A Memo from 2028: AI Will Bring a Super Economic Crisis Sweeping the World

Editor's Note: The development of AI during the just-concluded Spring Festival holiday was far more impactful than the sudden emergence of DeepSeek last year, only this time, anxiety outweighed excitement.

On February 20th, the first trading day of the Year of the Horse in Hong Kong, the market capitalization of the four-year-old AI company MiniMax briefly exceeded HKD 300 billion during the session; ZhiPu AI, which went public almost simultaneously, also showed a sharp upward trend. The combined market capitalization of the two companies surpassed HKD 620 billion, approaching or even surpassing established internet giants like Kuaishou, JD.com, and Ctrip.

This anxiety is not unique to China. Sentiment around AI in China and the US has almost formed a mirror image. On one side, the 2026 Spring Festival Gala has become a stage for concentrated displays of AI and robotics, with technological narratives entering public life at a high density and strong pace; on the other side, the US stock market fluctuated repeatedly, with declines deepening in sectors like software, SaaS, and dining. The market began to re-discuss the second-order effects of AI—not just which company would win, but how it would change employment, consumption, credit, and the macro cycle.

The public's aversion to technological overload and financial elites' concerns about structural risks are converging. Against this backdrop of sentiment, this article from Citrini Research emerged and quickly became a concrete expression of anxiety. The article sets up a hypothetical scenario: in June 2028, the disruptive impact of artificial intelligence leads to mass white-collar unemployment, a decline in consumer spending, defaults on software-supported loans, and an economic contraction.

As of now, this article has garnered over 23 million reads on Twitter, making it the most high-profile public event recently. It is even seen as the trigger for the sharp volatility in the US tech sector on Monday (February 23rd), with the Dow Jones Industrial Average falling the most, down over 800 points. The US stock market's advance-decline ratio was brutal, with only 27% of stocks rising. Software, payment, and food delivery stocks were hit hard. Companies mentioned in the article, such as DoorDash (DASH), American Express (AXP), KKR (KKR), and Blackstone (BX), all fell over 8%.

This article does not necessarily explain the market decline, but it certainly amplified the existing unease in the market. In a phase where AI disruption, geopolitics, and macro uncertainty overlap, a sufficiently dark and self-consistent narrative is enough to provide an outlet for fragile sentiment.

The authors of this article are Citrini Research and Alap Shah. Citrini Research was founded by James van Geelen; Alap Shah graduated from Harvard University with a degree in economics, worked as an analyst at Citadel LLC after graduation, and has served as CEO of Littlebird since September 2024.

The following is the original text:

If our long-term bullish view on AI remains valid, could it actually mean bad news for the overall economy?

The following content is not a prediction but a scenario analysis. It is neither a deliberately fear-mongering bearish narrative nor a doomsday fantasy about AI. The sole purpose of this article is to attempt a systematic modeling of a previously underexplored possible path. This question was first raised by our friend Alap Shah, and we jointly explored this line of thinking in our discussions. This piece was completed by us; the other two were written by him and can be found separately.

We hope this article helps readers better prepare psychologically for potential left-tail risks before AI gradually changes how the economy operates, even making the structure itself increasingly counterintuitive.

What follows is a macro memo written by CitriniResearch in June 2028, attempting to trace back and review the formation process of the global intelligence crisis and its cascading effects.

Macro Memo: The Economic Consequences of Intelligence Surplus

CitriniResearch

(February 22, 2026) June 30, 2028

The unemployment rate announced this morning is 10.2%, 0.3 percentage points higher than market expectations. Affected by this, the market fell 2%, and the S&P 500's cumulative drawdown from its October 2026 high has reached 38%.

Traders are almost numb to this. Just six months ago, such a magnitude of unemployment would have triggered circuit breakers.

In just two years, the economy has evolved from "risks are controllable, impacts are confined to specific industries" into a system that no longer matches the experience any of us grew up with. This quarter's macro memo attempts to reconstruct this evolution, conducting a post-mortem systematic dissection of the economic structure before the crisis truly arrived.

Once upon a time, market sentiment was still exuberant. In October 2026, the S&P 500 briefly approached 8000 points, and the Nasdaq broke through 30,000. The first wave of layoffs centered on human labor replacement started in early 2026, and it did achieve the effect the capital market desired: improved profit margins, earnings beats, and rising stock prices.

Record corporate profits were swiftly reinvested into AI computing power expansion.

Macro data superficially remained brilliant. Nominal GDP repeatedly recorded mid-to-high single-digit annualized growth, productivity significantly increased, and real output per hour growth hit its highest level since the 1950s. All this came from AI agents that don't rest, don't take sick leave, and require no welfare benefits.

The wealth of computing power owners rapidly swelled, contrasted by a clear weakening in real wage growth. Despite official emphasis on record productivity, more and more white-collar jobs were being replaced by machines, forcing workers into lower-paying positions.

When signs of softening began to appear on the consumption side, commentators introduced a new concept: "Ghost GDP"—output reflected in statistical reports but not truly entering the real economic cycle.

On almost all technical metrics, AI was exceeding expectations; capital market narratives were almost entirely centered on AI. The only deviation was that the economic structure itself did not benefit synchronously.

In hindsight, this logic isn't complicated. If the output of a GPU cluster in North Dakota is equated to the economic contribution of 10,000 white-collar workers in midtown Manhattan, its impact resembles an economic pandemic rather than a cure.

The velocity of money subsequently stagnated. The human-centric, consumption-based economy, accounting for about 70% of GDP, rapidly contracted. Perhaps we could have realized this earlier by asking a simple question: How much money do machines spend on discretionary consumer goods?

The answer is obvious: zero.

Then, negative feedback began to self-reinforce: AI capability improves → companies need fewer employees → white-collar layoffs expand → displaced workers cut spending → profit pressures force companies to further increase AI investment → AI capability continues to improve...

This is a cycle lacking a self-braking mechanism, a spiral process of systematic human intelligence replacement.

The income capacity of the white-collar group and the resulting consumption willingness were structurally eroded. And this income is precisely the foundation sustaining the $13 trillion residential mortgage market. Underwriters had to re-evaluate a long-taken-for-granted question: whether so-called prime mortgages still possess sufficient safety margins.

Meanwhile, 17 consecutive years without a real default cycle led to a massive accumulation of software asset transactions in the private markets, largely backed by private equity. These transactions almost universally rested on the same assumption: ARR (Annual Recurring Revenue) would remain stable, grow persistently, and possess compounding properties.

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

If the shock were confined to the software industry, the situation might still be manageable, but reality was not so.

By the end of 2027, almost all business models built on intermediary roles began to come under pressure. Companies that profited by providing friction-based intermediary services to humans experienced widespread collapse.

At a deeper level, the entire economic system is essentially a chain of highly correlated bets on 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 negative feedback mechanisms that had long existed.

The market had been waiting for the inflection point where "bad news is good news" for nearly a year. Government-level discussions on response plans began, but public confidence in the government's ability to implement effective bailouts was rapidly fading. Policy responses have always lagged economic reality, and at this stage, the lack of systemic solutions itself is deepening the deflationary spiral.

How It All Began

At the end of 2025, agentic programming tools underwent a leap in capability.

A seasoned developer, using Claude Code or Codex, could replicate the core functionality of a medium-sized SaaS product within weeks. While covering all edge cases was difficult, its maturity was sufficient for a CIO reviewing a $500,000 annual renewal contract to seriously consider the question—"Why don't we build it ourselves?"

Since most corporate fiscal years align with the calendar year, 2026 IT budgets were finalized as early as Q4 2025. Back then, "agentic AI" was still conceptual.

Therefore, mid-year reviews became the first real stress test. Procurement teams, for the first time with a full understanding of these systems' true capabilities, re-evaluated existing spending decisions.

That summer, we interviewed a procurement manager at a Fortune 500 company. He recalled a key budget negotiation: the sales side planned to use the previous year's template—a 5% annual price increase plus the standard "your team can't function without us" pitch. But the procurement manager stated bluntly that he was already in talks with OpenAI, considering having their frontline deployment engineers use AI tools to directly replace the existing vendor.

Ultimately, the contract was renewed at a 30% discount. In his view, this was a relatively ideal outcome. SaaS long-tail companies like Monday.com, Zapier, and Asana were in a much tougher spot.

Investors had long expected the SaaS long tail to be hit first. After all, they account for about one-third of enterprise tech stack spending, inherently the most exposed.

The real blind spot was that core software, seen as system-of-record, was considered safe enough.

It wasn't until ServiceNow's Q3 2026 earnings report that the reflexivity mechanism truly surfaced:

ServiceNow's new ACV growth slowed from 23% to 14%; announced 15% layoffs and initiated a structural efficiency plan; stock price fell 18%.

—Bloomberg, October 2026

SaaS isn't dead. Building in-house still involves trade-offs in operational costs and complexity. But the mere feasibility of building in-house fundamentally changed the starting point of pricing negotiations.

More importantly, the competitive landscape structurally changed. AI drastically lowered the barrier to feature development and product iteration, causing differentiation to rapidly collapse. Incumbents were forced into price wars, fighting each other while facing a wave of new challengers empowered directly by agentic programming capabilities, unburdened by historical costs.

Only at this moment did the market truly realize the high interconnectedness of these systems.

ServiceNow charges per seat. When its Fortune 500 customers lay off 15% of staff, that means 15% of licenses are simultaneously canceled.

The same AI-driven layoff logic that improved client profit margins was also eroding its own revenue base in an almost mechanical way. This company selling workflow automation was ultimately disrupted by more efficient workflow automation; and its response path could only be layoffs, reinvesting the saved costs into the very technology disrupting it.

What else could they do? Stand still and slowly die?

Thus, the most direct and ironic outcome emerged: the companies most threatened by AI became its most aggressive adopters.

In hindsight, this seems logical; but at the time (at least for me), it wasn't. Traditional tech disruption models typically involve incumbents resisting new technology, being eroded by more agile newcomers, and eventually declining slowly. Kodak, Blockbuster, Blackberry—all followed this pattern.

But 2026 was different. Incumbents didn't choose to resist; they were simply powerless to resist. When stock prices fell 40% to 60%, and boards demanded clear countermeasures from management, these companies at the epicenter of AI shock effectively had only one path: layoffs, reinvest the savings into AI, and use AI to maintain output at lower costs.

From an individual company perspective, such decisions are perfectly rational; but at the aggregate level, they brought disastrous consequences. Every dollar saved in labor costs was converted into investment strengthening AI capabilities, paving the way for the next round of layoffs.

And software was just the opening act.

While investors were still debating whether SaaS valuations had bottomed, a more critical change had already occurred. This reflexivity logic had long spilled over from the software industry. The logic supporting ServiceNow's layoffs applied equally to all companies with white-collar costs at their core.

When Friction Goes to Zero

By early 2027, using large language models had become the default. People unconsciously used AI agents, often without even knowing the concept of an "AI agent," much like how most people didn't understand cloud computing but were accustomed to streaming videos. To the average user, it seemed more like an underlying feature—autocomplete, spell check—something a device should naturally have.

The open-sourcing of Qwen's agentic shopping assistant became a key catalyst for AI taking over consumption decisions. Within weeks, almost all mainstream AI assistants embedded some form of agentic e-commerce functionality. The maturation of distilled models allowed these agents to run directly on end devices like phones and laptops, no longer entirely reliant on cloud computing, significantly compressing marginal inference costs.

The point that should have made investors more alert was that these agents didn't wait for explicit user commands; they ran continuously in the background according to preset preferences. Consumption was no longer a series of discrete choices made individually by humans but transformed into a 7×24 automated optimization process running continuously for every connected consumer. By March 2027, the average US individual's daily token consumption had risen to 400,000, a 10x increase from the end of 2026.

And the next link in this chain was already loosening.

Intermediation Layer

Over the past fifty years, the US economy superimposed a vast rent-seeking structure on top of human limitations. Decisions take time, patience is limited, brand familiarity often replaces careful comparison, and most people accept suboptimal prices to click a few less pages. Trillions in enterprise value were built on the premise that these behavioral frictions would persist.

The initial change seemed insignificant: agents began eliminating friction. Subscription services that auto-renewed after months of non-use, pricing models that quietly increased after trial periods—all were redefined as terms open for renegotiation. The core metric underpinning the entire subscription economy, Customer Lifetime Value (LTV), began to see substantial declines.

Consumer agents gradually rewrote the operating logic of almost all consumer transactions. A human buying a box of protein bars hardly has the energy to compare prices across five platforms one by one; a machine can.

Travel booking platforms were hit first due to their highly standardized business logic. By Q4 2026, AI agents could assemble complete travel itineraries—covering flights, hotels, ground transport, points optimization, budget constraints, and cancellation rules—faster, cheaper, and with overall efficiency surpassing traditional platforms.

Insurance renewals didn't escape either. Business models reliant on policyholder inertia for profits were rapidly dismantled by agents performing annual automatic price comparisons: that 15%–20% premium from passive renewals disappeared almost overnight.

Financial advisors, tax preparation, routine legal matters… any industry whose value proposition was built on handling complex, tedious tasks for clients was impacted. Because for an agent, "tedious" doesn't exist.

Even areas thought to be protected by interpersonal relationship value weren't spared.

The real estate industry long relied on information asymmetry between buyers and sellers, maintaining a 5%–6% commission structure. When AI agents accessed MLS data and could instantly call up decades of transaction records, this knowledge advantage was rapidly replicated.

A March 2027 sell-side report described this phenomenon as "agent vs. agent warfare." Median buyer commissions in major cities had compressed from 2.5%–3% to below 1%, with more transactions completing without any human buyer's agent involvement.

We overvalued the importance of relationships. Many so-called relationships are essentially friction dressed in friendliness.

And this was just the beginning of the intermediation layer's collapse. Successful companies had spent billions building moats on consumer behavioral biases and psychological inertia, but these mechanisms quickly failed against machines.

Machines only optimize for price and fit. They don't care about your favorite app, aren't drawn to fancy checkout pages, don't choose the easiest option out of fatigue, and don't habitually reorder from the same platform.

What was destroyed was a special kind of moat: habitual intermediation.

DoorDash became a classic case. Agentic programming significantly lowered the barrier to entry for food delivery platforms. A skilled developer could deploy a fully functional competitor within weeks. A flood of new platforms emerged, quickly attracting drivers away by allocating 90%–95% of delivery fees directly to them. Multi-platform management dashboards allowed gig workers to simultaneously access twenty to thirty platforms, almost eliminating previous lock-in effects. The market fragmented extremely quickly, compressing margins to near zero.

Agents accelerated the collapse on both ends: they spawned competitors and prioritized calling them. DoorDash's moat was essentially built on a simple premise: "You're hungry, you're too lazy to compare prices, this app is on your home screen."

But an agent has no "home screen." It simultaneously queries DoorDash, Uber Eats, restaurant websites, and dozens of new