The Four O's: Reading the Signs of an AI Bubble

Ruchir Sharma, chair of Rockefeller International, published a framework in the Financial Times this week that cuts through the ambient noise of market chatter. Amid what he calls “a bubble in bubble talk”—Google searches pairing AI with the b-word have surged to record levels—Sharma offers something rarer than opinion: a testable framework. His four O’s—Overvaluation, Over-ownership, Over-investment, and Over-leverage—provide a diagnostic lens honed across decades of observing manias from gold in the 1970s to dotcom in the late 1990s. Applied to artificial intelligence in December 2025, all four indicators flash warning signals. The uncomfortable question isn’t whether we’re in a bubble. It’s what happens when the bubble is also a revolution.

The timing of Sharma’s analysis proves instructive. Oracle’s disappointing Q2 FY2026 earnings on December 10 catalyzed a tech selloff that handed Wall Street one of its worst weeks in months. Despite narrowly beating earnings per share expectations, a slight revenue miss and a staggering projection of 40% higher capital expenditures for AI infrastructure—reaching $50 billion for fiscal 2026—sent shockwaves through markets. The selloff wasn’t irrational panic. It was markets beginning to ask a question they’d deferred for two years: when do the bets start paying off?

Here lies the paradox that makes AI different from tulips, railroads, or even the internet. The technology demonstrably works. ChatGPT crossed 300 million weekly active users. Coding assistants measurably accelerate software development. Autonomous systems navigate warehouses and highways. The speculative excess isn’t about whether AI functions—it’s about whether current prices already discount decades of future value creation. That distinction matters, because it means the correction, when it comes, need not invalidate the underlying thesis. Bubbles burst. Revolutions persist.

The four horsemen of market euphoria

Sharma’s framework emerges from pattern recognition across a century of market manias. Each of the four O’s captures a distinct pathology of speculative excess, and each has identifiable thresholds where historical probability shifts decisively toward correction.

Overvaluation represents the most intuitive metric: have prices risen beyond what fundamentals can justify? Sharma’s research shows that in major bubbles going back to gold in the 1970s and the internet boom of the late 1990s, inflation-adjusted prices rose tenfold over 10 to 15 years. US tech shares recently crossed that threshold. More granularly, a study of bubbles over the past century demonstrates that crash probability exceeds 50% when the industry at the heart of a mania beats the broader market by more than 100% over two years. AI-related stocks hover near that tipping point. These aren’t arbitrary numbers; they represent the historical boundary where speculation routinely overwhelms fundamentals.

The valuation landscape in December 2025 shows striking concentration. According to Fortune, in late 2025, 30% of the S&P 500 and 20% of the MSCI World index rested on just five companies—the greatest concentration in half a century. The Magnificent Seven now comprise approximately 35% of the S&P 500, exceeding dot-com concentration levels. Since the current bull market began in October 2022, roughly 75% of gains in the index have come from just seven stocks. This isn’t diversification; it’s a bet on one thesis expressed through a handful of names.

Over-ownership measures the herd behavior that transforms rational optimism into crowded trades. Sharma notes that American households now hold a record 52% of their wealth in stocks—higher than the peak in 2000 and far exceeding levels in the EU (30%), Japan (20%), and the UK (15%). Gallup data confirms that 62% of Americans report owning stock in 2025, matching last year’s reading and recovering from the sub-60% levels that persisted from 2010 to 2022. The ownership is heavily skewed: the top 10% hold a record 93% of US equities according to Federal Reserve data, while the bottom 50% own just 1%.

The behavioral symptoms intensify. Daily US share trading volume has risen 60% over five years to around 18 billion shares. The retail share of short-dated stock options has grown from one-third to more than half. On Robinhood, the five most heavily owned stocks are all Magnificent Seven members. Sharma identifies a sociological phenomenon he calls “financial nihilism”—young investors speculating aggressively because they’ve given up on buying homes through conventional saving. When the market drops on a given day, retail investors impulsively buy the next day. The trade has become reflexive, almost involuntary.

Over-investment examines capital allocation at the corporate level. Tech investment recently surpassed 6% of US GDP, exceeding the record set in 2000. The Magnificent Seven have invested $368 billion in AI-related capital expenditures this year, and Big Tech firms are on track to spend nearly $400 billion in 2025 alone. Nvidia’s management projects global data center capital expenditures will rise from $600 billion in 2025 to $3-4 trillion by 2030. AI spending has added half a percentage point to GDP growth in the first half of 2025—without it, growth would have been a mere 0.6% annualized rather than the reported 1.1%.

The spending scale beggars historical comparison. Current AI investment already exceeds the internet boom’s peak relative to GDP, and when adjusted for the shorter useful life of AI chips versus physical infrastructure, AI spending surpasses even the railroad buildout of the 1860s-1870s according to Morningstar. Meanwhile, S&P 500 companies recorded no year-over-year buyback growth as AI investment crowds out share repurchases. Capital that once returned to shareholders now flows into data centers whose returns remain uncertain.

Over-leverage historically provides the final accelerant. Sharma observes that corporations and households in the US do not yet look overleveraged—a crucial distinction from 2000 and 2008. But the Magnificent Seven are no longer the cash machines they were even a year ago. Amazon, Meta, and Microsoft are now net debtors, up from one in 2023. Only Google and Nvidia still generate the cash piles that characterized the group. The leverage has migrated to the government ledger, where record deficits pose risks if bond investors question America’s fiscal trajectory. And in financial markets, leverage has evolved beyond traditional margin loans into leveraged ETFs—funds that borrow to magnify bets—whose assets have grown sevenfold over the past decade to around $140 billion. According to National Bank Financial, these funds gathered $907 million in September 2025 alone, breaking their monthly inflow record.

Sharma’s verdict is unequivocal: to varying degrees, all four O’s suggest AI is a bubble that has reached an advanced stage. The framework doesn’t predict timing—no framework reliably does—but it establishes that by historical standards, the current moment exhibits the pathologies that precede corrections.

What the scoreboard reveals about market psychology

The evidence supporting Sharma’s framework extends beyond his specific metrics into the broader market psychology. Google DeepMind CEO Demis Hassabis observed in November: “It feels like there’s obviously a bubble in the private market. You look at seed rounds with just nothing being tens of billions of dollars. That seems a little unsustainable.” Goldman Sachs CEO David Solomon has warned of a “likely” 10-20% equity market drawdown within the next two years. The IMF and Bank of England have both sounded alarms. The wisdom of crowds has rendered its preliminary verdict.

The valuation divergence within AI tells its own story. Palantir Technologies trades at more than 180 times estimated profits; Snowflake at nearly 140 times projected earnings. OpenAI’s valuation more than tripled from $157 billion in October 2024 to $500 billion a year later. Yet Nvidia, Alphabet, and Microsoft all trade below 30 times earnings—relatively restrained multiples given the euphoria surrounding them. The market hasn’t uniformly lost its mind. It has selectively suspended disbelief for companies whose narratives outpace their numbers while maintaining discipline elsewhere.

The retail investor behavior patterns warrant particular attention. Schwab’s 2025 “ETFs and Beyond” study found that 26% of ETF investors plan to invest in leveraged products within the next year. Over 30 leveraged ETFs launched in Q2 2025 alone, many focused on individual stocks. The GraniteShares 2x Long COIN ETF surged 233% during the quarter; the Direxion Daily TSLA Bull 2X Shares ETF fell more than 56% over the first half of 2025. These instruments amplify both gains and losses, and their proliferation signals appetite for speculation that transcends reasoned analysis.

Sharma notes a peculiar psychological dynamic: AI enthusiasts argue that incessant bubble talk proves this isn’t a bubble, because peaks come when worry disappears and optimism becomes universal. The reasoning contains a kernel of truth—sentiment extremes often precede reversals—but Sharma counters that worry was in fact growing before the dotcom crash. One year before that implosion, the San Francisco Fed raised the spectre of 1929. Columnists, economists, and institutional investors echoed those fears. The presence of skepticism doesn’t inoculate markets against correction.

The “fully invested bear” represents perhaps the strangest creature in this menagerie. Because financial conditions remain loose and liquidity keeps driving up stocks, institutional investors who are skeptical of AI euphoria nonetheless keep buying. They recognize the bubble but cannot afford to step aside. Career risk—the fear of underperforming benchmarks while waiting for a correction that may take years—overwhelms conviction. When even the bears are long, who remains to buy at higher prices?

The ownership concentration creates fragility that standard diversification metrics miss. When 75% of market gains come from seven stocks, and those seven stocks all share exposure to the same thesis, the portfolio that appears diversified across hundreds of names actually concentrates risk in a single narrative. A disappointment in AI monetization doesn’t just affect Nvidia; it ripples through every Magnificent Seven constituent and every fund weighted toward them. The contagion channels exist; they await only a catalyst.

The capital expenditure trajectory poses particular questions. Companies are pouring money into AI data centers and the power plants to run them. For every survey showing demand for AI skyrocketing, another shows the opposite: fewer than 15% of US companies say they use AI, amid multiple signs that the adoption rate is slowing. Sharma quotes techno-optimists who say AI investment will pay for itself by cutting labor costs, potentially replacing up to 40% of human tasks “in the not very distant future” and pushing unemployment as high as 20%. But will humans sit by while this unfolds? Labor disruption at that scale could trigger political backlash that limits the degree to which AI investment pays off. The thesis requires not just technological success but social acceptance—a variable that spreadsheets cannot model.

The bull’s best defense

Intellectual honesty requires examining the counterarguments with the same rigor applied to the bubble case. The bulls marshal substantial evidence that this time differs from historical manias in material ways.

The most compelling distinction involves profitability. Federal Reserve Chair Jerome Powell has stated that AI differs from other technology bubbles like dotcom in that the corporations behind it are generating large amounts of revenue and that investment into AI data centers is generating large amounts of economic growth. This isn’t speculative vapor—it’s tangible output. The Magnificent Seven collectively earned hundreds of billions in profit last year. Nvidia’s revenue growth doesn’t rest on promises; it rests on chips that customers queue to purchase. When critics compare current valuations to 1999, they must acknowledge that the fundamental business quality differs categorically.

iShares analysis argues that today’s tech valuations are “far from bubble territory.” At the peak of the dotcom bubble, the top four tech leaders traded near 70 times two-year forward earnings. Today, the average two-year forward P/E for the biggest AI datacenter spenders—Microsoft, Alphabet, Amazon, and Meta—sits around 26 times. That’s elevated relative to historical averages but worlds apart from the late-1990s extremes. The comparison flatters the present moment.

The economic impact argument carries weight. AI spending adding half a percentage point to GDP growth represents real economic activity—construction jobs, equipment manufacturing, engineering employment. Goldman Sachs estimates every $1 in AI investment generates $4.90 in economic output. The multiplier effect means AI investment isn’t merely financial speculation; it’s infrastructure building with knock-on benefits throughout the economy. Data centers require steel, concrete, electrical equipment, and skilled labor. The spending creates real wealth even if the returns on that spending disappoint.

The technology adoption curve may be earlier than pessimists assume. When fewer than 15% of US companies report using AI, that statistic can be read two ways: either adoption has stalled, or the runway for growth stretches far ahead. The internet in 1998 faced similar adoption metrics; within five years, it had transformed commerce, communication, and media. Early disappointments gave way to pervasive integration. The question isn’t whether AI will be widely adopted but whether current prices already assume that outcome.

Some analysts point to the remaining buying power as evidence the mania has room to run. With $7.5 trillion still sitting in money market mutual funds, retail investors have ammunition remaining. Financial conditions remain loose. The trigger for bubble bursts, historically, has been rising interest rates and tightening financial conditions. Until the money inflating the bubble starts drying up, Sharma himself acknowledges, it could keep growing. The Fed’s rate trajectory thus matters more than any valuation metric.

The structural argument deserves consideration: AI isn’t a single company or even a single technology. It’s a general-purpose capability that enhances productivity across every sector. Unlike the dotcom boom, which produced many companies with no viable business model, AI companies today serve paying customers solving real problems. Microsoft’s Copilot integrates into enterprise workflows. Anthropic’s Claude assists with coding, analysis, and research. Google’s Gemini powers search and advertising optimization. The revenue models exist; the question is merely scale.

The global competitive dynamic adds another dimension. America cannot afford to underinvest in AI while China races ahead. The geopolitical imperative justifies spending that might otherwise seem excessive. Nvidia CEO Jensen Huang has called this a “Sputnik moment” for technology investment. When national security considerations enter the calculus, traditional return-on-investment analysis becomes secondary to maintaining capability. The spending may prove economically inefficient but strategically necessary—a cost of competition rather than speculation.

Finally, the self-awareness of market participants suggests the mania may be more self-limiting than predecessors. MIT Technology Review quotes an industry observer: “Everyone in tech agrees we’re in a bubble. They just can’t agree on what it looks like—or what happens when it pops.” That universal acknowledgment differs from historical episodes where denial persisted until the crash. When everyone expects a correction, does the correction still come? Markets have a perverse way of confounding consensus expectations.

Revolution outlasts the reckoning

The binary framing—bubble or not?—obscures what matters most. Both propositions can be simultaneously true: current valuations may be unsustainable, and AI may nonetheless represent a transformational technology that reshapes industries, economies, and societies over the coming decades. The correction, when it arrives, will punish excess without invalidating the underlying revolution.

History provides the template. The dotcom crash destroyed trillions in market value and wiped out hundreds of companies. But Amazon, which fell 94% from peak to trough, went on to become one of the most valuable companies in history. Google, founded during the bubble years, emerged as a dominant force in information access. The internet transformed commerce, communication, entertainment, and work—not despite the crash but through it. The speculative excess funded infrastructure that proved valuable; the correction merely reassigned ownership and reset expectations.

AI follows a similar trajectory. The data centers being built today will serve AI workloads for years regardless of who owns them after a correction. The talent being trained in machine learning doesn’t forget their skills when stock prices fall. The research advancing model capabilities continues independent of market sentiment. The bubble, if it bursts, will destroy paper wealth and punish those who bought at peaks. It will not unmake the technology.

The capabilities already demonstrated make regression impossible. Language models that draft legal contracts, generate code, analyze medical images, and accelerate drug discovery represent genuine advances that organizations will not abandon. The question of adoption timing—whether AI transforms work in two years or ten—affects valuations but not the fundamental trajectory. Once a technology crosses the threshold of practical utility, it diffuses through the economy according to its own logic.

The prudent approach acknowledges both the bubble risks and the revolutionary potential. For operators navigating this environment, several principles emerge:

Position for volatility without exiting the field. The correction, when it comes, may be severe. Oracle’s 40% projected capex increase spooked markets; imagine the reaction to an actual earnings miss or guidance cut. But sitting entirely on the sidelines means forfeiting the productivity gains AI enables. The middle path involves maintaining exposure while managing position sizes and avoiding leverage.

Distinguish between the thesis and the trade. AI will transform industries. Whether any particular AI stock at any particular price represents a good investment is a separate question. The revolutionary thesis doesn’t require every AI company to succeed or every valuation to prove justified. Being right about the technology and wrong about the timing can be financially indistinguishable from being wrong about everything.

Monitor the leverage indicators. Sharma identifies the trigger for bubble bursts as rising rates and tightening financial conditions. The Fed’s trajectory matters more than analyst price targets. When monetary conditions shift, the repricing happens quickly. The leveraged ETFs that amplify gains on the way up amplify losses on the way down. Position liquidity becomes paramount when the tide reverses.

Invest in capabilities, not just equities. For enterprises, the more durable play may be building internal AI competency rather than making financial bets on AI stocks. Training teams to use AI tools, experimenting with applications, and developing institutional knowledge creates value regardless of market fluctuations. The skills transfer; the stock certificates may not.

Maintain historical humility. Every generation believes its transformational technology differs from predecessors. Sometimes that belief proves correct; often it proves only partially so. The internet was transformational and the dotcom bubble was real. Both truths coexisted. AI can reshape civilization and AI stocks can be overvalued. The cognitive error is assuming one proposition excludes the other.

The four O’s framework provides a useful diagnostic but not a precise prediction. Sharma’s conclusion—that we are clearly in a bubble but cannot know when it bursts—represents the honest limit of analysis. Markets remain irrational longer than portfolios remain solvent. The bubble could continue growing for months or years before the reckoning arrives. Or the catalyst could emerge next quarter.

What remains unambiguous is that artificial intelligence has crossed the threshold from speculative potential to practical reality. The models work. The applications multiply. The productivity gains materialize. No market correction undoes those facts. The investors who suffered through the dotcom crash but held Amazon or bought Google in the aftermath were rewarded beyond any reasonable expectation. The same dynamic may unfold with AI—but identifying in advance which companies will be the Amazons and which the Pets.coms requires luck as much as analysis.

Bubble or not, the revolution is here. The prices will adjust, perhaps violently. The technology will remain. For those building with AI rather than merely betting on it, the distinction may be the only one that ultimately matters. Markets price securities; history judges capabilities. The four O’s warn us that a correction approaches. They cannot tell us it will reverse the transformation already underway.

The soap bubble is beautiful precisely because we know it cannot last. Its iridescent surface catches the light even as surface tension fights a losing battle against entropy. Eventually, inevitably, it pops. But the air that inflated it doesn’t vanish—it simply returns to the atmosphere, ready to fill the next bubble, and the next. Markets work the same way. The capital currently concentrated in AI will eventually redistribute, some to better investments, some to losses, some to the real economy. The technology will persist through every reallocation. That’s not a prediction about prices. It’s a recognition that revolutions, once begun, proceed on their own terms. The reckoning will come. The revolution will continue.