Is the AI Bubble Bursting? A Sober Look at the Concerns Behind the Boom

Introduction: A Hint of Worry Amid the Euphoria

"The AI bubble is bursting—we just haven't realized it yet." This slightly provocative statement has recently sparked heated debate on tech communities like Hacker News. While it's merely an opinion piece, it touches the most sensitive nerve in today's tech industry: are we standing at the peak of an unprecedented AI investment bubble?

Over the past two years, generative AI has reshaped the entire tech industry landscape at a staggering pace. From ChatGPT igniting a global frenzy, to tech giants committing tens of billions of dollars in capital expenditure, to countless startups receiving sky-high valuations simply by attaching the "AI" label—this boom seems unstoppable. But history has repeatedly shown that every technological revolution is accompanied by cyclical overinvestment followed by value reassessment. This article attempts a sober analysis of the AI bubble thesis.

The Core Logic Behind the AI Bubble Argument

The Severe Imbalance Between Capital Expenditure and Returns

The most direct indicator for determining whether an AI bubble exists is the match between input and output. Currently, tech giants like Microsoft, Google, Meta, and Amazon are spending hundreds of billions of dollars annually on AI infrastructure—primarily for GPU procurement and data center construction. In 2024 alone, the combined capital expenditure of America's four largest tech giants is expected to exceed $200 billion, with a significant portion directed at AI-related infrastructure. For reference, this figure already surpasses the total annual infrastructure investment of many mid-sized economies.

However, the direct commercial returns corresponding to these massive investments remain relatively unclear. The classic metric for measuring investment efficiency—Return on Invested Capital (ROIC)—faces unique challenges in the AI domain: because AI infrastructure has long and uncertain return cycles, traditional DCF (Discounted Cash Flow) models struggle to price it accurately. This forces investors to rely more on "narrative valuation" rather than "financial valuation"—which is precisely the fertile ground on which bubbles form. On the GPU supply side, NVIDIA's data center business revenue grew over 200% year-over-year in fiscal year 2024, but the extent to which this growth reflects genuine end-user demand versus corporate "stockpiling" behavior remains a point of contention.

Most enterprises' AI applications remain at the "nice-to-have" stage—they enhance product experiences but haven't yet proven they can generate revenue growth commensurate with the investment. When capital markets begin demanding answers to "where was the money spent, and how much came back," valuation logic may face a reckoning.

"Bubble" Doesn't Equal "Scam"

It's important to emphasize that the AI bubble thesis does not deny the value of artificial intelligence technology itself. The dot-com bubble of 2000 was the same: the internet truly did change the world, but that didn't prevent countless companies from collapsing due to inflated valuations. In the two years following the NASDAQ peak in March 2000, the index fell nearly 80%, and hundreds of internet companies went bankrupt—yet Amazon, eBay, and other companies with genuine business models ultimately survived and thrived. The long-term value of a technology and the short-term price of its assets are two dimensions that need to be distinguished.

The real question is: has the market paid too high a "present price" for AI's "future potential"? When expectations outrun reality, corrections are often unavoidable. In finance, this phenomenon is called "excessive discounting of forward earnings"—the market prices in returns that may take 5-10 years to materialize at an extremely low risk discount, and once the pace of realization falls short of expectations, valuation corrections follow.

Two Voices: For and Against the AI Bubble Thesis

The Bears' Concerns

Cautious observers point to several typical bubble signals:

• The circular financing problem: Some AI chip, cloud service, and model companies have formed complex capital and procurement interdependencies, with revenue "spinning" within the ecosystem, potentially inflating actual market demand. Specifically, large cloud providers offer AI startups cloud computing credits as part of their investment; these startups then use those credits to purchase cloud services—which get counted as revenue on the cloud provider's income statement. A similar pattern appeared during the 2000 telecom bubble in the form of "capacity swaps": telecom companies purchased network capacity from each other, causing paper revenues to surge while actual market demand was far below those levels. When this cycle breaks, once-impressive revenue growth can collapse within a single quarter.

• Weak monetization of consumer AI applications: Despite surging user numbers, the proportion of users willing to pay continuously for AI features remains limited. While ChatGPT set the record for fastest user growth in history, its paid user conversion rate still lags behind traditional SaaS products. More critically, AI features are being "bundled for free" by major platforms, further suppressing willingness to pay for standalone AI applications. Users have grown accustomed to using AI features for free within search engines, email clients, and office suites, leaving little motivation to pay extra for standalone AI tools.

• Computing hardware depreciation pressure: GPUs and other hardware iterate rapidly, putting massive capital expenditures at risk of swift devaluation. Looking at NVIDIA's product line—from A100 to H100 to B200—each generation achieves roughly 2-3x improvement in performance-per-watt, with iteration cycles of only 12-18 months. This means the GPU clusters that companies spend billions of dollars acquiring today will be dramatically outperformed in cost-efficiency by next-generation products within two years. Under the typical 3-5 year depreciation cycle for data center equipment, these assets may become economically "obsolete" before their accounting lifespan ends. For capital-intensive AI infrastructure investments, this technological depreciation constitutes a hidden but significant financial risk—similar to airlines being forced to write down the value of older fleets when more fuel-efficient new aircraft models enter service.

The Bulls' Rebuttal

The other side argues that simply comparing the current situation to a tech bubble is premature:

• AI capabilities are still improving exponentially, with each generational leap opening new application scenarios. From GPT-3 to GPT-4 to the latest multimodal models, each upgrade represents not just quantitative but qualitative change—new reasoning abilities, multimodal understanding, and long-context processing enable previously impossible applications (such as complex code engineering, scientific research assistance, and multi-step workflow automation). This "capability ladder" effect means the addressable market isn't fixed but continuously expands with technological progress.

• Unlike the 2000 dot-com bubble, today's leading companies mostly have real, robust cash flows to support their investments. Microsoft's annual free cash flow exceeds $70 billion, and Google parent Alphabet's exceeds $60 billion. While these companies' AI investments are aggressive, they're far from threatening their financial security. They're essentially using profits from mature businesses to make strategic investments in future platforms—fundamentally different from the internet companies of 2000 that had zero revenue yet commanded multi-billion-dollar market caps.

• Enterprise AI applications (such as code generation, customer service automation, and data analysis) are just beginning to penetrate, with commercial value gradually materializing. GitHub Copilot already has over one million paid users, and data shows developers experience approximately 30-55% productivity gains after adoption. Enterprises are willing to pay for quantifiable efficiency improvements—a clear value anchor that consumer AI applications lack.

A Rational Perspective: The Symbiotic Relationship Between Bubbles and Technological Revolutions

Looking to Historical Technology Cycles for Reference

It's worth reflecting on the fact that bubbles and technological revolutions are often two sides of the same coin. The railroad era, the electrical era, and the internet era all went through cycles of "overinvestment—bubble burst—infrastructure sedimentation—long-term prosperity." The railway tracks and fiber optic cables laid during bubble periods ultimately became the foundation for subsequent economic growth.

Venezuelan-British scholar Carlota Perez systematically demonstrated this pattern in her classic work Technological Revolutions and Financial Capital. She divides each technological revolution into four phases: Irruption, Frenzy, Turning Point (usually accompanied by a bubble burst), and Deployment. During the Frenzy phase, financial capital floods into the new technology sector, pushing asset prices far beyond fundamentals; after the bubble bursts, the excess infrastructure is "inherited cheaply," laying the material foundation for subsequent large-scale societal adoption. The railway miles built during Britain's 1840s railway bubble exceeded actual transportation demand by several times. After the bubble burst, countless investors lost everything, but those tracks supported Britain's industrial takeoff for the next half-century. The submarine cables and backbone networks laid during the 2000 internet bubble made subsequent bandwidth-intensive applications like YouTube and Netflix possible—yet most of the companies that laid those cables had already gone bankrupt.

If AI follows a similar cycle, then the massive investments being made today—even if they cannot generate commensurate returns in the short term—will form the computing infrastructure, software toolchains, and engineering talent pools that serve as the foundation for the next phase of widespread AI deployment. The question isn't whether these investments are "wasted," but who will bear the losses when the bubble bursts and who will reap the dividends during the deployment phase.

Even if AI asset prices experience a correction, the computing infrastructure, talent accumulation, and engineering practices it has catalyzed will continue to benefit the entire industry for years to come. In other words, the bursting of a bubble doesn't signify the failure of a technology—it's more likely to be an industry shakeout that separates the real from the fake.

Implications for AI Practitioners and Investors

For tech practitioners and entrepreneurs in the thick of it, the value of this discussion lies in maintaining clarity:

1. Return to fundamental value: Don't be misled by the valuation premium of the "AI" label. What's truly sustainable are products that solve real problems and generate real revenue. In the venture capital world, "AI wrappers" (applications built as thin layers on top of large model APIs) are experiencing valuation corrections because they lack technical moats and can easily be integrated by platform providers or replicated by competitors at low cost. In contrast, AI products that deeply integrate vertical industry data and build domain-specific capabilities demonstrate stronger pricing power and customer stickiness.

2. Focus on unit economics: The inference costs and gross margins of AI services are more reflective of long-term viability than user growth numbers. The operational cost structure of AI applications is fundamentally different from traditional SaaS: every user interaction requires GPU compute, generating real inference costs. For large language models, the cost of each API call is determined jointly by the number of input tokens, output tokens, and the parameter size of the model used. A typical GPT-4-level conversational interaction might cost anywhere from a few cents to tens of cents, meaning that if a user interacts frequently each day, monthly compute costs could match or even exceed their subscription fee. This is why "token economics"—the ratio of revenue to cost per token—is becoming one of the most critical operational metrics for AI companies. Companies that can optimize inference costs through model distillation (compressing large model capabilities into smaller, cheaper models), intelligent routing (directing simple requests to lightweight models), and caching strategies will build significant advantages in gross margins.

3. Prepare for cycles: Whether or not the AI bubble exists, industry hype will eventually return to rationality. Building moats and cash flow resilience ahead of time is crucial. Specifically, this means avoiding over-reliance on venture capital lifelines, establishing predictable revenue streams as early as possible, and reserving sufficient cash runway for scenarios where the funding environment deteriorates. The predicament of numerous growth-stage tech companies during the 2022 rate-hike cycle—valuations halved, forced layoffs, even shutdowns—should serve as a cautionary tale for all AI entrepreneurs.

Conclusion

"The AI bubble is bursting—we just haven't realized it yet"—whether this statement becomes prophetic remains to be determined. The short piece that sparked this discussion didn't provide exhaustive data-driven evidence; it was more about raising a thesis worthy of vigilance.

But regardless of the ultimate answer, maintaining prudence during prosperity and questioning returns are essential qualities for navigating technology cycles. The long-term value of artificial intelligence is certain. As investors and practitioners, what we need to do is find a judgment path grounded in real value—somewhere between euphoria and panic. As Warren Buffett's classic warning goes: "Only when the tide goes out do you discover who's been swimming naked." For the AI industry, whether the tide is receding remains uncertain, but ensuring you're not the one swimming naked is always the wise move.