Agents began searching for faster and cheaper settlement paths than traditional card organizations. Most ultimately opted to use stablecoins for payments via Solana or Ethereum Layer 2 networks, with settlements being almost instantaneous and transaction costs amounting to only a few cents.

Mastercard's Q1 2027 earnings: Net revenue grew 6% year-over-year; transaction volume growth slowed to 3.4% from 5.9% in the previous quarter; management mentioned "agent-led price optimization" and "pressure on discretionary spending categories." —Bloomberg, April 29, 2027

This financial report became an irreversible turning point.

The shift in agency-based commerce from product-level innovation to disruption at the settlement infrastructure level. MA fell 9% the following day, and Visa also came under pressure, but its decline was less severe due to its earlier investment in stablecoin infrastructure.

The agency-based business model, which bypasses the exchange fee settlement path, poses a more severe challenge to banks whose core business is bank card business, as well as single-business card issuers. These institutions have long collected the majority of their revenue from the 2%–3% exchange fee and built a complete business segment around points and rewards programs supported by merchant subsidies.

American Express faced the most concentrated pressure. On the one hand, the contraction of white-collar employment continued to weaken its high-value customer base; on the other hand, the bypassing of exchange fee settlement by agents directly shook its core revenue model. In the following weeks, the stock prices of Synchrony (SYF US), Capital One (COF US), and Discover (DFS US) also fell by more than 10%.

Their moat is essentially built on friction, and friction is rapidly dwindling to zero.

From industry risk to systemic risk

Throughout 2026, the market consistently viewed the negative impacts of AI as an industry-wide shock. The software and consulting sectors were hit hardest, with payment systems and other fee structures beginning to falter, but the broader macroeconomy appeared robust. While the labor market cooled, it did not experience an out-of-control decline. The prevailing consensus was that creative destruction is an inevitable phase in any technological innovation cycle: it will be painful in some areas, but the overall net benefits of AI will ultimately outweigh its negative impacts.

In our January 2027 macro memo, we pointed out that this is a flawed framework; the U.S. is essentially a white-collar-dominated service economy. White-collar workers account for 50% of total employment and contribute approximately 75% of discretionary consumer spending. The businesses and jobs that AI is eroding are not peripheral parts of the U.S. economy; they are the U.S. economy itself.

"Technological innovation destroys jobs while creating more jobs" was the most popular and persuasive counter-argument at the time.

It exists because it has almost never failed in the past two centuries. Even if we cannot clearly imagine the specific form of future jobs, they always appear as expected. ATMs reduced the operating costs of bank branches, yet banks opened more branches, and teller positions continued to grow over the next two decades. The internet disrupted travel agencies, yellow pages, and brick-and-mortar retail, but it also spawned entirely new industries and created a large number of new jobs.

But one premise has never been broken: all these new tasks still need to be done by humans.

AI is changing this premise. Today, AI has become a form of general intelligence and is rapidly advancing in tasks where humans would otherwise be redeployed. Replaced programmers cannot simply shift to AI management, because AI itself already possesses management capabilities.

Today, AI agents are capable of independently undertaking research and development tasks that last for weeks. Although business school professors still try to fit these data with new S-curves every year, the exponential growth has long exceeded our existing understanding of the boundaries of possibility.

They write almost all the code, and the best-performing agents are far smarter than almost any human in almost everything. And they're getting cheaper and cheaper.

AI has indeed created new jobs: prompt word engineers, AI security researchers, infrastructure technicians. Humans remain in the cycle, coordinating at the highest levels or controlling taste. But for every new job created, dozens of old ones are often rendered obsolete; and the salary level of the new job is only a fraction of that of the jobs it replaces.

US JOLTS data: Job openings fall below 5.5 million; the unemployment-to-job-opening ratio rises to approximately 1.7, the highest level since August 2020 – Bloomberg, October 2026

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

INDEED: Job postings in software, finance, and consulting industries have dropped significantly as the "Productivity Initiative" spreads—Indeed Hiring Lab, November–December 2026

White-collar job vacancies are shrinking rapidly, while blue-collar vacancies (construction, healthcare, skilled workers) remain relatively stable. The job losses are concentrated in positions involving memo writing, budget approval, and keeping the economy running at a mid-level (surprisingly, we're still in those roles). However, real wage growth for both groups has been negative for most of the year and continues to decline.

The stock market reacted far less to the JOLTS data than to another piece of news; they were more concerned about the fact that GE Vernova had sold all its turbine production capacity until 2040. The market was caught in a tug-of-war between negative macroeconomic data and positive AI infrastructure expansion, resulting in overall sideways movement.

However, the bond market (always smarter than the stock market, or at least less romantic) began pricing in the risk of a decline in consumption. The 10-year Treasury yield then began to fall from 4.3% to 3.2% over the next four months. Nevertheless, the overall unemployment rate did not rise out of control, and subtle structural differences were still being overlooked by some.

In normal economic recessions, problems often have self-correcting mechanisms. Over-construction leads to a slowdown in building, which in turn leads to lower interest rates, stimulating new construction. Excess inventory leads to destocking, which in turn triggers renewed buying. Cyclical mechanisms contain the seeds of their own recovery.

But this time, the root of the cycle is not cyclical.

AI is becoming better and cheaper. Companies first lay off employees, then invest the savings in more AI capabilities, thus creating the conditions for further layoffs; the laid-off employees subsequently reduce their spending; consumer-facing businesses experience declining sales and profit pressure, forcing them to continue increasing their investment in AI to maintain profit margins. Thus, AI becomes stronger and cheaper once again.

This is a negative feedback loop without a natural brake.

Intuitively, one might assume that the decline in aggregate demand would eventually slow down the pace of AI development. However, this is not the case, because this is not capital expenditure (CapEx) that relies on hyperscale computing centers in the traditional sense, but rather a type of operating expenditure (a structural alternative to OpEx).

A company that previously spent $100 million annually on human resources and only $5 million on AI might now see its human resource spending drop to $70 million, while its AI budget increases to $20 million. AI investment has multiplied, but this growth isn't due to expansion; rather, it's occurring amidst a decline in overall operating costs. In other words, every company's AI budget is increasing, while total spending is shrinking.

Ironically, even as AI is reshaping and weakening the economic systems it is embedded in, the AI infrastructure complex itself remains robust.

Nvidia (NVDA) continues to deliver record revenue, TSMC (TSM) maintains capacity utilization above 95%, and hyperscale cloud providers are still investing $150 billion to $200 billion per quarter in data center construction. Economies that best exemplify this trend, such as Taiwan and South Korea, have also significantly outperformed the broader market in their capital markets.

India presents a completely different picture. Its IT services sector exports over $200 billion annually, a core source of India's current account surplus and a key pillar in offsetting its long-standing merchandise trade deficit. This entire model is built on a simple comparative advantage: Indian developers cost only a fraction of what their American counterparts do.

However, the marginal cost of AI programming agents has collapsed to almost nothing more than electricity costs. Entering 2027, contract cancellations for Tata Consultancy Services (TCS), Infosys, and Wipro accelerated significantly. As the service trade surplus supporting India's external accounts rapidly evaporated, the rupee depreciated by 18% against the dollar in four months. By the first quarter of 2028, the International Monetary Fund (IMF) had begun initial contact with New Delhi.

The engines driving disruption are getting stronger every quarter, meaning the pace of disruption is accelerating. There is no natural bottom in the labor market.

In the United States, the question is no longer when the AI infrastructure bubble will burst. What's truly being repeatedly asked is: where will an economy built on consumer credit go when consumers are systematically replaced by machines?

Intelligent replacement spiral

2027 will be a year where the macroeconomic narrative is no longer subtle. The transmission mechanisms of those disconnected but clearly negative trends over the past 12 months will become apparent. You don't need to look at the Bureau of Labor Statistics (BLS) data; just attend a friend's dinner party.

The white-collar workers who were replaced didn't sit idle; they were downgraded. Many took on lower-paying service sector and gig economy jobs, which increased the labor supply in these areas and depressed wages there.

A friend of ours was a senior product manager at Salesforce in 2025. She had the 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 started driving for Uber, but her income dropped to $45,000.

The focus is not on individual stories, but on the second-order mathematical effect, multiplying this dynamic by the hundreds of thousands of workers in each major metropolitan area. The influx of surplus labor into the service sector and the gig economy has driven down the wages of already struggling existing workers, and industry-specific disruption has escalated into wage compression across the entire economy.

The remaining pool of human-centric labor is yet to be corrected, and it is happening right now as we write these words. This is because automated delivery and self-driving cars are sweeping through the gig economy, which absorbed the first wave of displaced workers.

By February 2027, it will be clear that professionals still employed will be spending in a manner consistent with their potential to be laid off. They will be working twice as hard (primarily with the help of AI) simply to avoid being laid off, with hopes of promotion or a raise dashed. Savings rates will rise slightly, while spending will remain weak.

The most dangerous part is the lag. High-income earners use their above-average savings to maintain the appearance of normalcy for two to three quarters. Hard data doesn't confirm the problem until it becomes old news in the real economy. Then, data that shatters this illusion is released.

U.S. initial jobless claims surge to 487,000, highest level since April 2020 | Labor Department, 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 applicants are white-collar professionals.

The S&P 500 fell 6% in the following week, as negative macroeconomic factors began to gain the upper hand in the tug-of-war.

In a typical economic recession, unemployment is widespread. The suffering of blue-collar and white-collar workers is roughly proportional to their proportion of total employment. The impact on consumption is also widespread and will quickly become apparent in the data, as low-income workers have a higher marginal propensity to consume.

During this cycle, unemployment is concentrated in the top ten percentiles of income distribution. These represent a relatively small percentage of total employment, but they drive a disproportionate share of consumer spending. The top 10% of earners account for more than half of all consumer spending in the United States. The top 20% account for approximately 65%. These individuals buy homes, cars, go on vacations, eat out, pay for private school tuition, and renovate their homes.

They form the basis of demand for the entire non-essential consumer goods economy.

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

By the second quarter of 2027, the economy had entered a recession. The National Bureau of Economic Research didn't officially determine the start date of the recession until several months later (as is their usual practice), but the data was clear—we had already seen two consecutive quarters of negative real GDP growth. But this wasn't a "financial crisis" yet…it was temporary.

Related betting dominoes

The size of private lending has expanded rapidly from less than $1 trillion in 2015 to over $2.5 trillion by 2026. A significant portion of this capital has been invested in software and technology deals, particularly leveraged buyouts of SaaS companies. The valuations of these deals are generally based on a key assumption: that revenue will maintain mid-to-high ten-digit growth over the long term.

These assumptions were effectively proven wrong between the first demonstration of proxy programming capabilities and the software stock crash in the first quarter of 2026. The problem is that asset valuations don't seem to have realized this.

While many publicly traded SaaS companies have seen their transaction multiples fall to 5–8 times EBITDA, private equity-backed software companies on their balance sheets are still maintaining acquisition valuations based on "revenue multiples," which have long since disappeared. Management is choosing to slowly reduce book value: from 100 cents to 92 or 85; meanwhile, comparable companies in the public market are already priced at 50 cents.

Moody's downgraded the ratings of 14 issuers' private equity-backed software debt totaling $18 billion, citing "long-term revenue headwinds from AI-driven competitive disruption"; this is the largest single-sector downgrade since the 2015 energy crisis. —Moody's Investors Service, April 2027

Everyone remembers what happened after the rating downgrade. For industry veterans who experienced the energy sector downgrade wave of 2015, this scenario is not unfamiliar.

By the third quarter of 2027, loans backed by software assets began to default; PE portfolio companies in the information services and consulting sector followed suit; and several leveraged buyout deals involving well-known SaaS companies, each worth billions of dollars, entered restructuring proceedings.

Zendesk is irrefutable evidence.

ZENDESK failed to meet its debt covenants as AI-driven customer service automation eroded its ARR (Annual Recurring Revenue); its $5 billion direct lending facility was flagged at 58 cents; marking the largest private lending software default in history. — Financial Times, September 2027

In 2022, Hellman & Friedman and Permira took Zendesk private for $10.2 billion. Its financing structure included a $5 billion direct loan, which was the largest ARR-backed credit facility in history at the time, led by Blackstone Group, with Apollo, Blue Owl and HPS also in the loan syndicate.

The core structural assumptions of this loan are very clear: it assumes that Zendesk's annual recurring revenue (ARR) will remain constant. Only under this premise is the entire capital structure reasonable, with a leverage ratio of approximately 25 times EBITDA.

However, by mid-2027, this premise will no longer exist.

Over the past six months, AI agents have begun to handle customer service autonomously. The categories defined by Zendesk (tickets, routing, managing human customer service interactions) are being replaced by systems that can resolve issues directly without generating tickets. The "annual recurring revenue" that forms the basis of loan underwriting is no longer recurring; it's simply revenue that hasn't yet been lost.

Thus, the largest ARR-backed loan in history ultimately became the largest private lending software default case in history. Almost simultaneously, every lending platform was repeatedly asking the same question: Who else misjudged long-term structural headwinds as cyclical fluctuations that could be weathered?

But this was one of the few points that held true in the initial market consensus (at least at the beginning), and this was a shock that should have been bearable and even weathered.

Private lending is not the banking system of 2008. Its institutional design was intended to avoid the chain reaction of forced sell-offs. These funds are mostly closed-end structures, with capital locked up for long periods, and limited partners typically committing for seven to ten years. There is no risk of depositor runs or the withdrawal of repurchase financing. Managers can hold onto damaged assets, gradually restructure them, and wait for recovery. The process may be painful, but it is theoretically controllable.

Executives from Blackstone, KKR, and Apollo have all emphasized that their software asset exposure accounts for only 7%–13% of total assets, and the risk is manageable. Sell-side reports and credit opinion leaders on financial social media have also repeatedly conveyed the same argument: private credit has perpetual capital and can absorb losses that would otherwise cripple highly leveraged banks.

"Permanent capital".

This term appears frequently in reassuring earnings calls and investor letters, almost becoming a mantra. But like most mantras, few people actually try to decipher its meaning.

Over the past decade, large alternative asset management firms have acquired life insurance companies and transformed them into financing platforms. Examples include Apollo's acquisition of Athene, Brookfield's acquisition of American Equity, and KKR's acquisition of Global Atlantic.

The logic seems elegant: annuity deposits provide a stable, long-term source of liabilities; the manager invests these funds in private credit assets they themselves have initiated, while earning interest spreads on the insurance side and collecting management fees on the asset management side, forming a "fee-cumulative" revenue structure. As long as one premise holds true, this mechanism works well.

The premise is that the principal of private credit assets must be safe.

When losses actually materialize, they impact the balance sheets of a class of entities that use long-term liabilities to hedge against non-current assets.

The so-called "permanent capital" is not patient institutional money in the abstract sense, nor is it a group of sophisticated investors willing to take on complex risks. It is the savings of American households, the funds of Main Street, existing in the form of annuities, invested in the private equity-backed software and technology debt that is now beginning to default.

The locked-up capital that cannot be withdrawn is actually the money of life insurance policyholders. However, the rules are different in this area.

Compared to banking regulation, insurance regulators have long been relatively lenient, even somewhat complacent. But this time, it serves as a real wake-up call. Regulators, already concerned about the concentration of private credit in life insurance companies, have begun to lower the risk capital requirements for these assets. This forces insurance companies to either raise capital or sell assets, but in a market environment that is already nearing a freeze, neither option is likely to be fulfilled at a reasonable price.

New York and Iowa regulators announced stricter capital requirements for certain privately rated credit held by life insurance companies; NAIC guidance is expected to increase the RBC factor and trigger additional scrutiny. — Reuters, November 2027

When Moody's downgraded Athene's financial strength rating outlook to negative, Apollo's stock price fell 22% in two trading days. Other institutions such as Bofeng and KKR subsequently came under pressure.

The complexity doesn't stop there. These institutions not only created perpetual motion machines of insurance but also constructed sophisticated offshore structures to enhance returns through regulatory arbitrage. After underwriting annuities, US insurance companies reinsured the risks to their controlled subsidiaries in Bermuda or the Cayman Islands, jurisdictions with more lenient regulations that allowed for lower capital requirements on similar assets. This reinsurance entity then introduced external capital through offshore special purpose entities, forming a new layer of counterparties, and jointly invested with the insurance company in private credit assets initiated by the same parent company's asset management division.

Rating agencies, some of which are themselves privately owned, have always been anything but transparent, which is hardly surprising. The intricate web of different companies and balance sheets creates a staggering level of opacity. In the event of default on underlying loans, it is virtually impossible to determine who ultimately bears the losses in real time.

The market crash in November 2027 marked a turning point in market perception.

What was initially seen as a routine cyclical pullback has evolved into a deeper and more unsettling structural problem. Federal Reserve Chairman Kevin Warsh described it at an emergency meeting of the Federal Open Market Committee that month as "a domino-like, interdependent betting structure built on expectations of white-collar productivity growth."

In fact, what triggers a crisis is never the loss itself, but the confirmation and acknowledgment of that loss. And within the financial system, there exists an area that is even larger and more important, yet increasingly worrying as to whether that "confirmation" is truly achieved.

Questions about mortgages

The ZILLOW Home Value Index shows a year-over-year decline of 11% in San Francisco, 9% in Seattle, and 8% in Austin; Fannie Mae has identified zip codes with over 40% of their workforce in the tech/finance sector experiencing "increased early delinquency rates." —Zillow / Fannie Mae, June 2028

This month, the Zillow Home Price Index showed that San Francisco home prices fell 11% year-over-year, Seattle fell 9%, and Austin fell 8%. This isn't the only worrying sign. Last month, Fannie Mae pointed out that some zip code areas with high-value mortgages saw higher early default rates. Borrowers in these areas generally have credit scores above 780, long considered "bulletproof" high-quality borrowers.

The U.S. residential mortgage market is approximately $13 trillion. Mortgage underwriting is based on the fundamental assumption that borrowers will largely maintain their current employment and income levels over the loan term. For most mortgages, this assumption is valid for up to 30 years.

The white-collar employment crisis is shaking this fundamental assumption through the continuous downward revision of income expectations. A question that seemed almost absurd three years ago is now unavoidable: are prime mortgages really as secure as cash?

Looking back at every mortgage crisis in U.S. history, the causes can be categorized into three types: first, excessive speculation (lending to people who simply cannot afford to buy a house, such as in 2008); second, interest rate shocks (rising interest rates make floating-rate mortgages unaffordable, such as in the early 1980s); and third, localized economic shocks (the collapse of a single industry in a single region, such as the Texas oil industry in the 1980s and the Michigan auto industry in 2009).

But this time, none of these three criteria apply. These borrowers are not subprime customers, but rather the gold standard group with an FICO score of 780; they make a 20% down payment, have clean credit records, stable employment, and their income is rigorously verified and fully documented at the time of lending. They are precisely the default credit cornerstone of all risk models in the financial system.

The problem in 2008 was that the loans were bad from the start. The difference in 2028 is that the loans were good from the start. It's just that the world subtly changed after the loans were issued.

People are taking out loans to bet on a future they can no longer believe in or afford.

As early as 2027, we observed some hidden pressures beginning to emerge: increased home equity loan withdrawals, rising 401(k) early withdrawals, and rapidly increasing credit card balances, while mortgage repayments remained normal. With layoffs, hiring freezes, and reduced bonuses, these families, who were originally considered high-quality borrowers, saw their debt-to-income ratios nearly double.

They can still make their mortgage payments on time, but at the cost of drastically cutting discretionary spending, constantly depleting their savings, and postponing all home maintenance and improvement plans. On paper, their mortgages are still on track; but in reality, they are on the verge of defaulting with just one additional shock. And the ever-evolving path of AI capabilities means that such a shock is not far off.

Subsequently, we saw mortgage default rates begin to rise significantly in cities with high concentrations of technology and finance, such as San Francisco, Seattle, Manhattan, and Austin, although the national average remained within the historical range.

We are entering a most sensitive phase. When marginal buyers (those who might take over the property) are financially sound, the market can absorb the decline in house prices; but now, marginal buyers themselves are experiencing the same pressure of income loss.

Risks are building, but have not yet escalated into a full-blown mortgage crisis. Default rates are indeed rising, but remain far below the 2008 peak. What is truly alarming is not the current level, but the trajectory it is taking.

Today, the intelligent substitution spiral has gained two more financial accelerators that directly impact the downturn in the real economy.

Labor substitution, mortgage concerns, and turmoil in the private equity market.

These three factors reinforce and amplify each other. While traditional policy tools (interest rate cuts, quantitative easing, QE) may be able to offset the pressure on the financial system, they cannot solve the problems of the real economy's engine because the problem does not stem from excessively tight financial conditions.

The engine of the real economy is being driven by another force. AI is making human intelligence neither scarce nor expensive. You can lower interest rates to zero and buy all mortgage-backed securities on the market, even taking over all defaulted software LBO debt…

But this doesn't change the fact that a Claude agent can do the work of a product manager with an annual salary of $180,000 for a monthly cost of $200.

If these fears materialize, the mortgage market will collapse in the second half of this year. In that scenario, we expect the current stock market pullback to eventually approach the magnitude of the global financial crisis (a drop of approximately 57% from peak to trough), which would bring the S&P 500 back to around 3500 points—the level last seen in November 2022, before the ChatGPT moment.

What is certain is that the income assumptions supporting $13 trillion in residential mortgage lending have been structurally undermined. What is uncertain is whether policy intervention will be timely before the mortgage market fully absorbs this reality.

We still hold onto hope, but we cannot deny that reasons for pessimism are accumulating.

Race against time

The first negative feedback loop occurs in the real economy: as AI capabilities improve, the scale of employment shrinks, consumption weakens, profit margins come under pressure, and enterprises increase their investment in AI, further enhancing AI capabilities.

Subsequently, this mechanism spread to the financial system, with income losses impacting mortgage performance, bank asset quality deteriorating, credit tightening, the wealth effect weakening, and the feedback loop accelerating. Both of these factors were further amplified by the government's incompetence and inadequate response to the crisis.

Our system was never designed to handle such a crisis from the outset. The federal government's revenue base is essentially a taxation mechanism on human time: individuals contribute labor, businesses pay compensation, and the government extracts taxes. In normal years, personal income tax and payroll tax form the backbone of government revenue.

However, as of the first quarter of this year, federal revenue was 12% lower than the Congressional Budget Office's baseline forecast. The decline in payroll taxes stemmed from a continued decrease in the number of people employed at existing wage levels; the weakening income tax reflected a structural suppression of residents' actual income. While productivity is indeed rising rapidly, the incremental gains are not flowing to workers but are being absorbed by capital and computing power.

Labor income as a percentage of GDP has fallen from 64% in 1974 to 56% in 2024, a slow decline that has lasted for four decades, driven by globalization, automation, and the long-term weakening of labor's bargaining power. In just four years after AI began its exponential leap forward, this percentage plummeted further to 46%, marking the largest drop on record.

Output hasn't disappeared, but it no longer circulates back to businesses through households, meaning it no longer passes through the IRS. The closed loop of the economic cycle is breaking, yet the market and society still expect the government to step in and mend this structural crack.

As with every previous economic downturn, fiscal spending has increased while fiscal revenue has decreased. However, the difference this time is that the spending pressure is not cyclical, but structural.

Automatic stabilizers were originally designed to address short-term unemployment shocks, rather than long-term, irreversible structural replacements. The system pays benefits on the premise that workers will eventually be reintegrated into the labor market.

But reality is rewriting this assumption; a significant portion of the workforce will not return to their jobs, or at least not at near-previous salary levels. During the COVID-19 pandemic, the government readily accepted a 15% fiscal deficit, widely viewed as a temporary shock.

Today, those who need government support are not facing a public health crisis that will eventually pass, but rather being replaced by a continuously evolving, irreversible technology.

As a result, the fiscal system is facing a sharp and unprecedented structural contradiction: while it must transfer more funds to households, the government is collecting less tax revenue from these households.

The United States will not default. It prints the money it uses to consume, and the same money it uses to repay borrowers. But pressure is beginning to emerge in other areas. This year, the municipal bond market has seen a worrying divergence. States without income taxes have generally performed well; however, general liability bonds (GO munis) issued by states heavily reliant on income tax revenue (mostly blue states) are beginning to be priced into a degree of default risk by the market. Politicians quickly realized this, and the debate over who should be bailed out quickly escalated into a partisan battle.

It is commendable that the current government recognized the structural nature of the crisis relatively early and began to push forward a series of bipartisan proposals, collectively known as the Transition Economy Act. Its core idea is to provide direct transfer payments to displaced workers by expanding the fiscal deficit and adding a proposed AI inference computing power tax.

More radical proposals have taken this a step further. The Shared AI Prosperity Act advocates for a public claim on the revenue generated from intelligent infrastructure itself, somewhere between a sovereign wealth fund and a royalty for AI outputs. The resulting dividends would be used to continuously transfer income to households.

Unsurprisingly, lobbying efforts from the private sector quickly dominated media headlines, warning of a dangerous slippery slope.

The political maneuvering behind the policy discussions is extremely clichéd, rife with grandstanding and brinkmanship. The right wing denounces transfer payments and redistribution as Marxism and warns that taxing computing power is tantamount to handing over technological leadership to China; the left wing warns that tax laws drafted with the help of vested interests will only result in disguised regulatory capture; fiscal hawks emphasize that deficits are unsustainable; and doves repeatedly cite the premature austerity policies implemented after the Global Financial Crisis (GFC) as a cautionary tale.

As this year's presidential election approaches, these divisions will only be amplified.

While politicians are locked in their disputes, the tearing apart of social structures is happening far faster than the legislative process itself. The Occupy Silicon Valley movement is a concentrated manifestation of this widespread discontent. Last month, protesters blocked the entrances to Anthropic and OpenAI's San Francisco offices for three consecutive weeks. Participation continues to grow, and the media attention these protests are receiving has even surpassed the unemployment data that sparked the protests.

It's hard to imagine anyone less hated by the public after the global financial crisis than bankers, but AI labs are rapidly approaching that position. And from the public's perspective, this hatred is understandable.

The founders and early investors of these companies amassed wealth at a pace that seemed gentle even in the Gilded Age. The benefits of this productivity explosion flowed almost entirely to the owners of computing power and the shareholders of the labs that relied on it, amplifying inequality in the United States to unprecedented levels.

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

The pace of AI capability evolution far outpaces the rhythm of institutional adjustments; policy responses continue to move at the speed of ideology, rather than the speed of reality. If governments cannot quickly reach a consensus on what the problem actually is, then the aforementioned feedback loop will write the next chapter for them.

The End of Smart Premium

Throughout modern economic history, human intelligence has consistently been the scarcest input factor. Capital is abundant (or at least replicable), natural resources, though finite, are often replaceable, and technological progress is slow enough for humanity to adapt. Intelligence—the ability to analyze, decide, create, persuade, and coordinate—is the only thing that cannot be replicated on a large scale.

It is precisely because of this scarcity that human intelligence naturally commands a premium. From the labor market to the housing mortgage system and tax design, almost all core economic institutions are built upon this premise.

We are now witnessing the end of this premium. In a growing number of tasks, machine intelligence is becoming a competent replacement for human intelligence, and it continues to evolve rapidly. A financial system that has operated and been continuously optimized for decades under the assumption of human intelligence scarcity is being forced to repric. This process is destined to be painful, chaotic, and far from over.

But repricing is not the same as collapse.

The economy may still find a new equilibrium. Reaching this equilibrium is one of the few tasks that can still only be accomplished by humankind, and we must do it right.

For the first time in history, the most productive asset in the economy has not created more jobs, but rather reduced them. Existing theoretical frameworks are no longer fully applicable because they were never designed to cope with a world where a previously scarce factor has suddenly become abundant. Therefore, we must construct new frameworks.

The only truly important question is whether we have enough time.

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

The S&P 500 is near all-time 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, and the premium for human intelligence will continue to narrow.

As investors, we still have time to examine how much of our portfolios are built on the assumption that we won't survive this decade. As society, we also still have time to choose to proactively shape the future, rather than passively accept the consequences.

The canary in the mine is still alive.

Original authors: Citrini and Alap Shah, Citrini Research