Blog · Review Essay · Last reviewed June 25, 2026

Technological Revolutions and Financial Capital and the AI Bubble Question

Carlota Perez's Technological Revolutions and Financial Capital is useful precisely because it refuses the simple choice between "real technology" and "mere bubble." It shows how a genuine technological revolution can arrive through speculative excess, institutional lag, inequality, infrastructure buildout, crash, and political choice.

An AI bubble, in this review, means a financial and institutional pattern in which equity valuations, private funding, capital expenditure, cloud commitments, procurement urgency, and cultural expectations run ahead of proven deployment value. That does not prove the underlying technology is fake. It means the social accounting has to separate useful infrastructure from valuation theater, durable capability from demo pressure, and public benefit from private rent.

The practical test is not whether AI stocks rise or fall. It is whether the installation phase leaves behind accountable capacity: verified use cases, measured productivity, auditable safety, public-interest access, energy and labor accounting, and exit rights when the promised deployment value does not appear.

The Book

Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages was published by Edward Elgar in 2002, with the paperback listing dated 2003. The publisher describes it as a long-view interpretation of economic good times and bad times, connecting technological change with finance. RePEc's EconPapers record lists the book under economics, finance, innovation, and technology, gives the 2002 date, and shows substantial later citation activity.

Perez is not writing pop futurology. UCL describes her as an Honorary Professor at the Institute for Innovation and Public Purpose whose work studies technical change, economic development, finance, markets, government, and institutional transformation. That background matters because the book's argument is not "new technology makes society better." It is closer to this: new technological potential becomes socially productive only after capital, infrastructure, firms, states, labor markets, and public expectations are reorganized around it.

The book appeared after the dot-com collapse, which gives it a useful temperament. It is neither anti-internet scolding nor boom-time boosterism. Perez treats bubbles as recurrent features of capitalist technological change. The fever can be wasteful and destructive, but it can also finance the networks, skills, standards, factories, cables, machines, and organizational experiments that later societies inherit.

Current Context

As of June 25, 2026, official company disclosures show an installation phase at industrial scale without settling the bubble question. Microsoft reported in April 2026 that its AI business had surpassed a $37 billion annual revenue run rate. Alphabet's first-quarter 2026 transcript put capital expenditures at $35.7 billion, overwhelmingly for technical infrastructure to support AI opportunities, with roughly 60% of that technical-infrastructure investment in servers and 40% in data centers and networking equipment. Amazon's first-quarter 2026 release said trailing-twelve-month free cash flow fell to $1.2 billion, primarily because property-and-equipment purchases rose by $59.3 billion year over year, an increase it said mainly reflected AI investments. Those disclosures prove neither mania nor durable productivity. They show the scale of capital being converted into compute, buildings, power contracts, cloud capacity, and accounting expectations.

That evidence is installation evidence, not verdict evidence. It says capital is being organized around a forecast. It does not by itself prove economy-wide productivity, fair value, a crash, or a durable golden age. A serious bubble claim has to name the layer it is judging: public-market valuation, private funding, chip and data-center capex, cloud commitments, enterprise procurement, labor-replacement claims, or the cultural story that adoption is inevitable.

The physical and market-structure context is just as important as the revenue story. The International Energy Agency's 2025 Energy and AI report estimated global data-center electricity use at about 415 TWh in 2024 and projected about 945 TWh by 2030 in its Base Case. The FTC's 2025 staff report on cloud-provider and AI-developer partnerships identified areas to watch around access to compute and talent, contractual and technical switching costs, and cloud partners' access to sensitive technical and business information. Perez's categories become concrete here: the AI boom is not only a software cycle. It is a finance, energy, cloud, chip, labor, and dependency cycle.

The grid context now belongs in the same ledger. In June 2026, FERC directed the six regional grid operators under its jurisdiction to justify or reform rules for integrating data centers and other large energy users. NERC's May 2026 large-load reliability guideline treats emerging large loads, including data centers, as planning and operations risks that require better forecasting, coordination, modeling, monitoring, and restoration analysis. If speculative infrastructure creates durable grid obligations before deployment value is settled, the bubble question becomes a public-infrastructure question.

The disclosure environment also matters. The SEC's 2024 AI-washing cases against two investment advisers show the basic regulatory point: claims about AI capability and use can be securities and investor-protection claims, not just branding. For this page, that means "AI boom" evidence has to stay sorted. Company-reported run rate, capital expenditure, cloud commitment, model benchmark, customer pilot, investor deck, and audited cash flow each answer a different question.

Great Surges

Perez's basic unit is not the single invention. It is the technological revolution: a cluster of interdependent technologies, infrastructures, organizational forms, and business practices that can spread across the whole economy. Edward Elgar's listing names the historical sequence she studies: the Industrial Revolution, steam and railways, steel and electricity, oil and automobiles, and the information revolution.

Each surge has a rhythm. A new cluster breaks through. Finance discovers it. Entrepreneurs, engineers, investors, and speculators pile in. Infrastructure is built unevenly. Old institutions lag behind the new technical possibilities. Inequality and dislocation grow. A bubble forms around future expectations. The bubble breaks. Then comes the turning point: the same technical potential can either remain trapped in casino logic or be redirected into broader deployment.

This is why the book belongs beside histories of cybernetics, information empires, platform capitalism, and technological politics. It gives those themes a macroeconomic skeleton. The interface, model, platform, data center, chip supply chain, cloud contract, and workplace dashboard are not only technical artifacts. They are parts of an investment regime and an institutional settlement.

Finance as Accelerator

The strongest part of the book is Perez's distinction between financial capital and production capital. Financial capital is mobile, impatient, and attracted to new spaces where old rules have not yet stabilized returns. Production capital is slower and tied to firms, infrastructure, labor, supply chains, and actual deployment. Early in a surge, finance can be useful because it funds experiments that no cautious incumbent would underwrite. Later, if finance becomes self-referential, the technological revolution is treated mainly as a story for valuation.

That distinction is a clean tool for reading AI. A model demo, a valuation, a GPU order, a data-center lease, a benchmark score, an enterprise pilot, a headcount reduction, and a durable productivity gain are different things. A bubble narrative collapses them into one rising line. Perez gives us a better question: which investments are building deployable productive capacity, and which are only trading claims on the expectation that capacity will appear?

The accounting question is not only how much money is spent. It is whether the asset being built is a reusable productive substrate or a brittle dependency whose value depends on one model family, one vendor, one pricing regime, one power contract, or one forecast of demand. A data center, a chip cluster, or an enterprise agent program can be productive capital in one institutional setting and stranded cost in another.

The answer is not obvious. Waste and usefulness can coexist. The dot-com bubble overbuilt capacity and destroyed wealth, but it also left fiber, server practices, web standards, logistics models, payment habits, and talent networks. The AI boom may produce the same ambiguous inheritance: excess compute contracts, stranded data centers, disappointed startups, and also new chips, inference infrastructure, workflow redesign, labor conflicts, safety institutions, and public habits around machine assistance.

The Bubble Ledger

The practical object Perez helps name is a bubble ledger: a record that separates six claims often fused by hype. Technical capability asks what the system can actually do under specified conditions. Productive deployment asks whether a workflow, organization, or public service measurably improves after adoption. Financial claim asks whether revenue, margins, cash flow, and useful life justify the capital committed. Dependency claim asks whether value rests on replaceable capacity or on lock-in across vendor, data, cloud, chip, power, and integration layers. Institutional residue asks what infrastructure, skills, standards, dependencies, and rules remain after the speculative phase. Social cost asks who absorbed layoffs, energy burden, water demand, surveillance, degraded service, deskilling, or public risk while the experiment ran.

The ledger prevents two lazy readings. The first says "bubble" and stops looking for real capability. The second says "real capability" and stops accounting for waste, lock-in, or harm. Perez's framework is valuable because it keeps both facts in view: installation phases can build the rails of a later economy while also mispricing risk, rewarding imitation, and giving financiers first claim on a social future they did not democratically design.

For AI, the ledger has to track more than model accuracy. It should connect capex to actual use cases, compute to energy and grid constraints, procurement to exit rights, productivity claims to baselines, labor savings to job redesign, model access to cloud concentration, and safety controls to incident records. Otherwise the boom leaves behind infrastructure without memory of the public trade-offs that built it.

A useful deployment proof has four parts. First, it states the task and baseline before AI enters the workflow. Second, it measures the whole deployed system, including retrieval, prompts, tools, human review, costs, errors, security, and user impact. Third, it records who can stop, appeal, roll back, or replace the system. Fourth, it survives renewal review after prices, model versions, vendor terms, and organizational dependence have changed. Without those pieces, a pilot can become infrastructure while still resting on a demo.

Installation Account

The practical artifact this review adds to Perez's cycle is an installation account: a living record of what the boom is actually installing. For each major AI investment, procurement, partnership, or data-center commitment, the account should separate the productive claim, financial claim, dependency claim, infrastructure claim, public-burden claim, and safety claim. The point is not to predict the crash. It is to prevent a forecast from becoming infrastructure before the public record can distinguish capacity from theater.

A strong installation account starts with the use case and baseline. What task is being changed, what non-AI process is being compared, what cost and quality measures existed before deployment, and what evidence would count against renewal? It then maps the substrate: model or vendor, cloud region, chip or accelerator dependency where knowable, data-center or power assumption, retrieval sources, tool permissions, logs, security controls, worker impact, accessibility path, and fallback if the vendor, price, policy, model, or grid connection changes.

The CISA and G7 Software Bill of Materials for AI - Minimum Elements guidance is useful here because it treats AI systems as supply-chain objects with models, datasets, infrastructure, security properties, and performance indicators, not as a single magic capability. A bubble ledger without a component record can see the money but not the machine. An AI bill of materials, system inventory, safety case, and incident log make the investment inspectable after the sales deck is gone.

The account should also record who bears downside risk. If a city signs a chatbot contract, a utility upgrades substations, a school changes assessment around generated text, a hospital adopts a clinical assistant, or a newsroom automates summaries, the account should say who can appeal, who pays if the forecast fails, who keeps the logs, who can shut the system off, and what happens to workers, users, patients, students, readers, and ratepayers when the vendor's promised value does not materialize.

This turns the AI bubble question into a governance question. Not "will the market crash?" but "what evidence, records, rights, and exit paths exist before the installation phase makes itself ordinary?" That connects this review to the AI bill of materials, safety cases, public AI registers, interconnection queues, AI system inventories, and incident reporting.

The AI-Age Reading

Read in 2026, the book is almost impossible to separate from the AI bubble question. Yet it resists the usual argument format. "Is AI a bubble?" is too small. The better question is what kind of bubble, attached to what kind of technological potential, producing what kind of institutional residue.

The disclosures above make the buildout concrete even when the payoff remains contested. They are best read as installation evidence: capital is being committed before institutions have fully settled the uses, revenue models, labor bargains, competition rules, safety practices, or public accounting needed for stable deployment.

Generative AI has many marks of an installation-phase technology: high expectations, large infrastructure spending, rapid firm formation, benchmark theater, uncertain revenue models, labor anxiety, regulatory catch-up, and a scramble to define the new common sense of work. The phrase "AI" also bundles unlike things: foundation models, chips, data centers, coding agents, recommender systems, synthetic media, robotics, office automation, surveillance tools, tutoring systems, search interfaces, and military applications. Perez's framework helps separate a general-purpose technological wave from the financial stories attached to particular firms.

Her own later research project places artificial intelligence as a new wave within the information-technology transformation rather than as a standalone civilizational reset. That distinction is useful. AI may be revolutionary without being a wholly separate revolution. It can deepen the logic of the information age: networking, intangible value, services, software coordination, flexible organization, data extraction, and automated classification.

The practical risk is that institutions mistake installation for destiny. A company buys agents because competitors bought agents. A school adds detection software because generated text unsettled assessment. A government signs vendor contracts because procurement needs an AI line item. A newsroom automates summaries because the traffic model demands speed. Each decision may be locally rational while still building a fragile social settlement around unproven assumptions.

The Institutional Question

Perez is at her best when she insists that deployment is political. A technological revolution does not naturally distribute its benefits. It needs rules, standards, public investment, labor bargains, competition policy, education, safety systems, infrastructure planning, and social protections. Without those, the new paradigm can increase productivity in narrow zones while leaving broader society with precarity, monopoly, speculative rents, and resentment.

This is the bridge to recurring concerns about legibility, institutions, labor, and machine-mediated reality. AI systems make work and behavior newly readable. They also make institutions dependent on vendors, models, metrics, and interfaces that can narrow what counts as competent action. The question is not whether AI can automate a task. The question is whether the surrounding institution can govern the changed task after the interface becomes normal.

Deployment requires boring powers that hype culture dislikes: audit rights, procurement discipline, appeal processes, labor consultation, data governance, public records, safety cases, incident reporting, interoperability, and the right to refuse systems that make people legible without making power accountable. Perez's framework makes those concerns central rather than secondary. The golden age, if it comes, is not produced by better demos. It is produced by institutions that can turn technical capacity into shared capability.

Governance and Safety

The governance lesson is not "stop investment." It is "make installation accountable before it becomes destiny." An AI buildout touches balance sheets, power grids, water systems, labor markets, public procurement, cloud competition, education, healthcare, media, and security. A serious institution should therefore ask what useful capacity is being built, which dependencies are being created, who can inspect the claim, who can exit, and what harms are being treated as acceptable installation costs.

The energy layer makes this visible. The International Energy Agency's 2025 Energy and AI report projects that global data-center electricity demand could more than double by 2030 to around 945 terawatt-hours, with AI as a significant driver. That does not make AI illegitimate. It does mean data-center siting, grid connection, water use, emissions, reliability, and local community impacts belong inside AI governance rather than in a separate facilities appendix.

Large-load governance now turns AI procurement into grid governance. FERC's June 2026 action and NERC's May 2026 guideline show that data-center commitments can affect interconnection queues, transmission planning, resource adequacy, real-time operations, and restoration planning. If ordinary ratepayers, utilities, or local governments absorb upgrade risk while private deployment value remains speculative, the relevant safety question is not model alignment alone. It is whether public systems are being asked to underwrite private forecasts without public rights, records, or recourse.

The market-structure layer matters too. The FTC's 2025 staff report on partnerships between cloud service providers and AI developers describes cloud providers as suppliers of key AI inputs, including compute, and discusses cloud commitments, preferential terms, information rights, exclusivity concerns, and switching costs. Perez's finance/production distinction becomes concrete here: if production capacity is mediated by a few capital-intensive cloud and chip bottlenecks, speculative finance can harden into durable dependency.

Bubble governance should therefore be countercyclical. During the boom, require records before irreversible dependence forms: use case, baseline, model or vendor owner, data rights, energy assumption, cost horizon, failure mode, fallback process, worker and user impact, audit access, exit terms, and decommissioning plan. During a downturn, preserve the same records so distressed vendors, abandoned pilots, or underused data centers do not turn into surveillance resale, unsafe cost cutting, stranded public obligations, or quiet concentration through acquisition.

The safety controls should be operational. Public agencies and large firms should map AI expenditures to specific use cases, baselines, owners, energy assumptions, data rights, vendor exit terms, audit rights, incident reporting, worker consultation, accessibility, security, and post-deployment review. NIST's AI Risk Management Framework is useful because it treats AI risk as lifecycle work across govern, map, measure, and manage. A benchmark score or pilot demo is not a deployment case. A deployment case has to say what changed, for whom, at what cost, with what fallback if the forecast fails.

Bubble safety also includes disclosure hygiene. Investors, procurement officers, public agencies, schools, hospitals, and employers should be able to tell whether "AI" means a frontier model, a rules engine with branding, a vendor-hosted workflow, a human labor pipeline, a fine-tuned internal system, or a spreadsheet wrapped in a chatbot. AI-washing is not only a market-integrity problem. It is a deployment-safety problem because exaggerated capability claims can move systems into consequential workflows before the evidence exists.

Where the Book Needs Friction

The danger of a powerful cycle theory is that it can become too smooth. The past does not guarantee the next turning point. Climate limits, geopolitical fragmentation, supply-chain shocks, demographic change, ecological damage, monopoly power, and model opacity may make the current information-age transition less cooperative than earlier patterns suggest.

There is also a moral risk in treating bubbles as historically functional. If speculative excess later leaves useful infrastructure, that does not excuse the harms distributed along the way: lost savings, layoffs, distorted public priorities, extractive labor, environmental damage, and communities forced to host infrastructure before they share in the gains. A bubble can build roads to the future while also deciding who gets run over during construction.

The book also gives less attention to the cultural and psychological dimensions that now matter for AI: synthetic intimacy, automated persuasion, generated belief environments, identity formation, and the way conversational systems can become private reality tutors. For that, it needs to be read alongside media theory, platform governance, cult-dynamics work, and studies of human-machine cognition.

What This Changes

The book's most useful warning is that a technological revolution is not validated by a stock chart, a demo, or even a real productivity breakthrough. It is validated, if at all, by the institutions that form around it.

For AI, that means looking past the surface drama of boom and bust. The important residues may be less visible: data-center politics, cloud dependency, labor deskilling, agent permissions, model audit norms, synthetic-media habits, procurement templates, safety reporting, classroom redesign, and the slow normalization of machine-readable work. The crash, should one come, will not automatically restore judgment. It may simply leave behind a partially built cognitive infrastructure with weaker owners and stronger incentives to monetize whatever remains.

Perez gives readers a disciplined way to stay double-minded. The technology can be real and the valuation absurd. The bubble can be destructive and still build infrastructure. The crash can be painful and still open a political choice. The decisive question is what gets deployed after the fever: a society organized around extraction and automated authority, or one that turns new tools toward wider competence, repair, and accountable power.

Source Discipline

This review separates book metadata, theory, company disclosure, regulator evidence, grid guidance, and interpretation. Publisher and bibliographic sources establish the book. Company earnings materials establish what firms chose to report about AI revenue, capital expenditure, and infrastructure investment; they are not independent audits of social value. The IEA establishes an energy-demand scenario, not a moral verdict. FERC and NERC establish large-load governance and reliability concerns, not proof that any specific data center is harmful or useful. The FTC establishes cloud and AI partnership concerns, not a final antitrust judgment. NIST establishes a risk-management vocabulary, not proof that any specific deployment is safe.

For bubble claims, evidence burdens should stay separate. A high valuation does not prove fraud. A real product does not prove a fair price. A useful pilot does not prove economy-wide productivity. A capex plan does not prove durable demand. A skeptical essay does not prove collapse. SEC AI-washing actions and investor alerts show why claims about AI use and capability require support, but they do not turn every optimistic AI statement into misconduct. Source discipline means naming which claim the source supports and leaving the other claims open.

This is not investment advice and it does not predict a crash date. The page reads public disclosures, regulator reports, and standards guidance as evidence about institutional formation: what is being built, what is being promised, what risks are visible, and what records should exist before a speculative installation phase becomes ordinary infrastructure.

The same rule applies to the site's recurring concern with machine-mediated reality. If AI systems reshape work, memory, media, and institutional defaults, the question is not whether a model is magical. It is whether the deployment record shows accountable benefit, controllable dependency, measured harm, and a path to revise or exit when the forecast fails.

This page makes no claim that AI systems are conscious, divine, AGI, or inevitable. It treats AI as a set of engineered systems, businesses, infrastructures, labor arrangements, and institutional choices whose consequences can be inspected without turning capability claims into metaphysics.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books