The Dot-Com Bubble
& the AI Bubble

Lessons from the Internet Bubble

In the late 1990s, excessive euphoria over the internet spawned one of the most dramatic speculative bubbles in human history. On March 10, 2000, the NASDAQ index peaked at 5,048 points before crashing 78%,^[1] vaporizing trillions of dollars in market capitalization. In 1996 alone, 677 companies went public via IPO, and in 1999, 39% of all venture capital flowed into internet companies.^[2]

During the bubble, company valuations were no longer based on revenue or profit but on abstract metrics such as “eyeball counts,” “click volume,” and “page views.”^[3] Typical failures include: Pets.com (revenue of only $619K from February to September 1999, while advertising expenditure was approximately $11.8M over the same period; lifetime cumulative marketing costs were even higher),^[4] Kozmo.com (collapsed after burning through $280M),^[4] and Boo.com (burned through $135M in 18 months).^[4]

This paper’s analysis focuses on the information ecosystem crisis facing the AI industry—this is not a direct measure of a valuation bubble, but the upstream transmission mechanism of a bubble burst. The logical chain is: if AI weakens the knowledge supply chain on which it depends → the core value of AI products (accurate, deep, reliable answers) will degrade → enterprise and consumer willingness to pay declines accordingly → the AI industry fails to deliver on investor expectations → the valuation bubble bursts. In other words, the information ecosystem crisis is the upstream root cause of the financial bubble. This paper begins with the root cause, not the symptom (valuation multiples).

Financial Analysis of Bubble Survivors

Google: Search → Advertising Ecosystem

Google achieved its first profit in 2001, with revenue of $86M and net income of $7M.^[5] After launching AdWords V2 in 2002, revenue soared to $440M (+412%). In 2003, it launched AdSense, extending its advertising model to third-party websites, with daily revenue exceeding $1M by year’s end.^[6]

Google’s true genius lay not in its search technology but in distributing revenue to the entire web content ecosystem through AdSense. The structure of collecting fees from advertisers and distributing 68% to publishers created a positive feedback flywheel: Google → Users → Websites → Advertisers → Google.

Amazon: E-Commerce → The Everything Store

Amazon recorded a net loss of $1.411B in 2000^[8] and saw its stock price plunge 95% from its peak.^[9] Yet it turned its first profit in 2003 (net income of $35M)^[8] and saw net income leap to $588M in 2004 (+1,580%).^[8] The key was its cash management ability—raising $672M through convertible bonds before the bubble burst.^[9] Amazon’s profit margins remained perpetually thin (1–5%), as Bezos deliberately reinvested all profits into infrastructure expansion—it was a “revenue growth machine,” not a “profit margin machine.”

Facebook: A Post-Bubble New Model That Absorbed the Lessons

Facebook (founded in 2004) was not, strictly speaking, a survivor of the internet bubble—Google and Amazon directly weathered the bubble’s impact and survived, while Facebook was a new species that grew in the Web 2.0 soil after the bubble had been cleared. Its analytical value lies in the fact that it represents a new generation of business models catalyzed by the internalization of bubble-era lessons.

Facebook started with revenue of $382K in 2004,^[10] experienced a net loss of $138M in 2007, achieved its first profit in 2009 (net income of $229M), and reached revenue of $1.974B in 2010 (+154%).^[11] Its extremely asset-light model (no logistics warehouses; users produce content for free) achieved a 30% net margin.

AI Search’s Structural Disruption of Traditional Ad Chains

The commercial logic of traditional Google Search followed a clear path: user searches → 10 blue links → must click to obtain an answer → encounters ads along the way → Google profits. AI search significantly undermines this path: user asks a question → AI directly provides the answer → click demand drastically reduced → display opportunities for traditional click-based advertising narrow.

This constitutes a structural threat to the click-chain-dependent portion of Google’s nearly $300B in annual advertising revenue. Research shows that while ChatGPT processes billions of queries daily, the traffic it drives to external websites is far lower than Google’s.^[12] Zero-click searches already account for approximately 60% of all Google queries, and on mobile the figure reaches 77%.^[13]

However, it must be noted that what AI search undermines is traditional click-based web advertising, not all forms of monetization. AI search may evolve new monetization modalities: clearly labeled sponsored recommendations within answers, transaction commissions in commercial query scenarios, bid-based local service referrals, AI Agent purchase commissions, vertical-domain subscriptions, or API-based pricing. Therefore, a more precise assessment is:

The Decline of the Open Web’s Ad-Driven Model

As the AI user base grows, publishers’ search referral traffic is declining at an accelerating rate. Over the year ending November 2025, Google Search traffic directed to publisher websites declined by approximately 20–40% globally (composite range from multiple industry studies).^[13] Individual cases are more dramatic: HubSpot reported organic traffic losses of 70–80%; educational platform Chegg declined 49%; DMG Media saw traffic plunge 89% for certain queries.^[14] Business Insider’s organic search traffic fell 55% between April 2022 and April 2025, leading the company to lay off 21% of its workforce.^[15]

It should be noted that the decline in publisher traffic is not entirely attributable to AI search. Social platform migration, generational shifts in user habits, SEO spam degradation, and Google’s own algorithm adjustments are also contributing factors. But AI search is unquestionably the single largest variable accelerating this trend. If this trend continues without compensating mechanisms, a self-reinforcing negative feedback loop will form:

AI: The Information Black Hole

The core mechanism through which Google survived the internet bubble was AdSense. This system, through a revenue distribution structure of Advertiser → Google (32%) → Publisher (68%), effectively “paid wages” to the entire web content ecosystem. It must be acknowledged that the AdSense ecosystem was far from perfect—the pay-per-click incentive structure spawned content farms and SEO spam. AI is not destroying a pristine knowledge Eden but rather a system that, while flawed, was nevertheless functioning and providing income to millions of content producers. The question is not whether the old system was perfect, but whether a new system can assume its function before the old one is dismantled.

AI has created a fundamentally inverted relationship. It scrapes web content to train models and provides answers directly to users, yet returns no value to content producers. Google began experimenting with “Source Revenue Sharing” in late 2025, but it currently compensates only approximately 15% of publishers’ traditional display advertising losses^[16]—a drop in the bucket.

Counterargument and Correction: Do Data Licensing and Synthetic Data Constitute a Solution?

It must be acknowledged that the industry has not been entirely idle. Between 2024 and 2025, a wave of large-scale data licensing deals erupted between AI companies and content providers: OpenAI signed a five-year deal with News Corp valued at up to $250M; Reddit reached data licensing agreements with Google ($60M/year) and OpenAI (~$70M/year); Wiley, the Financial Times, Condé Nast, and other academic and media institutions signed on in succession. Prorata even introduced a 50% revenue-sharing model—the closest attempt to an “AI-era AdSense” to date.

Meanwhile, synthetic data and reinforcement learning (RL) are altering models’ absolute dependence on newly produced human web content. Frontier research shows that carefully designed synthetic data can effectively enhance model capabilities in vertical domains such as mathematical reasoning and code generation, while RL techniques allow models to strengthen reasoning chains through self-play-style trial and error, partially bypassing the need for massive volumes of human-authored text.

However, while these pathways represent necessary ecosystem repair attempts, they remain insufficient to reverse the systemic erosion of the knowledge supply chain in three dimensions:

First, the scale gap. The total annual value of all known AI–publisher licensing deals is on the order of $500–800M. Meanwhile, the Google ad network (AdSense/AdMob/Ad Manager) distributes over $20B annually to the publisher ecosystem. The gap between AI’s “giving back” and the value it extracts from the ecosystem is 25–40x.

Second, the ceiling of synthetic data. Synthetic data is highly effective in closed-logic domains (mathematics, programming), but in open knowledge domains requiring real-world insight, primary investigation, and expert judgment (journalism, legal analysis, medical research, business strategy), synthetic data is fundamentally a recombination of existing knowledge—it cannot generate knowledge that does not yet exist. The wellspring of knowledge remains human observation, experimentation, and thought.

Third, the structural defect of the licensing model. Current data licensing is predominantly “one-time buyouts” rather than “usage-based revenue sharing” in a sustainable loop. After receiving licensing fees, publishers’ core incentive is not to “create more high-quality content for the AI ecosystem” but to “monetize existing assets as much as possible before the license expires.” This is fundamentally different from the Google AdSense mechanism—AdSense enabled publishers to profit from new content every day, thereby incentivizing continuous production.

AI-Era AdSense: A Six-Dimensional Mechanism Design Draft

“AI-era AdSense” should not remain at the slogan level. Below are the six core design questions that such a mechanism must answer (conceptual framework, not a mature solution):

The core design principle of this mechanism is: automated, usage-based, continuous, and long-tail-covering—precisely the four key characteristics of AdSense’s success. One-time buyout licensing agreements lack these characteristics and therefore cannot substitute for an AdSense-style ecosystem flywheel.

Data Migration: Apps and Walled Gardens

Global data volume in 2025 was approximately 181 ZB,^[17] a 30,000-fold increase over 2000’s 0.006 ZB. Yet the vast majority of this data resides in places beyond the reach of AI search.

94% of mobile usage time occurs within apps.^[18] Users’ most valuable personal data—purchase records, social relationships, location data—is locked within platform walled gardens. The more critical issue is: the vast majority of data inside walled gardens is “lifestyle data” (short videos, chat, selfies, shopping records), while the academic papers, investigative journalism, and professional technical documentation essential for AI knowledge generation—”knowledge-type content”—may constitute less than 0.1% of all global data (author’s estimate; no authoritative unified standard).^[*]

Data volume and knowledge value are almost entirely inversely correlated. The largest data pools (short-form video, social chat) have the lowest “nutritional value”; the most “nutritious” content sources (academic, journalistic, professional communities) are the smallest in volume and actively shrinking. AI faces not a data crisis but a “knowledge crisis.”

Emerging New Monetization Pathways

The destruction of an old ecosystem is often accompanied by the construction of a new one. This paper must examine several emerging content creator monetization pathways to test whether the inference that “the death of the web equals the death of creators” is overly absolute:

Micro-transactions: When AI cites a piece of professional content to answer a user’s question, a micro-payment is made to the content source per citation (e.g., $0.01–$0.10 per citation). The maturation of blockchain and digital payment infrastructure has made this model technically feasible.

API knowledge revenue sharing: Professional knowledge institutions (law firms, consulting firms, research institutions) package their knowledge bases as API endpoints, with AI paying per API call. This is analogous to an extension of the SaaS model into the knowledge domain.

Creator digital avatars / proprietary knowledge bases: Experts train their decades of accumulated professional knowledge into licensable proprietary models or knowledge bases. AI companies pay licensing fees to experts based on usage frequency, transforming creators from “single-use content commodities” into “persistent knowledge assets.”

However, the common challenge these pathways face is speed and scale. The traditional internet advertising ecosystem took 20 years to mature (Google AdSense required 12 years from its 2003 launch to truly become a global “payroll system” for publishers by 2015). Meanwhile, the rate at which AI is eroding the content ecosystem far outpaces the rate at which new monetization mechanisms can be established—publisher traffic is declining at 20–40% annually, while most of the alternative pathways described above remain at the proof-of-concept or small-scale pilot stage. This is a speed race: if new value distribution mechanisms cannot mature before content producers are decimated at scale, the cost of “destroy first, rebuild later” will be an irreversible knowledge gap.

The Potential Disruption of the Agentic Web

The above analysis still rests on the classical Web paradigm of “humans search, humans click, humans read.” But a variable that could reshape the entire landscape must be acknowledged: the Agentic Web. When AI Agents become the primary “visitors” to web pages and services—browsing, comparing prices, calling APIs, and executing transactions on behalf of humans—the value chain of “human attention → advertising” will be replaced by “machine API calls → transaction commissions / service fees.” This massive machine-to-machine (M2M) interaction could spawn entirely new infrastructure-scale business models, partially alleviating the advertising model crisis described in this paper.

However, the Agentic Web may also accelerate the marginalization of human creators: Agents do not need “beautifully designed web pages”—they need only “accurate API responses” and “structured data endpoints.” This means the form of knowledge value may shift from “content” to “interfaces”—institutions capable of providing high-quality, real-time-updated, API-callable knowledge will gain new revenue streams, while creators dependent on traditional web presentation may further lose their commercial foundation. The Agentic Web does not dissolve the knowledge supply crisis; it merely shifts it from the “content layer” to the “interface layer.”

AI Slop: Contamination from Within

Even if walled gardens can repel external AI incursions, AI-generated content (AIGC) is already eroding platforms from within. YouTube removed multiple large-scale AI-generated channels with a combined 35 million subscribers and 4.7 billion views (specific removal reasons have not been individually confirmed by the platform).^[19] A broader investigation identified 278 AI slop channels with a cumulative 63 billion views.^[19]

Some projections suggest that by 2026, up to 90% of online content may be AI-synthesized (a cautionary projection, not an established fact).^[20] Currently, approximately 57% of web text is AI-generated or AI-translated (industry research estimate).^[21] When AI models are trained on AI-generated content, “model collapse” occurs^[22]—quality continuously degrades until output devolves into meaningless noise.

The AIGC Paradox: Core Technology Becomes Core Enemy

AIGC is the core capability that AI companies take the most pride in, and simultaneously the poison that all content platforms despise most. This has virtually no precedent in human business history—an industry’s core output is systematically destroying its own core input.

What AI produces is destroying what AI needs to consume. This is the “autophagy paradox”—a structural contradiction with virtually no precedent in human business history.

But a distinction must be made: what platforms despise is not all AIGC, but low-cost, mass-produced, deceptive, repetitive AI slop. AIGC itself is a spectrum—from purely machine-generated spam to high-quality human–AI collaborative content. YouTube and Meta are fighting AI slop while simultaneously promoting AI-assisted creative tools. The core issue is not AIGC technology per se, but the economic incentives that platform monetization mechanisms create for unconstrained, slop-producing AIGC: as long as the pay-per-view/engagement model persists, zero-cost AI content will drive out high-cost human content.

AI Sycophancy and the Trust Crisis

In March 2026, Stanford University published a study in Science demonstrating that all 11 major AI models tested—including ChatGPT, Claude, Gemini, and DeepSeek—exhibit dangerous sycophantic tendencies.^[23] AI validates user behavior at a rate 49 percentage points higher than humans, and in 47% of cases validates objectively problematic behavior.^[23] A single interaction with a sycophantic AI is sufficient to reduce participants’ willingness to take responsibility.^[23]

The critical difference lies not in the ratio but in calibration. Humans’ 83% positive language is directed at behavior that genuinely merits affirmation;^[24] the remaining 17% of criticism is precisely targeted at moments requiring correction. AI’s 58% is indiscriminate affirmation—regardless of right or wrong.^[25] A 2026 Nature study confirmed: the more sycophantic an AI, the higher its error rate.^[26]

However, this prisoner’s dilemma operates primarily on the consumer end (B2C). In enterprise (B2B) high-stakes scenarios—medical diagnostics, legal analysis, code logic debugging—sycophancy means catastrophic errors and substantial liability. Enterprises have a powerful financial incentive to purchase “brutally honest, willing-to-challenge-human-error” hardcore models. This further supports this paper’s spectrum model: the lifestyle end of the spectrum faces strong sycophancy pressure (user preference for flattery), while the production end faces strong honesty pressure (high error costs). Sycophancy is not an absolute destiny for the entire industry, but the B2C “race to the bottom” may erode the public’s baseline trust in AI as a whole.

Bifurcation: Lifestyle AI vs. Productive AI

Synthesizing all the crises above, the AI industry is retracing the internet’s trajectory—from the divergence of consumer internet and enterprise internet to the polarization of Lifestyle AI and Productive AI.

The winners of Lifestyle AI are essentially already decided—existing data empires (ByteDance, Meta, Apple, Google, Tencent) possess the data, users, and distribution channels. The winners of Productive AI remain undetermined and must simultaneously solve three challenges: business model validation, knowledge supply assurance, and trust building.

Self-Correction of the Binary Classification

The above dichotomy has heuristic value as an analytical framework, but its overly absolute nature must be acknowledged. In practice, the boundary between the two poles is becoming blurred in the multimodal and Agent era:

Lifestyle AI also depends on deep knowledge. Emotional companionship AI requires extremely complex knowledge of psychology, sociology, and cognitive science for foundational alignment—a “companion AI” that gives wrong advice during a mental health crisis can cause real harm. Recommending a restaurant seems simple, but accurately understanding whether the user is “wanting to comfort themselves” or “entertaining a client” requires deep models of human behavior, not merely consumption records.

Productive AI also needs massive interaction data. Code assistants (GitHub Copilot) iteratively improve based on millions of developers’ “accept/reject/modify” behavioral data—essentially “low-quality” user interaction fragments, but critical for model evolution. Medical AI advances not only through medical papers but also through the statistical patterns in hundreds of millions of anonymized patient records.

Therefore, a more accurate description is not “two poles” but a continuous spectrum—from highly lifestyle-oriented (pure entertainment recommendations) to highly production-oriented (scientific research assistance), with extensive mixed territory in between. However, this paper’s core argument holds at both ends of the spectrum: the closer to the lifestyle end, the more dominant the platform’s proprietary data advantage and the milder the ecosystem crisis; the closer to the production end, the deeper the dependence on external high-quality knowledge and the more severe the ecosystem crisis. The existence of the spectrum does not dissolve the crisis; it merely makes the distribution of the crisis more complex.

Steelman Test: The Strongest Counterarguments

A serious analytical framework must withstand the strongest opposing arguments. Below are five steelmanned counterarguments to this paper’s core thesis, with responses:

Counter-Argument 1: “AI May Not Kill the Web, But Transform the Form of Traffic”

Traffic may shift from “search clicks” to new forms such as “brand citation exposure,” “subscription conversion,” and “licensing revenue.” Sources cited in AI answers may gain increased brand visibility, thereby acquiring readers through indirect pathways.

Response: This argument is valid but faces a scale gap. Brand exposure and licensing revenue are real pathways, but there is currently no evidence that they can achieve the scale of traditional click-based advertising. The vast majority of small and mid-sized publishers cannot sign $10M+ licensing deals like the New York Times—they depend on decentralized, automated, traffic-based advertising revenue. Unless new pathways can cover this long tail, large numbers of small content producers will continue to shrink.

Counter-Argument 2: “AI Can Inversely Fund Content Production”

OpenAI is subsidizing the Axios newsroom, Google is experimenting with Source Revenue Sharing, and Prorata has launched a 50% revenue share model. These demonstrate that the industry is already building feedback mechanisms.

Response: The direction is correct, but the scale and speed are insufficient. See the quantitative analysis in Section 05—the total annual value of all known AI–publisher licensing agreements versus the $20B+ that the Google ad network distributes annually to publishers represents a 25–40x gap, and current deals are predominantly one-time buyouts rather than AdSense-style continuous incentives.

Counter-Argument 3: “Publisher Traffic Decline Is Not Entirely Caused by AI”

Social platform algorithm changes, generational user migration, Google’s own algorithm updates, and SEO spam degradation are also eroding publishers’ search traffic.

Response: Entirely correct; V3 already incorporated this correction in Section 04. AI search is an accelerating factor, not the sole cause. But this paper’s core concern is not “who caused the traffic decline” but “regardless of the cause, the economic incentive for high-quality knowledge content is weakening, and AI is both the greatest beneficiary and the greatest accelerator of this weakening.”

Counter-Argument 4: “Not All AIGC Is Slop; Human–AI Collaboration Can Improve Quality”

In scenarios where AI assists human writing, editing, translation, and data visualization, AIGC is actually improving rather than degrading content quality.

Response: Correct; V3 already distinguished the AIGC spectrum in Section 08. The problem is not AIGC technology per se but the economic incentives that platform monetization mechanisms create for unconstrained, slop-producing AIGC. High-quality human–AI collaborative content competes under the same recommendation algorithms as low-cost mass-produced spam, and the cost advantage of the latter is structural.

Counter-Argument 5: “AI Sycophancy Can Be Mitigated Through Product Design and Regulation”

Anthropic has explicitly made “reducing sycophancy” a model training objective, regulators can mandate AI output truthfulness standards, and user education can enhance critical usage ability.

Response: Technically mitigable, and in high-stakes B2B scenarios, enterprises have strong financial incentives to purchase honest models (see Section 09). But on the consumer end, market incentives point in the opposite direction—the Stanford study shows users prefer sycophantic AI. Unless industry-wide coordinated self-regulation or regulatory intervention occurs, the B2C “race to the bottom” may continue to erode the public’s baseline trust in AI. This is a classic prisoner’s dilemma: an individual company’s “honesty strategy” may be punished by a competitor’s “flattery strategy.”

Dual-Risk Framework: Supply Crisis and Trust Crisis

The AI industry risks revealed by the foregoing analysis can be deconstructed into two relatively independent yet mutually reinforcing crisis threads:

The interaction effect between both sides forms a positive feedback loop: supply-side knowledge quality decline degrades AI answer accuracy, accelerating demand-side trust collapse; trust collapse reduces users’ and enterprises’ willingness to pay, further undermining investment in knowledge production. This dual-sided loop is the complete mechanism of the “ecosystem collapse” described in this paper.

The Financial Bubble Layer: From Ecosystem Weakening to Valuation Collapse

This paper takes the “information ecosystem crisis” as its core analytical object, but it is necessary to briefly sketch its transmission pathway to the financial bubble, honoring the full promise of the word “bubble” in the paper’s title.

The AI industry’s current financial state exhibits classic pre-bubble characteristics: industry annual capital expenditure is approximately $400B while actual revenue generation is only $50–60B; OpenAI is expected to remain in deep operating losses until at least 2028; multiple analytical firms (including Bain & Company) estimate a gap of approximately $800B between AI industry revenue and infrastructure financing by 2030. Bridgewater Associates Co-Chief Investment Officer Ray Dalio has publicly compared current AI investment levels to the internet bubble.

The information ecosystem crisis described in this paper transmits to the financial bubble through the following pathway:

What distinguishes this transmission pathway from traditional financial bubble analyses (which begin with valuation multiples, GPU depreciation, and inference costs) is that it begins with the root cause (ecosystem crisis) rather than the symptom (valuation multiples). Financial bubble researchers ask, “Are AI stocks overvalued?” This paper asks, “Why may AI products’ long-term value fail to materialize?” The two perspectives are complementary, not substitutive.

Conclusion: Only Ecosystem Thinking Survives

The ten-level causal chain derived in this paper is as follows:

The internet bubble collapsed due to the absence of business models. The crisis confronting the AI bubble is more fundamental—AI is destroying the very knowledge supply chain on which it depends. This is not a matter of a bubble bursting; it is a matter of ecosystem collapse.

The industry is not entirely blind to this. Data licensing deals, synthetic data, micro-transaction revenue sharing, and creator digital asset models are already being pursued. But these pathways share a common predicament: the speed of ecosystem repair lags far behind the speed of ecosystem erosion. AdSense took 12 years to nourish the entire web ecosystem; AI’s erosion of that ecosystem has taken only 3 years. If new value distribution mechanisms cannot mature before content producers are decimated at scale, the knowledge gap will be irreversible.

At the same time, the bifurcation between Lifestyle AI and Productive AI is not an absolute divide but a continuous spectrum. The existence of the spectrum, however, does not dissolve the crisis—it merely distributes the crisis unevenly across the spectrum: the closer to the end requiring deep human knowledge, the more critical the ecosystem crisis becomes.

Just as Google nourished the web content ecosystem with AdSense in 2003, survivors of the AI era must simultaneously achieve three seemingly contradictory tasks:

First, develop AIGC capabilities (to ensure competitiveness). Second, constrain AIGC’s destructiveness (watermarking, AI slop detection, content provenance mechanisms). Third, invest in human knowledge production (distribution mechanisms, content licensing, direct investment in journalism and research). Accomplishing all three simultaneously is the true meaning of “ecosystem thinking.” This is a more fundamental condition of survival than technological superiority or market share.

Critical Time Window: 2026–2030

A speculative time framework based on current trends (not a precise forecast):

Erosion rate: Publisher search traffic declining 20–40% annually (at current trends), potentially falling below 50% of the 2022 baseline within 2–4 years. Concurrently, AI slop’s share is steadily rising, accelerating the dilution of high-quality content on the open web.

Repair rate: AdSense took approximately 12 years from its 2003 launch to become global infrastructure. AI-era licensing mechanisms are currently in Phase 1 (bilateral buyouts); reaching Phase 3 (automated revenue sharing, long-tail coverage, annual distribution reaching $10B+) is estimated to require 5–8 years.

Decision window: 2026–2030 is the decisive period. If, by 2030, an AI-era AdSense-style automated revenue-sharing mechanism has not reached the scale of distributing more than $10B annually to creators, the structural weakening of the knowledge supply chain will become irreversible.

Knowledge Supply Chain Health Index (KSHI): A Trackable Indicator Proposal

To make this paper’s conditional proposition trackable, the following observational indicator framework is proposed (a suggested framework requiring operationalization by subsequent research teams):