Is the AI Stock Bubble Just the Dot-Com Bubble 2.0?

Metrics at a Glance: 1999 vs. Today

Before diving into the details, it helps to anchor the discussion with the headline numbers. The table below compares four dimensions that matter most when assessing whether a market cycle is driven by fundamentals or speculation.

At first glance, the numbers suggest we are in a fundamentally healthier position. Nvidia trades at a fraction of Cisco's peak multiple. The companies driving AI infrastructure are profitable, cash-rich incumbents rather than speculative startups. But surface-level comparisons can be misleading. The details matter, and that is where the real debate lives.

1. The Dot-Com Bubble: Anatomy of a Crash

To understand whether we are repeating history, you first need to understand what actually happened between 1995 and 2002. The standard shorthand—“investors got excited about the internet, valuations went crazy, then everything crashed”—captures the vibe but misses the mechanics.

The dot-com era was built on a genuinely transformative technological shift: the mass adoption of the internet. Between 1995 and 2000, internet users in the United States grew from roughly 16 million to over 120 million. The Nasdaq Composite rose from under 1,000 in 1995 to a peak of 5,048 on March 10, 2000. It was the fastest accumulation of paper wealth in modern financial history, and it happened because the underlying thesis—the internet will change everything—was correct. The problem was timing, not direction.

The telecom industry alone laid approximately 80 million miles of fiber optic cable during the bubble years, most of it driven by companies like WorldCom, Global Crossing, and Qwest. The total capital expenditure across the telecom sector exceeded $500 billion. When the bubble burst, an estimated 95% of that fiber sat unused—what industry insiders called “dark fiber.” It was eventually lit, but not until 2004–2007, when YouTube, Netflix streaming, and cloud computing created actual demand for bandwidth at scale. The infrastructure was built for a future that arrived five years late, and the companies that built it went bankrupt waiting.

At the stock level, the numbers are staggering: Cisco hit a market cap of $555 billion in March 2000 on $18.9 billion in trailing revenue, implying a price-to-sales ratio above 29x. Pets.com raised $82.5 million in its IPO despite having $619,000 in revenue and a business model that lost money on every shipment. Webvan burned through $800 million building automated grocery warehouses before filing for bankruptcy in 2001. The common thread was not just high valuations but the complete absence of a path to profitability. When the Federal Reserve raised rates from 4.75% to 6.5% between mid-1999 and mid-2000, the cost of capital exposed every business model that depended on infinite cheap funding.

By October 2002, the Nasdaq had fallen 78% from its peak to 1,114. Approximately $5 trillion in market value was erased. It took the index 15 years—until 2015—to reclaim its 2000 high. Not all companies died: Amazon survived by raising $672 million in convertible bonds just months before the crash, and emerged as one of the most valuable companies in history. But for every Amazon there were dozens of eToys, Boo.com, and TheGlobe.com—companies remembered only as footnotes in market history textbooks.

2. The AI Landscape: What Makes This Cycle Different

The current AI cycle shares the core feature of the dot-com era: it is built on a genuinely transformational technology. Large language models, diffusion models for image and video generation, and multimodal AI systems represent a step change in what software can do. Unlike the internet of the 1990s—where the user experience was slow, clunky, and confined to desktop computers—today's AI products are already in the hands of hundreds of millions of users. ChatGPT reached 100 million monthly active users faster than any consumer application in history, outpacing TikTok, Instagram, and every other platform.

The investment numbers reflect this conviction. In 2024, the four largest US hyperscalers—Microsoft, Amazon, Google, and Meta—spent a combined $230 billion on capital expenditures, the majority directed toward AI infrastructure: GPU clusters, data center expansion, and power infrastructure. For 2025, consensus estimates project that figure rising above $320 billion. To put that in perspective, $320 billion is roughly the annual GDP of Chile or the entire market capitalization of Coca-Cola. This is not venture capital speculation; it is the largest concentrated industrial investment cycle since the build-out of the interstate highway system.

Nvidia sits at the center of this spending. The company reported $130 billion in revenue for its fiscal year 2025, with data center GPU sales accounting for the vast majority. Its gross margins exceeded 75%, a level typically associated with enterprise software companies, not hardware manufacturers. Nvidia's market capitalization crossed $3 trillion in 2024, briefly making it the most valuable company in the world. The bull case is straightforward: Nvidia is selling the picks and shovels for a gold rush that has barely begun. The bear case: $3 trillion implies that AI infrastructure spending continues growing at 30%+ annually for the better part of a decade with no competitive erosion of margins—an assumption that history suggests is aggressive.

3. The Capex Parallel: Fiber Optics vs. GPU Data Centers

This is the comparison that drives the most urgent bubble debate, and for good reason. The structural similarity is impossible to ignore: in both eras, the dominant narrative was that massive infrastructure investment was necessary to support an inevitable technological future. In both eras, the bottleneck was physical rather than digital. In both eras, the companies building the infrastructure saw their stock prices decouple from traditional valuation frameworks as investors priced in a decade of growth in a single year.

During the dot-com bubble, telecom companies argued that internet traffic was doubling every 100 days—a claim that AT&T executives repeated in earnings calls and investor presentations. The logic was seductive: if traffic keeps doubling, you can never have enough fiber. The flaw was that while internet traffic did grow exponentially, it grew from a very small base, and the doubling rate slowed as the base expanded. By the time the fiber was lit, the companies that laid it had already restructured or liquidated.

Today, the equivalent argument is that AI model scaling laws demand exponentially more compute. The release of each new generation of frontier models—from GPT-3 to GPT-4 to Claude 3.5 to Gemini Ultra—has required roughly 3x to 10x the training compute of its predecessor. If that scaling trend continues, the demand for GPUs appears insatiable. OpenAI CEO Sam Altman has publicly discussed plans for data centers requiring 5 gigawatts of power—roughly the output of five nuclear reactors—dedicated to a single training run. Whether these ambitions materialize or not, the narrative drives investment decisions today.

But there is a critical distinction between fiber overbuild and GPU overbuild that many commentators miss: fiber has a 20+ year useful life with near-zero marginal cost once deployed, while GPUs depreciate rapidly. An H100 GPU cluster purchased in 2024 for $300 million may be worth half that in 2026 when Nvidia's next-generation Rubin architecture ships with 4x the performance per watt. The depreciation risk means that if enterprise AI revenue takes longer to materialize than expected, the assets lose value at a pace that fiber never did. In the dot-com era, the fiber was still there, waiting to be used. In the AI era, last year's GPUs are already obsolete.

4. The Valuation Debate: Expensive, But How Expensive?

Valuation is where the dot-com comparison gets complicated—and where reasonable people can look at the same numbers and reach opposite conclusions. The most commonly cited defense of current AI valuations is that the dominant players are far cheaper on traditional multiples than their dot-com counterparts. That is true. It is also incomplete.

Cisco vs. Nvidia. At its March 2000 peak, Cisco traded at approximately 150x trailing earnings and 29x trailing sales. By any standard, these were absurd multiples for a hardware company. Nvidia, by contrast, currently trades at around 30–45x trailing earnings and 20–25x trailing sales—still expensive for a semiconductor company (the industry median is roughly 20x earnings) but nowhere near Cisco's mania levels. On forward estimates, Nvidia looks even more reasonable: consensus expects earnings to grow over 50% in the next year, implying a forward P/E in the mid-20s. If those estimates prove accurate, Nvidia is arguably cheap.

But this is where the analysis needs to go deeper. The question is not whether Nvidia is cheaper than Cisco was. The question is whether the earnings growth that justifies Nvidia's current multiple is sustainable. Nvidia's revenue grew from $27 billion in fiscal 2023 to $130 billion in fiscal 2025—a nearly 5x increase in two years. That growth rate is unprecedented for a company of Nvidia's scale. If growth reverts to anything resembling the semiconductor industry average of 5–10% annually, the current valuation is difficult to justify. The bull case requires believing that we are still in the early innings of a decade-long AI infrastructure build-out, not the middle of a cyclical capex peak.

Beyond Nvidia, the broader AI ecosystem presents its own valuation puzzles. Microsoft trades at roughly 33x earnings, a premium to its 10-year average of 28x, with much of the premium attributed to its AI leadership position via OpenAI. Palantir, which has repositioned itself as an AI platform company, briefly traded above 200x earnings in late 2024 before correcting. The “Magnificent Seven” tech stocks—Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla—collectively accounted for over 60% of the S&P 500's gains in 2024, a level of concentration not seen since the Nifty Fifty era of the early 1970s.

5. The Balance Sheet Argument: Why This Time Might Be Different

The single strongest argument against the bubble thesis is the quality of the balance sheets driving the current cycle. In 1999, the companies spending the most on internet infrastructure were, by and large, unprofitable startups and leveraged telecom operators. They depended on continuous access to capital markets—equity offerings, convertible bonds, vendor financing—to fund their operations. When the capital markets closed in 2000, they had no internal resources to fall back on.

Today, the picture is fundamentally different. Microsoft generated over $85 billion in free cash flow in its most recent fiscal year. Alphabet generated $70 billion. Meta generated $50 billion. Amazon, despite its notoriously thin retail margins, generated over $40 billion in free cash flow driven by AWS. These are not companies that need to raise money to survive a downturn. They are companies that could continue funding AI infrastructure at current levels for years even if their AI businesses generate zero incremental profit—funded entirely by their legacy cloud, advertising, and enterprise software businesses.

Consider Meta's position. The company has committed to spending $60–65 billion on capex in 2025, primarily on AI infrastructure. Critics call this reckless. But Meta generated $52 billion in free cash flow in 2024 after significant AI spending. Its net cash position exceeds $40 billion. If AI spending proves unproductive, Meta can simply reduce capex and return to generating enormous cash flows from its advertising business. The same logic applies to Microsoft (Office, Azure, Windows), Alphabet (Search, YouTube, Cloud), and Amazon (AWS, Retail). Each has a “day job” that prints money.

This is why the balance-sheet argument is so powerful: the dot-com crash was not primarily a valuation correction—it was a liquidity crisis. Companies ran out of money and went bankrupt, triggering cascading defaults across the telecom supply chain. In the current cycle, even if AI valuations correct significantly, the core companies have the financial resources to absorb the shock without existential risk. A 50% drawdown in Nvidia would be painful for shareholders but would not trigger the kind of systemic credit event that defined the 2000–2002 unwind.

6. Market Concentration: The Hidden Vulnerability

If the balance sheets are the bull case, market concentration is the bear case that keeps portfolio managers up at night. The AI investment thesis depends on a remarkably narrow set of assumptions about a remarkably narrow set of companies.

Consider the supply chain: virtually all advanced AI training chips are designed by Nvidia and manufactured by TSMC. Nvidia's CUDA software ecosystem, built over 15 years, creates a moat that competitors have struggled to cross. AMD and Intel are investing heavily in AI accelerators, and hyperscalers like Google (TPUs) and Amazon (Trainium) are developing custom silicon, but Nvidia's market share in data center AI chips remains above 80%. A disruption to Nvidia—whether from competitive pressure, geopolitical restrictions on chip exports, or simply a cyclical slowdown in hyperscaler procurement—would reverberate through the entire AI investment ecosystem.

On the demand side, the concentration is equally stark. Four companies—Microsoft, Amazon, Google, and Meta—account for the vast majority of AI infrastructure spending. Oracle, Tesla, and a handful of sovereign wealth funds and state-backed entities round out the buyer base. This is not a broad-based industrial cycle with thousands of independent purchasers making decentralized decisions. It is a highly concentrated procurement cycle where a shift in strategy at any one of the top four buyers could meaningfully impact the entire supply chain.

This concentration creates a specific kind of risk that was less pronounced in the dot-com era: correlated decision-making. If Microsoft's board concludes that AI capex is not generating adequate returns and decides to slow spending, does Meta follow? Does Amazon? These companies watch each other closely, and capex decisions at the hyperscalers are often influenced by competitive dynamics as much as by pure ROI calculations. The fear of falling behind can sustain spending for years. But once the spending slows, the deceleration can be sharp and synchronized—and the companies further down the supply chain (Nvidia, memory manufacturers, data center operators, power infrastructure providers) would feel the impact simultaneously.

7. The Revenue Question: Where Is the ROI?

This is the single most important question in the entire AI investment debate, and it does not yet have a clear answer. In the dot-com era, the answer became clear in retrospect: there was essentially no revenue to justify the investment. Pets.com generated less than $1 million in revenue against hundreds of millions in market cap. By the time the bubble burst, it was obvious that the emperor had no clothes.

Today, the picture is more nuanced. Microsoft reported that its Azure cloud business grew 33% in the most recent quarter, with 13 percentage points of that growth attributed to AI services. That is real revenue—roughly $3–4 billion per quarter in incremental AI-driven cloud revenue, annualizing to $12–16 billion. Microsoft's GitHub Copilot has over 1.8 million paid subscribers generating what analysts estimate to be $400–500 million in annual recurring revenue. OpenAI itself is reported to have reached an annualized revenue run rate above $5 billion in 2025, up from $1.6 billion a year earlier.

But here is the problem: $16 billion in annual AI revenue against $320 billion in annual AI capex is a 5% return on invested capital. Even if you assume the infrastructure has multiple years of useful life, the payback period on current spending levels stretches well beyond what most corporate boards would tolerate without evidence of accelerating customer adoption. For context, the median S&P 500 company generates a return on invested capital of approximately 14%. The AI infrastructure build-out, at current revenue levels, is generating returns well below that threshold.

The bull response is that we are measuring too early. AI revenue is growing at triple-digit rates from a small base, and the infrastructure being built today is for the applications of 2027 and 2028, not 2025. This is the same argument made by telecom executives in 1999, and it was correct in direction—the demand did arrive, just not in time to save the companies that built the infrastructure. The difference this time, as discussed in Section 5, is that the companies doing the building can afford to wait. Microsoft and Google will not go bankrupt if AI revenue takes three more years to scale. But their shareholders may not be patient, and the stock prices that embed expectations of rapid, sustained growth could correct significantly even without a liquidity crisis.

There is also a subtler revenue challenge: the unit economics of AI inference remain challenging for many use cases. Running a large language model to answer a customer service query costs roughly 10–100x more than a traditional keyword-based chatbot. While AI responses are better, they are rarely 10–100x better from the customer's perspective. Enterprises are experimenting with AI, but converting experiments into seven-figure annual contracts requires demonstrable ROI that many AI applications have not yet delivered at scale. Until inference costs come down dramatically—which they are, driven by hardware improvements and model optimization—or until AI applications deliver dramatically more value, the revenue curve may be shallower than the infrastructure build-out implies.

8. The Open-Source Wildcard and Talent Constraints

Beyond the financial metrics, two structural factors are reshaping the AI investment landscape in ways that have no dot-com era parallel: the rise of open-source AI models and the acute scarcity of AI research talent.

Open-source models are compressing the value of proprietary AI. When Meta released Llama 2 as an open-weight model in mid-2023, it fundamentally altered the economics of the AI industry. Suddenly, any company with sufficient engineering resources could download, fine-tune, and deploy a capable large language model without paying per-token API fees to OpenAI, Anthropic, or Google. Subsequent releases—Mistral, Falcon, Llama 3, DeepSeek, Qwen—have continued to close the performance gap between open and closed models. DeepSeek's V3 model, released in late 2024 by a Chinese quantitative hedge fund, reportedly achieved performance competitive with GPT-4 at a fraction of the training cost, sending shockwaves through the industry and contributing to a single-day 17% decline in Nvidia's stock price.

The open-source dynamic creates a paradox for AI investors. On one hand, widespread availability of capable AI models accelerates adoption and creates demand for infrastructure—good for Nvidia and the hyperscalers. On the other hand, it commoditizes the model layer, compressing the margins of companies like OpenAI and Anthropic that have raised tens of billions of dollars on the premise that they would maintain a durable technological lead. If the best AI models are free and open, the value accrues to infrastructure providers and application-layer companies rather than model developers. This is fundamentally different from the dot-com era, where proprietary technology (Cisco's routers, Sun Microsystems' servers) maintained pricing power even as the broader ecosystem expanded.

The talent bottleneck is equally significant. The global supply of researchers and engineers capable of training frontier AI models is estimated at fewer than 10,000 people, concentrated overwhelmingly at a handful of companies: Google DeepMind, OpenAI, Anthropic, Meta AI, and a few well-funded startups. Compensation for top researchers has reached $5–10 million annually, a level that even well-capitalized startups struggle to sustain. This concentration of talent acts as both a moat for the incumbents and a ceiling on how fast the broader AI industry can advance. You cannot simply spend more money to produce more AI breakthroughs; you need people who understand transformer architectures, reinforcement learning from human feedback, and distributed training at massive scale, and those people are in extraordinarily short supply.

The talent constraint also introduces a fragility that was less present in the dot-com era. In 1999, if a talented engineer left Cisco, there were thousands of networking engineers available to replace them. In 2025, if a core research team leaves OpenAI, there may be no one on the planet with equivalent expertise who is available for hire. This concentration risk means that the AI industry's trajectory depends on the decisions and stability of a remarkably small group of individuals—a structural vulnerability that traditional financial analysis does not easily capture.

9. Infrastructure Bottlenecks: Power, Chips, and the Physical World

One of the most underappreciated dimensions of the AI build-out is the extent to which it is constrained by physical infrastructure, not just capital. Unlike software, which can scale near-infinitely at near-zero marginal cost, AI infrastructure is fundamentally a physical business: chips must be manufactured in fabrication plants that take 3–5 years to build, data centers require concrete and steel and cooling systems, and everything requires electricity.

The power grid is emerging as the binding constraint. A single state-of-the-art AI data center can consume 500 megawatts or more—roughly the electricity consumption of 400,000 homes. Northern Virginia, the world's largest data center market, is facing severe transmission constraints, with utility Dominion Energy warning that new large-scale data center connections may face multi-year delays. In Ireland, data centers already consume over 20% of the country's electricity, and the grid operator has imposed a de facto moratorium on new connections in the Dublin area. Similar constraints are emerging in Singapore, the Netherlands, and key data center corridors in the American West.

These physical constraints have an interesting implication for the bubble debate: they may prevent the kind of unfettered overbuild that characterized the fiber era. In the late 1990s, if you had capital, you could trench fiber. The physical barriers were minimal—construction crews and cable were readily available, and right-of-way permits were relatively easy to obtain. Today, if you have capital and want to build a gigawatt-scale AI data center, you need years of regulatory approval, grid interconnection agreements, environmental impact assessments, and possibly dedicated power generation. The bureaucracy of physical infrastructure acts as a natural brake on speculative overinvestment.

On the semiconductor side, the constraint is manufacturing capacity. TSMC, which manufactures essentially all advanced AI chips, operates a handful of fabrication plants in Taiwan, with new facilities under construction in Arizona, Japan, and Germany. Each fab costs $20–40 billion and takes 3–5 years to reach production capacity. Nvidia's ability to ship GPUs is ultimately limited by TSMC's ability to manufacture them, and TSMC's capacity is allocated years in advance. This is another difference from the dot-com era: the supply of the critical infrastructure component (advanced logic chips) is not elastic in the short to medium term, which supports pricing power but also caps the speed at which the infrastructure build-out can proceed.

10. Geopolitics: The Wild Card

No analysis of the AI investment cycle is complete without addressing geopolitics, which introduces risks that simply did not exist in the dot-com era. The US-China technology competition has turned advanced semiconductors into a strategic asset subject to export controls, sanctions, and supply chain decoupling.

In October 2022, the US Commerce Department imposed sweeping export controls on advanced AI chips to China, effectively barring Nvidia from selling its H100 and A100 GPUs to Chinese customers. Subsequent rounds of controls in 2023 and 2024 further tightened the restrictions, closing loopholes and expanding the scope to include chips with slightly lower performance thresholds. China, in response, has accelerated its domestic chip development efforts, with Huawei's Ascend series emerging as the most credible alternative to Nvidia's offerings—though still 2–3 generations behind in performance.

The geopolitical dimension creates several distinct risks for the AI investment thesis. First, it segments the global market: Chinese AI companies cannot access the most advanced hardware, which limits their ability to compete but also creates a parallel AI ecosystem that may eventually produce competitive alternatives. Second, it introduces tail risk around Taiwan, where TSMC manufactures the chips that the entire AI industry depends on. Any disruption to TSMC's production—whether from natural disaster, political conflict, or extended supply chain dislocation—would have immediate and severe consequences for the global AI infrastructure build-out. Third, export controls could escalate: if the US extends restrictions to additional countries or to additional categories of technology, the addressable market for American AI hardware could shrink, impacting the growth assumptions embedded in current valuations.

These risks are difficult to quantify but impossible to ignore. The dot-com era was fundamentally about domestic technology adoption within a relatively stable geopolitical environment. The AI era unfolds against a backdrop of great-power competition that could reshape the industry's trajectory in ways that have nothing to do with the technology itself.

11. Scenarios: How This Could Play Out

Rather than offering a binary prediction—bubble or not bubble—it is more useful to consider the range of plausible scenarios and their implications for investors. Here are four paths that the current AI cycle could follow.

Scenario A: The Productivity Boom (Bull Case). Enterprise AI adoption accelerates through 2026–2028 as inference costs fall and AI agents become capable of handling complex, multi-step business processes. Corporate AI revenue grows to hundreds of billions annually, justifying the infrastructure investment. Nvidia's growth decelerates but stabilizes at healthy levels, and the broader AI ecosystem—cloud providers, enterprise software vendors, AI-native startups—generates enough value to sustain elevated valuations. The S&P 500 broadens as AI-driven productivity gains boost margins across sectors beyond technology. Probability: 20%.

Scenario B: The Soft Landing (Base Case). AI revenue grows meaningfully but more slowly than the most optimistic projections. Hyperscaler capex peaks in 2026–2027 and then normalizes as the most urgent infrastructure build-out is completed. Nvidia's growth rate compresses from triple digits to 15–20%, and its multiple contracts to reflect the maturation of the AI hardware cycle. Tech stocks experience a multi-year period of sideways returns as earnings catch up to valuations—similar to what Microsoft and Cisco experienced between 2000 and 2003, but without the 80% drawdowns. Capital flows out of mega-cap tech and into other sectors, driving a rotation rather than a crash. Probability: 45%.

Scenario C: The Correction (Bear Case). Enterprise AI adoption disappoints relative to the infrastructure investment. Hyperscaler capex decelerates sharply as boards demand evidence of ROI that has not materialized. Nvidia's revenue declines 30–40% as orders are cut, and its multiple contracts from 30x to 15x, implying a 60%+ drawdown from peak. The correction is contained within the tech sector—the banking system is not exposed in the way it was during the 2008 financial crisis—but the wealth destruction is significant. The S&P 500 declines 20–30%, driven almost entirely by the mega-cap tech names that drove it up. Probability: 25%.

Scenario D: The Systemic Event (Tail Risk). A geopolitical shock—major conflict involving Taiwan, severe escalation of US-China technology restrictions, or a financial event triggered by AI-related credit losses—turns the AI correction into a broader market crisis. This scenario is unlikely but non-trivial given the concentration of the semiconductor supply chain and the geopolitical premium embedded in the current environment. Probability: 10%.

Conclusion: Different Song, Similar Rhythm

After walking through the evidence, the most defensible conclusion is that the AI investment cycle is not a replay of the dot-com bubble, but it carries enough structural similarities to warrant caution. The companies driving the current cycle are financially stronger. The technology is more advanced and has clearer near-term applications. The physical infrastructure constraints provide a natural ceiling on speculative overinvestment. These are meaningful differences that should prevent the kind of systemic collapse that followed the dot-com peak.

However, three factors should give investors pause. First, the concentration of AI spending among four hyperscalers means that a shift in strategy at any one of them—let alone a synchronized pullback—would have outsized effects on the entire ecosystem. Second, the gap between infrastructure investment and demonstrated revenue is large and widening, and the history of technology cycles suggests that this gap eventually closes, usually in a way that is uncomfortable for late-cycle investors. Third, the geopolitical dimension introduces risks that were absent in previous technology cycles and that are inherently difficult to price into valuation models.

The most likely outcome, in our assessment, is a version of Scenario B: the AI build-out continues for several more years, revenue grows but more slowly than the infrastructure build-out implies, and valuations compress gradually rather than collapsing suddenly. This is not a prediction of smooth sailing—a 20–30% drawdown in AI-exposed stocks at some point in the next three years would be entirely consistent with historical patterns of technology adoption cycles. But it is a prediction that the current cycle ends with a correction, not a crash.

The dot-com comparison is useful not because it predicts what will happen, but because it reminds us of what can happen when infrastructure investment runs ahead of application-layer demand. In 1999, the infrastructure builders went bankrupt while the applications they enabled—streaming video, cloud computing, social media—created trillions in value for the companies that survived. The lesson for today's investors is not that AI is a bubble. It is that the companies building the picks and shovels may not be the ones that capture the most value, and that timing matters as much as conviction.

For those tracking this story as it unfolds, we recommend watching six specific indicators over the coming quarters. First, hyperscaler capex guidance for 2026 and 2027—any deceleration in the growth rate of planned spending would be the earliest signal that the infrastructure cycle is maturing. Second, enterprise AI revenue disclosures in earnings calls from Microsoft, Salesforce, ServiceNow, and Adobe—these are the best real-time proxies for whether businesses are actually paying for AI at scale. Third, Nvidia's data center revenue growth rate—as the base gets larger, the growth rate will naturally decelerate, but a move below 20% year-over-year would signal that the hyperscaler procurement cycle has peaked. Fourth, the performance gap between open-source and proprietary AI models—if open models match or exceed GPT-5 or Claude 4 within six months of release, the economic moat of the model companies narrows dramatically. Fifth, the number of corporate AI use cases that have moved from pilot to production at scale—anecdotes are plentiful, but systematic deployment data remains scarce. Sixth, and perhaps most importantly, inference costs—as these decline, the addressable market expands, potentially unlocking the revenue growth that justifies the infrastructure investment.

The most dangerous phrase in investing is “this time is different.” The second most dangerous is “this is just like last time.” The AI investment cycle is neither a repeat of 1999 nor a clean break from history. It is a new chapter in a recurring pattern—transformative technology attracting enormous capital, with outcomes determined less by the quality of the technology than by the relationship between investment and revenue, between ambition and evidence, between the stories we tell about the future and the numbers that arrive each quarter.