Legacy of Technology Bubbles

The history of the digital economy demonstrates a persistent pattern: speculative bubbles, while destroying significant volumes of capital, simultaneously form the material, technological, and human infrastructure that becomes the foundation of the next technological cycle. The dot-com crash of 2000–2002, the transformation of the internet during the Web 2.0 era, and the bubble now forming in the artificial intelligence market can be viewed as links in a single process, in which the overvaluation of expectations systematically outpaces the emergence of demand, yet the investments made during the boom period do not vanish without a trace but are passed on to subsequent market participants at a substantially reduced cost.

The dot-com crash is customarily assessed above all through the scale of financial losses — by widely cited estimates, on the order of five trillion dollars in market capitalization. However, from an infrastructural standpoint, the bubble performed the function of large-scale advance financing for the internet of the future.

Telecommunications infrastructure. During the period 1998–2001, telecommunications companies laid excessive volumes of fiber optic cable in anticipation of explosive traffic growth. By 2001, only about 2.6% of the laid capacity was in use; a significant portion of the fiber remained “dark.” The bankruptcies of Global Crossing, WorldCom, and Exodus led to backbone networks, data centers, and thousands of miles of cable being sold off at 10–20% of the cost of laying them. The result was a sustained decline in the cost of bandwidth, without which video hosting, IP telephony, content delivery networks, and mass streaming would have been economically impossible.

Computing capacity and real estate. Server and networking equipment was sold after the crash for fractions of a percent of its original cost; specialized buildings intended for hosting and data centers were subsequently repurposed for colocation, CDNs, and cloud services. The example of Amazon is illustrative: excess capacity that sat idle outside of peak loads became the foundation for the launch of AWS in 2006. Thus, the concept of cloud computing is a direct consequence of the overinvestment of the dot-com era.

Broadband access. The bet on mass broadband (DSL, cable modems) ensured the replacement of dial-up access with a permanent high-speed connection, which created the audience base for all subsequent development of the internet.

Surviving companies and the refinement of business models. Alongside bankrupted ventures (Pets.com, Boo.com), companies with real economics survived the crash: Amazon, eBay, Priceline, Akamai, PayPal. The launch of Google AdWords (2000) and AdSense (2003) demonstrated the viability of an advertising model oriented toward measurable actions rather than reach.

Human capital. The most underappreciated residue of the bubble turned out to be people. Thousands of engineers, managers, and entrepreneurs freed up after 2001 formed the personnel base of the next cycle. The so-called “PayPal Mafia” is illustrative: former PayPal employees subsequently founded or financed LinkedIn, YouTube, Palantir, SpaceX, Tesla, Affirm, and a number of other companies.

The technology stack and culture. The scarcity of capital stimulated a shift toward open-source software (the LAMP stack), as well as the development of RSS, REST APIs, and AJAX. The venture industry changed its criteria for evaluating projects: reach metrics gave way to a requirement of unit economics. Finally, a stable consumer habit of online shopping and electronic payments took hold, one that outlived its own creators.

Taken together, the dot-com crash represented not so much the destruction of value as its redistribution: the costs were borne by first-wave investors, while the benefits accrued to the companies and users of subsequent years.

If the dot-com bubble left behind predominantly physical infrastructure, then the Web 2.0 period (approximately 2004–2012, a term coined by Tim O’Reilly in 2004) formed the social, economic, and behavioral environment in which the internet turned into a mass medium of everyday life.

The technological level. The website evolved from a static page into an application (Gmail, 2004; Google Maps, 2005). Open APIs made possible the model of “the internet as a construction kit,” in which payments, maps, computation, and communications are provided as interchangeable service blocks. Services such as Heroku and GitHub lowered the barrier to entry into development to a minimum. At the same time, interface patterns became entrenched — the feed, the “like,” the repost, infinite scroll (invented by Aza Raskin in 2006) — patterns that have not undergone fundamental change over the course of a decade and a half.

The economic level. The central innovation became the platform in place of the product: the value of a service is determined by the network effect rather than by the good being directly sold. The freemium model and the principle of “free in exchange for attention and data” radically changed consumer expectations. The YouTube Partner Program (2007) laid the foundation for the creator economy. Professional practices of growth management took shape: A/B testing, funnels, retention metrics. The platform model spread to labor relations (the gig economy) and to the direct monetization of an audience (the creator economy).

The cultural level. The user was transformed from an anonymous consumer into a co-author of content and into a profile with a trackable reputation. The social graph — a machine-readable representation of the connections between people — became the foundation of recommendation systems, targeted advertising, and viral distribution. The principle of the “wisdom of the crowd” (Wikipedia, Stack Overflow, Reddit) decentralized the production of knowledge. The advent of the smartphone met an audience already trained and familiar with accounts, feeds, and cloud storage.

Side effects. The legacy of Web 2.0 also includes a problematic component: centralization in the hands of a small number of platforms despite originally decentralist rhetoric; an attention economy built on dopamine-driven retention mechanisms; a data-extraction model in which the user of a free product becomes the product themselves. The advertising machine of “behavior — profile — prediction — personalization” gave rise to phenomena discussed today as problems of digital well-being, polarization, and disinformation.

Comparing the epochs, one can say that Web 1.0 built the internet as an information network, the dot-com bubble built its physical infrastructure, Web 2.0 built its social platform, the mobile era transferred the internet onto a personal device, and cloud computing shaped it into a universal computing environment. The question of what the contribution of the artificial intelligence era will be remains open.

The analogy with the dot-coms suggests a distinction between two oppositely directed disillusionments. In the case of the dot-coms, disillusionment affected primarily the investor; in the case of artificial intelligence, the end user for the first time acts as the direct customer of the technology, which changes the structure of the cycle.

Disillusionment among the mass user is most likely in the segment of general-purpose chatbots, which make up the overwhelming majority of consumer contact with AI. The gap between expectation (a reliable personal assistant that retains context and carries out actions) and reality (confidently stated factual errors, the absence of long-term memory and learning from user data, an inability to autonomously carry out tasks in external systems) produces a fatigue effect: the cost of verifying results begins to exceed the time saved. At the same time, chatbots themselves as a tool will most likely endure — in the same way that search engines survived the dot-com crash without having lived up to maximalist expectations.

Disillusionment among business and investors is more likely to touch specialized models and agentic systems, which account for the bulk of capital investment: autonomous transport, systems that lay claim to replacing qualified specialists (in medicine, law, and development), and accelerated drug discovery. The key constraint here is the “last-percent problem”: a reliability of 90%, sufficient for a conversational interface, proves unacceptable in environments that require legal accountability and a fault-tolerance level of 99.99%. The cost of achieving the last percent of reliability exceeds the cost of the first ninety and is tied not to computing power but to the transformation of processes, regulation, and the distribution of responsibility. Empirical data confirm the gap between adoption and transformation: about 60% of workers have already gained access to AI tools, yet only about a third of companies report a deep restructuring of their business, and the share of organizations that fully trust third-party models has fallen from 24% (2023) to 16% (2026). The most probable trigger for a correction is the question of the return on investment in computing infrastructure — given that AI-related debt in the United States is estimated to have approached 500 billion dollars, while projected investment in data centers is estimated at 3 trillion dollars by 2028.

The probable legacy of the correction structurally reproduces the experience of the dot-coms. First, excess computing capacity — graphics processors, gigawatt-class data centers, energy infrastructure — will become cheaper after the correction of demand, in the manner of “dark fiber,” and the cost of inference will approach the cost of a commodity resource. Second, open models will play a role analogous to that of Linux and Apache after 2002: the market will shift from “the most powerful model” to “the cheapest model sufficient for the task.” Third, an infrastructural stack invisible to the user will remain in place: vector databases, data-preparation pipelines, quality-evaluation methods, RAG architectures. Fourth, the change in interface grammar will persist — dialogue and the formulation of intent as a means of interacting with a computing system. Fifth, several economically justified applications will hold up: programming assistants, enterprise search, customer support, tools for generating visual content; directions that require legal accountability from AI will likely retreat. Sixth, there will remain a generation of specialists socialized within the transformer paradigm, as well as a problematic legacy — masses of synthetic content that raise the question of authenticity verification and that are giving rise to a market for the means of checking it.

The comparison of the three epochs allows us to formulate a general mechanism: a speculative bubble performs the function of premature collective financing of infrastructure for which demand has not yet formed. The dot-com crash subsidized cheap bandwidth and computing power; the collapse of numerous social platforms consolidated a free social graph; a possible correction of the AI boom will, in all likelihood, subsidize nearly free access to computable intelligence. It is significant that the principal beneficiary of such a cycle has historically turned out to be not the company that sold the technology itself, but the one that embedded it as an invisible layer within an ordinary product — in the same way that Amazon became not an “internet company” but a store that makes use of the internet. Such a perspective suggests that the ultimate outcome of the current boom may not be the disappearance of artificial intelligence, but its transition from the status of an independent industry to the status of infrastructure — comparable in its degree of invisibility to electricity or broadband access.

The history of computing technology knows no shortage of cases in which devices created for entertainment or civilian purposes tur