In the four weeks through early February, more than $17 billion in broadly syndicated U.S. tech loans fell to distressed trading levels.1 Private credit stocks slid. Portfolio managers repositioned. The AI economy, for the first time, had a credible bear case — but the underlying drivers were scattered across disconnected headlines, analysis, and research. No one had assembled a coherent thesis connecting them.
Then someone did. Citrini Research’s “The 2028 Global Intelligence Crisis,” published February 22, provided a glimpse into the [not so distant] future where the economy has been brought to its knees by AI-driven disruption in 2028: the S&P 500 down 38% from its peak, unemployment at 10.2%, the mortgage market cracking.2 It is one of the more sophisticated attempts to connect AI’s technological trajectory to its macroeconomic consequences. It is also, as the authors themselves note, a scenario — not a forecast.
That distinction has been largely lost. And it matters.
Scenario planning is used across financial institutions and regulatory bodies to stress-test assumptions and surface hidden risks. It is a tool for thinking, not a verdict. Markets and commentators have stripped that context away, treating the piece as a hard directional call. The resulting reaction has been filled with impulsive portfolio repositioning and surface-level reporting that misses the bigger picture.
Many of the risks identified are worth debating. The two-year timeline is not.
The "what if" storyline begins with a mundane corporate budget meeting. By mid-2026, AI coding tools have become capable enough that an internal team can replicate the core functionality of a six-figure software subscription in weeks. Vendors slash prices. The SaaS industry, built on the assumption that switching costs would keep customers locked in, finds that assumption evaporating.
From there, the disruption becomes a financial crisis.
The underlying driver: the private credit market which now stands at an estimated $2.1 to $2.3 trillion, according to the IMF and Moody's.³ ⁴ It is heavily concentrated in private-equity-owned software companies. S&P Global Ratings reported in February 2026 that software represents roughly 20% of assets in middle-market private credit vehicles.⁵ Many of these deals were structured around the assumption that subscription revenue would remain "recurring." If it stops recurring, the loans go bad.
UBS analysts put the risk exposure even higher, estimating that 25% to 35% of all private credit portfolios face elevated AI disruption risk — and projecting that in a severe scenario, default rates could approach 13%.⁶ This is far above historical stress expectations.
Meanwhile, the overlap between private credit and the broader financial system has deepened. Over the past decade, alternative asset managers including Apollo, Blackstone, and KKR have acquired or built insurance arms, channeling policyholder premiums and retirement savings into private credit portfolios.⁷ About one-third of the life insurance industry's $6 trillion in assets now sits in some form of private credit, according to Moody's.⁷ What was once a niche institutional asset class now touches Main Street balance sheets.
The "Ghost GDP" concept is similarly sharp. If AI productivity gains accrue almost entirely to the owners of compute and capital — rather than flowing through wages to workers who then spend — headline economic output could look healthy while the consumer economy quietly hollows out. Personal consumption expenditures account for roughly 68% of U.S. GDP.⁸ The IMF has flagged a version of this concern, noting that AI productivity gains are currently making the wealthy wealthier while benefits for lower-wage workers remain uncertain.⁹
Finally, the 2028 scenario also identifies a significant geopolitical risk that most Western analysis tends to ignore: the potential for AI to hollow out the economies of nations built on digital services exports. As one of the leaders in this space, India's IT and business process outsourcing sectors contribute roughly 7.5% of the nation's GDP and have been the single largest driver of middle-class expansion for two decades.¹⁰ As AI automates coding, customer support, and back-office tasks, the risk is not just sectoral — it threatens the economic model that built modern India.¹¹ Disruption at that scale does not stay local.
These are genuine, long-term risks worth discussing and drawing greater attention to. The scenario analysis is right to name them.
The problem is the timeline. When held against where the AI industry actually stands today — the energy constraints, the enterprise failure rates, the deployment costs — the two-year collapse does not hold up. The direction of risk is credible. The speed at which they will materialize is not.
The Kilowatt Constraint. The most significant physical barrier to the presented scenario is energy. According to the International Energy Agency, data centers could consume over 945 terawatt-hours of electricity annually by 2030. This is roughly double their 2025 consumption and equivalent to Germany's entire current electricity demand.¹² Gartner and S&P Global project a doubling or even tripling of data center power demand by 2030.¹³ ¹⁴
This surge is running into a grid that cannot handle it. Grid Strategies reported in November 2025 that the five-year forecast for U.S. electric power demand growth has increased sixfold in just four years — driven almost entirely by data center expansion.¹⁵ Building new high-voltage transmission lines takes seven to twelve years once permitting is factored in. OpenAI CEO Sam Altman himself has stated that future AI progress depends on an energy breakthrough.¹⁶
This planned build-out faces a threat that gets ignored entirely: the public. Data centers have become America's hottest NIMBY issue. According to Data Center Watch, $64 billion worth of U.S. data center projects have been blocked or delayed by local resistance, with at least 142 activist groups organizing across 24 states.¹⁷ In Warrenton, Virginia, residents voted out every town council member who supported an Amazon data center proposal. In Monterey Park, California, a community rallied against a planned facility and won.¹⁸ The infrastructure required to power the AI revolution is not just an engineering challenge. It is a political one.
The Corporate Reality. The piece is right that incumbents will feel the pressure. AI-native startups, unburdened by legacy infrastructure or institutional bureaucracy, will move faster, ship cheaper products, and force established companies into uncomfortable strategic reckonings. That competitive threat is real.
But then the scenario makes an enormous leap: that these same incumbents from Fortune 500 companies with decades of legacy IT systems, rigid procurement cycles, deeply entrenched processes, and chronic talent shortages in AI and data engineering will [miraculously] respond by transforming themselves at startup speed.
This is not how large organizations work.
Recognizing a threat and overcoming the structural barriers to addressing it are fundamentally different acts. A CEO can declare an AI mandate on Monday. It doesn't simply manifest into reality. Gartner projected in June 2025 that more than 40% of agentic AI projects will be canceled outright by the end of 2027 — not because the technology is unimpressive, but because of escalating costs, unclear business value, and inadequate risk controls.¹⁹ The competitive pressure the piece describes will accelerate urgency. It will not dissolve the bureaucracy, the data quality problems, or the talent gaps that stand between urgency and execution.
The $200 Claim. One of the scenario's most quotable lines — that a Claude AI agent can do the work of a $180,000 product manager for "$200 a month" — refers to a retail subscription price that bears no relationship to the true cost of deploying AI at enterprise scale. But to understand why, it helps to distinguish between the two ways companies are actually using AI today.
The first is the AI agent model — tools like Microsoft Copilot and Salesforce Einstein that operate inside existing software platforms, augmenting human workers. They make employees faster, but they don't replace them. They also don't replace the underlying software. Every agent that automates a workflow in Salesforce or Microsoft 365 still requires the underlying license. And vendors are responding to AI adoption not by cutting prices, but by raising them. Microsoft announced M365 commercial price increases of 5% to 33% effective July 2026, with Copilot remaining a $30 per user add-on.²⁰ Zylo's 2026 SaaS Management Index found that enterprise spending on AI-native applications surged 108% year-over-year — and 393% at companies with more than 10,000 employees.²¹ Seventy-eight percent of IT leaders reported unexpected charges tied to AI pricing models in the past 12 months.²¹ Under this model, the correlation between reduced software spending and scaled enterprise AI is not being realized on the trajectory required for the depicted 2028 scenario.
The second is the agentic worker model — autonomous systems that handle multi-step workflows with minimal human oversight, and that the scenario assumes will make the per-seat software model obsolete. This is where the disruption thesis lives. And where the evidence simply doesn't support the narrative.
Tech investor Jason Calacanis disclosed on the All-In podcast in February 2026 that running agentic workers for his businesses was costing $300 per day, per worker — roughly $100,000 per year — at only 10% to 20% of intended capacity.²² He described provisioning each agentic worker with its own software license — not eliminating seats, but adding them. Social Capital CEO Chamath Palihapitiya noted on the same episode that AI models "need to be at least two times as productive as another employee" just to justify their cost. Investor Mark Cuban added that eight Claude AI agents could cost twice as much as a single employee to accomplish the same daily output.²² These are not AI skeptics. They are among the technology's most prominent advocates, and they point to the same conclusion: inference cost is a fraction of the total cost of ownership, which includes infrastructure, integration, monitoring, and the specialized talent required to keep these systems running.
The reliability problem makes the economics worse. Research published by Patronus AI in late 2025 found that an agent with a 1% error rate per step compounds to a 63% chance of overall failure by the hundredth step.²³ Even at a 95% per-step success rate, a 20-step workflow yields an overall completion rate below 36%. An agentic worker that fails two-thirds of the time on complex workflows is not replacing a $180,000 employee. It is creating a new cost center that still requires human oversight and, at a minimum, that human's software license.
Eventually, however, both models will put pressure on existing software economics. That pressure will come from three directions:
Reduced headcount shrinking the number of per-seat licenses companies need
Developing a more technically skilled workforce that can extract greater value from AI tools (which in turn accelerates further headcount reduction)
The eventual development of proprietary systems built by that same skilled workforce to replace incumbent platforms altogether
Each driver reinforces the others, and together they represent a genuine long-term threat to the SaaS model as it exists today and potentially the larger economy.
But the path from here to there is not a straight line, and it is not a short one. Building enterprise-grade software that replaces Salesforce or ServiceNow requires significant development, testing, security hardening, organizational change management, and the internal talent to maintain what gets built long after the initial deployment.
Finally, an important reminder that any replacement system will still carry its own costs: agentic workflows, API access, and infrastructure overhead that includes the workforce capable of executing this transition (which happens to be in short supply and high demand).
The compounding effect of these three drivers depends on each preceding stage maturing first — you cannot build proprietary replacements without the skilled staff, and you cannot free up the budget for skilled staff without the headcount efficiencies that have barely begun. The ROI for this transition is not as straightforward or immediately value-generating as the memo suggests. Will we eventually get there? Yes. Will it happen in 24 months? No.
The private credit market's concentration in software is a genuine vulnerability. Not because a 2028 crisis is inevitable, but because the gradual repricing of SaaS multiples and the slow erosion of recurring revenue predictability will create a long tail of credit stress over the coming years. The question for investors is not whether defaults will happen, they will. The issue is when, to what extent, and whether current portfolio valuations reflect that reality.
For business leaders, the strategic imperative is not to wait for AI to mature but to start building the necessary institutional knowledge to ensure readiness for when it does. The companies that begin integrating AI today, even imperfectly, are developing the workflows, the data infrastructure, and the internal expertise that will separate them from competitors in meaningful ways. Whatever scenario ultimately plays out, the leading organizations will be those that began the slow, unglamorous work of integration years earlier.
The scenario ends with a note of genuine humility: "You're not reading this in June 2028. You're reading it in February 2026." It is, at its core, a warning to act before any crisis arrives.
Ultimately, the current risk is not what was depicted. It's misreading it. A doomsday prediction often triggers two costly mistakes: rushed, uninformed decisioning that creates operational chaos, or strategic paralysis that leaves companies exposed. Both responses can accelerate aspects of the "what if" 2028 outcomes. The real cost of misreading the future isn't about a failed forecast. It's that it has the power to distort the decisions being made now.
Footnotes
¹ CNBC. (2026, February 3). "Private credit stocks plummet on concern about exposure to software industry disrupted by AI." cnbc.com
² Citrini Research. (2026, February 22). "THE 2028 GLOBAL INTELLIGENCE CRISIS." citriniresearch.com
³ International Monetary Fund. (2024, April 8). "Fast-Growing $2 Trillion Private Credit Market Warrants Closer Watch." imf.org
⁴ Moody's. (2026, January 21). "Private Credit Outlook 2026." moodys.com
⁵ S&P Global Ratings. (2026, February 12). "AI Disruption Worries Spill Over To Private Credit Markets." spglobal.com
⁶ Bloomberg / UBS. (2026, February 2). "Private Credit Defaults Would Hit 13% in UBS Worst Case for AI." bloomberg.com
⁷ Bloomberg. (2025, November 16–21). "America's Insurance" investigative series. bloomberg.com
⁸ Federal Reserve Bank of St. Louis / BEA. "Shares of Gross Domestic Product: Personal Consumption Expenditures (Q4 2025: 67.9%)." fred.stlouisfed.org
⁹ Fortune. (2026, January 24). "AI productivity gains are making the rich richer, and they'll wipe out jobs — but the IMF chief sees a silver lining for low-wage workers." fortune.com
¹⁰ Government of India / Press Information Bureau, citing NASSCOM data. (2026, February 12). "AI@Work: Driving Productivity, Jobs, and Innovation." pib.gov.in
¹¹ The Economic Times. (2026, February 16). "AI Impact Summit: Profit comes first, not jobs for IT companies, warns Vineet Nayyar as AI boom could worsen job crisis." economictimes.com
¹² International Energy Agency. (2026, January). Electricity 2026. iea.org
¹³ Gartner. (2025, November 17). "Gartner Says Electricity Demand for Data Centers to Grow 16% in 2025 and Double by 2030." gartner.com
¹⁴ S&P Global. (2025, October 14). "Data center grid-power demand to rise 22% in 2025, nearly triple by 2030." spglobal.com
¹⁵ Grid Strategies. (2025, November). National Load Growth Report 2025. gridstrategiesllc.com
¹⁶ Reuters. (2026, January 16). "OpenAI's Altman says energy breakthrough is needed for future AI." reuters.com
¹⁷ Data Center Watch. (2025). "$64 Billion of U.S. Data Center Projects Have Been Blocked or Delayed Amid Local Opposition." datacenterwatch.org
¹⁸ The Guardian. (2026, February 7). "A California community rallied against a data center — and won." theguardian.com
¹⁹ Gartner. (2025, June 25). "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027." gartner.com
²⁰ Microsoft. (2025, December 4). "Advancing Microsoft 365: New capabilities and pricing update." Microsoft 365 Blog. microsoft.com
²¹ Zylo. (2026, January 29). "2026 SaaS Management Index." As reported by CFO Dive, "SaaS bills climb as AI shakes up pricing." cfodive.com
²² All-In Podcast, Episode 219. (2026, February 15). "E219: Agentic AI costs, the real math on replacing workers." youtube.com
²³ VentureBeat. (2025, December 17). "AI agents fail 63% of the time on complex tasks, Patronus AI says." venturebeat.com