Apollo Chief Economist Warns: AI Assets May Face Painful Repricing

A Warning from Wall Street

Recently, the chief economist at Apollo Global Management issued a warning that the currently red-hot market for AI-related assets may face a "painful repricing." This assessment has drawn attention in tech communities such as Hacker News, and the market sentiment it reflects deserves serious consideration from every practitioner and investor watching the AI industry.

Founded in 1990 by Leon Black, Josh Harris, and Marc Rowan, Apollo Global Management is headquartered in New York and is one of the world's largest alternative asset management institutions, with assets under management exceeding $600 billion across private equity, credit, real estate, and other domains. To understand the weight of Apollo's warning, one must first appreciate the unique nature of the "alternative asset management" industry. So-called alternative assets are defined in contrast to traditional public-market assets such as stocks and bonds, encompassing private equity (PE), private credit, hedge funds, real estate, infrastructure, and more. The core advantage of such institutions lies in their informational dimension: they participate deeply in both the primary market (directly investing in unlisted companies) and the secondary market (engaging with listed companies through instruments such as credit), enabling them to form a 360-degree view of the entire capital ecosystem.

It is worth pointing out that the informational edge of alternative asset managers does not come from "insider information," but rather from the systematic accumulation of data that arises from deep participation across markets and asset classes. Apollo's private credit business makes it a direct creditor to numerous AI data center projects, giving it firsthand access to core financial data such as project cash flows, operational efficiency, and repayment pressure—information that often does not appear in public financial reports. Private credit, a form of direct lending that bypasses the traditional banking system, grew rapidly in the tightly regulated environment following the 2008 financial crisis, and now exceeds $1.7 trillion globally. Unlike bank loans, private credit institutions typically participate directly in negotiating loan terms and continuously monitor the borrower's operational conditions, giving them microdata of far finer granularity than public-market analysts possess.

It is necessary here to elaborate on the mechanics of the private credit market and its structural evolution in recent years. After the 2008 financial crisis, regulatory frameworks such as Basel III substantially raised capital adequacy requirements for traditional banks, causing them to sharply retrench in areas such as mid-market corporate lending and leveraged buyout financing, thereby leaving a vast market gap for private credit institutions. Private credit typically carries interest rates 200 to 400 basis points higher than bank loans of comparable maturity, but in exchange, borrowers must accept stricter financial disclosure obligations and more detailed operational monitoring covenants—and it is precisely these covenants that serve as the institutional source enabling Apollo to grasp the true operating conditions of AI infrastructure projects. In the case of data center financing, private credit contracts typically require borrowers to submit detailed operational metrics each quarter, such as server rack utilization rates, power usage effectiveness (PUE), and contract revenue coverage ratios; once a given metric breaches a preset threshold, lenders have the right to demand early repayment or additional collateral. This "embedded monitoring" mechanism makes private credit institutions one of the few external observers able to continuously and in real time sense the genuine pressures on projects amid the AI infrastructure investment boom.

When Apollo's lending teams observe shared signals of declining capital-use efficiency and rising repayment pressure across dozens of AI infrastructure projects worldwide, this microdata from the "creditor's vantage point" forms the underlying basis for its chief economist's warning. This also explains why the public statements of economists at institutions like Apollo—which rank among the largest globally in private credit—carry significant bellwether weight on Wall Street.

Apollo's business model also possesses a distinctive characteristic rarely noticed by outsiders: as an institution spanning both public and private markets, its investment portfolio naturally covers the upstream, midstream, and downstream of the AI value chain—from equity funds holding shares in listed companies such as Nvidia, to infrastructure debt providing construction financing for AI data centers, to growth-stage funds directly investing in AI startups. This full-value-chain perspective enables Apollo to simultaneously observe the true returns of AI investment across different layers and stages. When signals across all layers exhibit systematic divergence, the judgment of its economists carries a comprehensive credibility that few other institutions can match.

So-called "repricing" essentially refers to a systematic adjustment in the market's expectations of the value of a certain asset class. When an economist at an institution managing hundreds of billions of dollars like Apollo uses the word "painful," it typically means that current valuations have clearly diverged from fundamentals—and once market expectations converge, the magnitude of the price correction could be quite severe.

The Hidden Concerns of an AI Valuation Bubble

The Gap Between Capital Frenzy and Commercial Returns

Over the past two years, the technological wave centered on generative AI has attracted enormous capital inflows. In 2024 alone, the combined capital expenditures of the four major tech giants—Microsoft, Google, Amazon, and Meta—exceeded $200 billion, the vast majority of which flowed into AI infrastructure: GPU clusters, data center construction, and power supply. Nvidia's market capitalization once surpassed $3.3 trillion, OpenAI completed a funding round at a valuation of roughly $66 billion, and numerous AI startups secured substantial funding as early as their seed rounds. The entire market exhibits the classic characteristics of being "narrative-driven."

"Narrative-driven" is an important concept in behavioral finance, systematically articulated by Nobel laureate Robert Shiller in his book Narrative Economics. Its core thesis is that in markets with incomplete information, investors often rely on compelling stories rather than cold financial data to make decisions. By studying vast amounts of historical market data, Shiller found that emotionally resonant narratives spread through social networks like viruses, altering people's economic behavioral expectations and thereby producing real price effects at the macro level.

From the perspective of neuroscience and cognitive psychology, the reason narratives can overwhelm financial data has deep biological roots. When the human brain processes narrative information, it activates the same neural circuits as firsthand experience, whereas processing statistical data primarily relies on rational cognitive regions—the emotional activation intensity of the former far exceeds the latter. This makes grand narratives such as "AI will reshape everything" often more effective in driving investment decisions than price-to-earnings analysis. Behavioral finance scholars call this phenomenon the "availability heuristic": people tend to assess probability based on the cases most readily recalled from memory (such as the explosive rise of ChatGPT or the surge in Nvidia's stock price), rather than systematically analyzing fundamentals. This cognitive bias is further amplified in today's highly developed social media environment—algorithmic recommendation mechanisms naturally favor spreading content that triggers strong emotional reactions, causing the speed and reach of AI success narratives to far exceed any historical period.

When a narrative (such as "AI will reshape every industry") forms a social consensus and spreads virally, capital pours in and asset prices detach from fundamentals. During the dot-com bubble of the late 1990s, narratives of the "eyeball economy" and "user growth" propped up the valuations of numerous companies with no profit model whatsoever; the 2021 SPAC boom likewise relied on "imagined future potential" rather than current cash flows. The current AI market landscape bears a striking resemblance to these historical cases—a large number of valuations rest on extremely optimistic assumptions about future earnings rather than verifiable current financial performance. It is worth noting that the simultaneous frenzy in both public and private markets has also formed a kind of "dual valuation bubble": high valuations in public markets provide a comparative anchor for private markets, while financing cases in private markets in turn endorse the optimistic narratives of public markets. The two reinforce each other, making a potential correction all the more destructive.

Examined at the macro-data level, the current AI investment boom also exhibits a rare structural feature: the market is systematically underestimating the time lag between the scale of capital expenditure and quantifiable commercial returns. Take the most comparable historical case—the fiber-optic infrastructure investment boom (1996–2001), during which more than 390 million miles of fiber were laid worldwide. Yet it took roughly a full decade after the investment peak ended before broadband penetration truly triggered the large-scale commercialization of "killer apps." The payback cycle of AI infrastructure investment may be equally protracted—meaning the market's current assumptions about the AI profitability timeline may contain a systematic optimistic bias.

Herein lies the core concern of Apollo's economist: Has the capital market's optimistic expectations for AI already run far ahead of the commercial returns AI technology can actually generate?

Every technological revolution in history has been accompanied by valuation bubbles. The dot-com bubble of the late 1990s and the overinvestment in fiber-optic infrastructure in the 2000s both went through a phase of "narrative leading, reality lagging." Once the market realizes that profits commensurate with the investment are unlikely to materialize in the short term, the repricing of AI assets becomes inevitable.

Concentration Risk Cannot Be Ignored

The current rally in U.S. equities depends heavily on the pull of a handful of AI-related tech giants. This highly concentrated market structure means that once the AI narrative of these leading stocks is called into question, the entire market could suffer collateral damage—which is precisely why institutional economists repeatedly warn of systemic risk.

Taking U.S. equities as an example, a handful of core AI beneficiaries such as Nvidia, Microsoft, Google, and Meta already account for weightings in the S&P 500 index that fall within historical extremes. This means that price movements in these companies have a disproportionately amplified effect on the entire index. From a risk management standpoint, a highly concentrated market harbors a "single point of failure" vulnerability: once the market begins to doubt the sustainability of demand for Nvidia's GPUs, or once a leading AI company's commercialization data falls short of expectations, the resulting sell-off could rapidly transmit to the entire market through passive investment vehicles such as index funds, triggering so-called "systemic deleveraging."

This transmission mechanism warrants particular vigilance in modern financial markets. The widespread adoption of passive investment vehicles (ETFs, index funds) has altered the risk structure of markets: when core AI stocks exceed a certain weighting threshold, any negative expectation targeting these companies could trigger mechanical selling by passive funds, forming a negative feedback loop of "valuation decline → index decline → passive fund redemptions → further selling." The danger of this mechanism lies in its nonlinear nature—once the redemption-trigger threshold for passive funds is breached, the system will automatically execute sell orders at a pace far exceeding the speed of human intervention, and the involvement of high-frequency trading algorithms will further amplify the magnitude and speed of price swings. By some estimates, the scale of passive investment in the U.S. now exceeds that of active management, meaning this kind of mechanical transmission effect has never appeared in any past market fluctuation, making the transmission speed and severity of this potential repricing difficult to predict accurately based on historical experience.

Further worth noting is that market concentration risk also involves a kind of "hidden leverage amplification" effect in the derivatives market. Core AI targets such as Nvidia are among the most actively traded instruments in the current options market. The enormous open interest in options contracts means that market makers must continuously adjust their spot positions to maintain delta hedging—when falling stock prices prompt market makers to massively sell to close out put positions, the hedging demand in the derivatives market forms a "gamma squeeze" reverse effect at key price levels, further amplifying volatility in the spot market. This coupling effect between the public market and the derivatives market is one of the most difficult-to-quantify hidden dangers within the current AI asset concentration risk.

Why a "Painful" Repricing?

From an asset-pricing perspective, a "painful" repricing typically exhibits the following features:

• Contraction of valuation multiples: Metrics such as price-to-earnings (P/E) and price-to-sales (P/S) ratios rapidly retreat from extremely high levels toward historical averages. To understand this mechanism, one must recognize that the P/S ratios of some AI software companies already exceed 30 to 50 times, meaning investors are assuming their profit margins and growth rates will remain extremely high for years to come. Historical data show that during the tech-stock valuation compression of 2022, high-P/S SaaS companies fell by an average of over 60%, while current valuation multiples of AI-related companies are generally higher than those at the peak of the 2021 SaaS bubble;
• Reversal of capital flows: Hot money chasing AI concepts retreats, shifting toward asset classes with greater certainty;
• Chain-reaction transmission: Spreading from the secondary market to the primary market, making it harder for AI startups to raise funds and causing valuations to shrink;
• Impact on confidence: Investors begin to waver on the long-term prospects of the entire sector.

It is worth emphasizing that repricing does not mean AI technology itself loses value. Even after the dot-com bubble burst, companies with genuine core competitiveness (such as Amazon and Google) still rose from the rubble. After the 2000 dot-com bubble burst, the Nasdaq index fell about 78% from its peak, and Amazon's stock price plunged more than 90% from its 1999 high. What ultimately allowed these companies to rise was the infrastructure capabilities and user scale they built during the bubble, forged during the "winter" of the bursting bubble through extremely rigorous optimization of their business models—it was precisely during that period that Amazon forged its unmatched supply chain efficiency and cost control capabilities. From a broader perspective, overinvestment during a bubble is not entirely worthless: the vast fiber-optic infrastructure laid in the 1990s, though it caused enormous losses, provided the material foundation for the spread of broadband internet in the mid-2000s. Likewise, the hundreds of billions of dollars in current AI data center investment, even if unable to generate commensurate commercial returns in the short term, will objectively lay a computing-power foundation for broader AI applications in the future. Market repricing is often a necessary process of "separating the wheat from the chaff," and AI companies that can pass the "stress test" may instead build deeper moats at the bottom of the cycle.

Practical Takeaways for Practitioners

Return to Genuine Commercial Loops

For AI entrepreneurs and technical teams, this warning conveys a clear signal: technological capability must be converted into sustainable commercial value. The era of raising funds simply by telling an "AI concept" story is receding. Only products that can complete a paying loop and generate positive cash flow have the true confidence to survive the cycle.

At the practical level, AI business models are primarily divided into several layers: the underlying compute and foundation model layer (such as Nvidia and the OpenAI API), the industry-vertical application layer (such as AI coding assistants and medical imaging diagnostics), and the enterprise solution layer. From an investment standpoint, the layer closest to "willingness to pay" is often enterprise applications—once B2B customers form a dependency, renewal rates are high and their willingness to pay is strong. The core metrics for evaluating the commercial value of B2B AI products typically include: net revenue retention (NRR), the ratio of customer acquisition cost to lifetime value (CAC/LTV), and the conversion cycle from pilot to full deployment. An NRR above 120% often means the product has been deeply embedded into customer workflows, forming genuine usage stickiness—this is an important dividing line between "real commercial value" and "conceptual bubble." AI products aimed at B2C consumers, by contrast, generally face the challenges of low payment conversion rates and difficult user retention.

The true profitability of the current AI industry is far less optimistic than market valuations reflect—most AI application companies are still in the "burn cash for growth" phase. Companies that can truly achieve rapid growth in ARR (Annual Recurring Revenue) sufficient to cover compute costs remain a minority in the entire industry. Behind this lies a structural challenge rarely noticed by outsiders: AI companies face unique "inference cost" pressure—that is, the GPU compute expense consumed by each model invocation.

The structural pressure of inference costs is far more complex than the surface numbers suggest and differs fundamentally from traditional software economics, warranting an in-depth discussion here. Under the traditional SaaS model, the marginal cost of a software product is nearly zero—whether an enterprise management software serves 100 companies or 100,000 companies, the cost of replicating its core code is negligible, which is precisely why traditional SaaS companies can see gross margins climb steadily to 70% to 80% at scale. AI inference, however, breaks this economic logic: every user interaction with a large language model requires executing billions or even hundreds of billions of floating-point operations on GPU clusters, consuming real electricity and compute resources, with marginal costs growing nearly in proportion to the number of users.

The inference cost of current mainstream large language models is roughly several to dozens of times the training cost, and as users demand higher-quality answers, model parameter counts continue to climb and the compute consumed per inference keeps rising. More critically, this "diseconomy of scale" characteristic fundamentally overturns the internet-era business logic of "the bigger, the more profitable," posing a severe challenge to traditional SaaS valuation frameworks in the AI era. Take OpenAI as an example: industry estimates put its 2023 inference costs at hundreds of millions of dollars, and even with the rapid growth of ChatGPT Plus paying users, it remained in an overall loss position—the core reason being that the inference cost generated per active user exceeded its subscription revenue.

The industry has seen some exploration of compressing inference costs through technical means such as "model distillation" and "mixture of experts" (MoE)—model distillation reduces inference overhead by transferring the knowledge of a large model to a smaller one, while MoE improves efficiency by activating only a subset of parameters during each inference—but the trade-off between effectiveness and cost in these approaches remains an industry challenge. It is worth noting that emerging model vendors such as DeepSeek have, through aggressive engineering optimization, compressed inference costs to a fraction of those of mainstream competitors. This trend both signals the possibility of computing-power democratization and poses a potential competitive threat to leading model vendors that rely on high inference pricing to maintain gross margins—the rapid decline in inference costs is a double-edged sword. While it lowers the barrier to deploying AI applications, it also compresses the pricing space across the entire value chain. AI applications that can truly achieve profitability need to make simultaneous breakthroughs in model efficiency optimization, differentiated pricing strategies, and user lifetime value (LTV) management—which is precisely the core commercial proposition that most AI companies have yet to solve.

Viewing Market Cycles Rationally

For investors and industry observers, maintaining clear-headed independent judgment is especially important. As a foundational technology, AI's long-term transformative significance is beyond doubt, but this does not mean that any target at any point in time is worth chasing at a high price. Distinguishing between "the long-term value of the technology" and "the short-term price of the asset" is the key to preserving principal amid a potential AI market repricing.

In practice, this distinction requires investors to possess the ability to identify "value anchors"—that is, which segments of the AI value chain have already had their commercial value fully or even excessively reflected in current prices, and which segments remain undervalued by the market. Historical experience shows that the companies that truly and continuously create value in each technological revolution are often not the earliest concept companies, but rather the "second wave" of beneficiaries who quietly accumulate capabilities during the infrastructure-building period and monetize quickly during the application-explosion period. This pattern has deep structural roots: in the early stages of a technological revolution, uncertainty is extremely high, and the trial-and-error costs borne by pioneers are ultimately shared by the entire market; whereas second-wave beneficiaries can build on the failed experiences of their predecessors to achieve scale at lower trial-and-error costs and with a clearer commercial path. In the internet era, when AWS launched in 2006, six years had already passed since the dot-com bubble burst, and the infrastructure operations experience Amazon accumulated during the bubble became the core competitiveness of its commercial cloud services. In the mobile internet era, WeChat likewise completed its ecosystem building after the early mobile-app frenzy cooled, drawing on a deep understanding of user needs. This pattern is worth close attention as a reference for investors in the AI era.

From a more actionable standpoint, there are several "value depressions" worth watching in the current AI value chain: first, the layer of deep integration of AI applications into traditional enterprise workflows—compared with betting directly on the models themselves, the "system integrators" and "industry adaptation layers" that help traditional industries deploy AI often carry more reasonable valuations; second, the non-GPU segments of AI infrastructure—supporting infrastructure such as network interconnection, cooling systems, and power supply is relatively overlooked in the current narrative, yet its business models are more certain and its valuations lower; third, traditional software companies that have been unfairly sold off due to the threat of AI displacement but can actually achieve efficiency leaps with the help of AI tools. Identifying these structural opportunities requires investors to move beyond the coarse-grained judgment of being "broadly bullish on the AI sector" and to build a fine-grained analytical framework based on the real supply-and-demand relationships of each segment of the value chain.

Conclusion

This warning from Apollo's economist is less a pessimistic view of AI's technological prospects than a calm correction of current market sentiment. The technological revolution is real, but the pricing mechanisms of capital markets have always had a tendency to overreact. When bubble and value intertwine, rational judgment becomes especially precious.

Whether or not the "painful repricing" truly arrives, for everyone caught up in the AI wave, returning to commercial fundamentals and squarely facing valuation risks is a prudent posture worth upholding.