From Prisoner's Dilemma to Reflexivity: The Market Game of AI Infrastructure Investment

Last week, two risk threads in AI infrastructure surfaced simultaneously: CoreWeave suffered a double blow to its stocks and bonds, with financing costs surging; Meta's massive AI capital expenditures and its swing to negative free cash flow once again triggered market concerns over the input-output ratio. The former exposed the financing fragility of highly leveraged infrastructure companies, while the latter prompted investors to re-examine the ever-increasing capital expenditures of tech giants.

Cloud computing giants are mired in a classic prisoner's dilemma—everyone understands the cost of collective overinvestment, yet none dares to hit the brakes unilaterally. According to Reuters' analysis of LSEG consensus forecast data, the five hyperscale cloud providers—Microsoft, Alphabet, Amazon, Meta and Oracle—are expected to see total capital expenditures exceed their free cash flow by 2027, with every additional $1 of operating cash flow corresponding to about $1.57 of additional capital expenditures. This scissors gap is pushing tech platforms once known for being "asset-light" toward a new business model highly dependent on capital markets.

Investor Citrini, in an interview on the Monetary Matters podcast, directly pointed out the core risk of this shift: "In the first three years of this AI trade, sentiment-driven selloffs always had a floor—because most of the spending came from hyperscalers' operating cash flow. But now, more and more infrastructure is being financed on rather astonishing expectations." When the source of financing shifts from internal cash flow to equity issuance and debt financing, market sentiment itself gains the ability to directly influence the scale of physical investment.

If market sentiment takes a sharp turn for the worse, that incremental capital will dry up—because this is no longer a company spending its own money, but the market deciding.

This is precisely the "reflexivity" risk (reflexivity refers to the circular, two-way interaction and influence between cause and effect): confidence supports financing, and financing supports investment; once confidence weakens, the contraction of capital will in turn hit investment and growth expectations, forming a negative feedback loop.

No one dares to stop: the game structure of the prisoner's dilemma

Cornell Capital Group's research report "The AI Prisoner's Dilemma" describes the internal logic of this arms race quite clearly: any giant that places a heavy bet on AI while competitors do not can gain market dominance; if it does not invest while rivals do, it faces the risk of being eliminated. Therefore, regardless of how competitors choose, aggressive investment is the "rational strategy" for every company.

The numbers confirm the intensity of this logic. In 2024, the five companies Amazon, Alphabet, Microsoft, Meta and Apple had combined capital expenditures of about $227 billion; in 2025 this figure jumped to $391 billion; management guidance for 2026 indicates the total could exceed $635 billion. Goldman Sachs forecasts that cumulative capital expenditures by hyperscale cloud providers between 2025 and 2027 will reach $1.15 trillion to $1.4 trillion—more than double the $477 billion from 2022 to 2024.

At the individual company level, 2026 guidance is even more aggressive: Amazon targets $200 billion in spending, Alphabet guides $175 billion to $185 billion, with CEO Sundar Pichai even saying "that number may not be enough," and Meta guides up to $135 billion.

Analyst Amin Mrini, in his report "The $700 Billion Prisoner's Dilemma," points out that the five companies' combined committed 2026 capital expenditures are about 3.8 times their combined free cash flow—"the most cash-generative businesses in history are on a trajectory of spending nearly $4 for every $1 of free cash flow they generate, filling the gap with record debt issuance."

The structural root of this dilemma lies in the incompressibility of infrastructure. Data centers, power agreements and custom chips all require multi-year advance procurement and construction cycles. By the time demand is clearly visible on the revenue side, the response window has long closed. Mrini writes: "It is precisely this lag that turns expectations into obligations."

A mirror of history: the fate of railways, aviation and fiber optics

Cornell Capital's report cites three historical episodes, providing a sobering reference frame for this AI arms race.

During the 19th-century American railway boom, railway stocks and bonds totaled $10.6 billion (about $400 billion in today's terms), far exceeding the $1.2 billion in government debt at the time. The railways did transform America, but in the "Long Depression" after the Panic of 1873, nine out of ten railway companies failed, and during the Panic of 1893 a quarter of railway mileage went into receivership. The beneficiaries were manufacturers, agricultural producers and merchants, not railway shareholders.

The story of commercial aviation is even more straightforward. Over the past 80 years, the global aviation industry has accumulated about $86 billion in net profit, while the three years of the COVID-19 pandemic alone wiped out nearly $190 billion. Warren Buffett once summed up this structural predicament: "If a far-sighted capitalist had been present at Kitty Hawk, he would have done his successors a huge favor by shooting Orville Wright down." Buffett himself ultimately sold out of his holdings in four U.S. airlines in 2020 at a loss of about $5 billion.

The telecommunications bubble of the late 1990s is the most direct analogy. At the time, more than $500 billion poured into fiber-optic cable and wireless networks, most of it financed by debt. By 2002, more than 23 major telecom companies had filed for bankruptcy, about $2 trillion in market value had been wiped out, and less than 5% of the fiber laid was lit and in use. Yet it was precisely the cheap bandwidth left behind by those bankrupt investors that incubated Netflix, YouTube and the entire cloud computing industry.

Cornell Capital's conclusion cuts to the heart of the matter: "The paradox of transformative technology is this: the more companies spend and the more intense the competition, the faster the technology is commoditized. Commoditization is a blessing for consumers, but a corrosive force for investors' profit margins."

The scissors gap in cash flow: a stress test at the data level

Reuters' analysis of LSEG data provides empirical support for this logic.

Oracle is currently the most visibly stressed case. Its capital expenditure as a share of operating cash flow soared from 47% in fiscal 2022 to 174% in fiscal 2026 (ending in May this year)—capital expenditure that year was $55.7 billion, operating cash flow only $32 billion, and free cash flow has turned negative. The company's share price has fallen 36% this year, and it plans to raise $45 billion to $50 billion through debt and equity financing for cloud infrastructure expansion.

Microsoft recorded $35.8 billion in operating cash flow in its most recent fiscal quarter, while capital expenditure (including finance leases) reached $37.5 billion over the same period. Amazon's operating cash flow grew 30% year on year to $148.5 billion over the past 12 months, but free cash flow has fallen to $1.2 billion.

Shay Boloor, chief market strategist at Futurum Equities, said: "Investors underestimate the extent to which AI is fundamentally changing the business models of big tech companies. These companies have historically been viewed as asset-light platforms, because revenue growth could far outpace capital needs, but AI is pushing them toward a hybrid model in which the economics of software, advertising and cloud computing increasingly depend on massive physical infrastructure spending."

Mrini's calculations are even more severe: extrapolating capital expenditure to 2027 at the current guidance pace, and using a 15% pre-tax return on investment as the benchmark, AI computing revenue would need to reach about $2.9 trillion by 2030 to cover the cost of deployed capital—more than three times the size of the entire current cloud computing market. Starting from the current base of about $211 billion, that implies sustaining an annual growth rate of about 68% for five consecutive years, "without any precedent in infrastructure markets."

Reflexivity: when the financing structure changes the nature of the risk

Citrini's description of this shift in the interview is one of the most precise frameworks for understanding the risk in the current AI trade.

He divides the change in the financing structure of AI investment into two phases. In the first phase, hyperscalers funded spending from their own operating cash flow, which meant that no matter how market sentiment fluctuated, as long as these companies kept investing, the fundamentals of beneficiary companies such as semiconductor makers had a "floor." "Whether it was tariffs, the yen carry trade, or whatever it was we were worrying about in September 2023 that I can no longer even remember, the fundamentals kept improving."

The second phase, the present one, is fundamentally different. Citrini says plainly: "Now, almost every incremental dollar of AI spending comes from equity financing—whether venture capital or public markets—and debt financing. That means AI can become more useful, usage can keep expanding, but you may have paid too high a price for the compute that supplies it; or sentiment can have a more pronounced impact—because unlike before, there is no fundamental floor, and if market sentiment turns, that incremental money will dry up, because it is no longer just one company spending, but the market deciding."

The core of this mechanism is that the change in the source of financing turns market sentiment from an exogenous variable into an endogenous variable that can directly affect the scale of investment. Citrini calls this dynamic "reflexivity"—confidence supports financing, financing supports investment, investment supports revenue expectations, and revenue expectations in turn support confidence. And once any link in this chain cracks, the entire positive feedback loop can reverse.

He offers a specific trigger scenario: "I think the turning point in this cycle will arrive with a whimper, not a bang. It won't be a company default triggering a crash; rather, we will cut the combined ARR forecast from the labs from $400 billion to $360 billion, and then, because that growth has already been fully priced in, that incremental change will spread outward and drag the whole system down."

The Jevons Paradox fails: the key gap in demand elasticity

A core optimistic assumption underpinning AI infrastructure investment is the "Jevons Paradox"—falling compute costs will spur an explosion in demand, thereby absorbing all new capacity. Mrini acknowledges in the report that this logic exists, but at the same time points out its limitations in the AI context: enterprise integration is slow, regulatory constraints are real, and efficiency gains occur mainly at the model layer, so the demand response may flow to application developers and model providers rather than to the infrastructure layer that sells compute. "The Jevons Paradox can hold at the level of AI as a technology while failing at the level of AI infrastructure as a business model."

Citrini cited economic index data published by Anthropic: for every 10% reduction in task cost, usage increases by only about 3%. Perhaps when the model is not working well, you actually need more tokens, because you have to keep revising. "This is not the Jevons Paradox at all." He also pointed out that current AI revenue growth is about 25% to 30% per year, while capital expenditure growth is close to 60% to 70%, and the gap between the two widens with each guidance cycle.

He gave clear conditions for "when we would turn bullish on semiconductors again": "If we see paid usage continue to grow, the second derivative reaccelerate, and financing can truly be supported by customer cash flow—that requires a real increase in adoption, especially adoption of the most frontier models."

Mrini observed that the reason Meta has gained market recognition is that its AI investment is already reflected in improvements in ad targeting and ad performance measurement, bringing revenue growth and margin expansion. Here, compute is embedded in a business with pricing power, and efficiency gains can directly reduce the cost of service.

By contrast, Google and Microsoft are still at a stage where the investment is clearer and returns at the application layer are not yet fully reflected in the income statement. Both companies are embedding AI into search, office software, Android, enterprise collaboration and intelligent agent tools, and both are competing for the orchestration and coordination entry point for intelligent agent workflows. But what the market sees in the short term is still, first and foremost, the data center bill.

Amazon and Oracle are closer to infrastructure providers. Although Amazon has initiatives such as Alexa+ and its in-house Trainium, most of its investment still goes to AWS compute sold by the hour; Oracle is more dependent on cloud infrastructure and faces more significant financing pressure.

Reuters cited Freddy Lavric, a senior trader at Winthrop Capital Management, as saying: "Over the next two to three years, companies need to prove that AI is bringing incremental revenue, expanding margins and improving cash flow. If these financial benefits do not materialize, the market will begin to question whether this round of investment has gone too far."

The problem is that all participants believe they cannot afford to be absent, which may collectively create excess supply; as all participants increase financing intensity, market sentiment may shift from being an observer to a variable that determines the scale of investment.

The prisoner's dilemma explains why the giants keep raising their bets: everyone fears falling behind.

Reflexivity explains why the market has become fragile: when internal cash flow cannot cover new investment, stock prices, bonds, willingness to finance and capital expenditure begin to influence one another. In optimistic times, this mechanism can amplify expansion; once revenue materializes more slowly than expected, it will likewise amplify contraction.

Whether AI infrastructure ultimately becomes a high-return asset or an expensive commoditization treadmill depends on a simple yet extremely difficult question: can demand growth outrun the decline in compute prices, and can that answer be proven before the equipment invested in today is fully depreciated? Risk disclosure and disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.