THE AI BUBBLE: WHY MORE THAN $1 TRILLION PRODUCED SO LITTLE

Aug 30

18 min read

We believe artificial intelligence can be a useful and powerful tool, but it has serious—and sometimes dangerous—limitations. It can assist human beings, accelerate certain tasks and improve access to information, but it does not possess genuine understanding, independent judgment or human intelligence.

We do not believe AI will ever replace human intelligence or eliminate jobs on the frightening scale repeatedly predicted by technology executives and investors.

The claims remain promises rather than demonstrated reality.

Readers are invited to explore our full analysis below and watch Ed Zitron’s complete interview (at the bottom of this article) to consider the evidence and reach their own conclusions.

Artificial intelligence is being promoted as the next great technological revolution—a force expected to transform employment, education, medicine, finance, business, government and nearly every part of modern life.

British technology critic Ed Zitron believes the industry’s promises, financial claims and actual products do not come close to matching the extraordinary amount of money being spent. For perspective, $1 trillion is equal to one million multiplied by $1 million.

Zitron is not arguing that generative-AI tools such as ChatGPT, Claude, Gemini and Microsoft Copilot have no practical value. He acknowledges that they can assist with computer troubleshooting, basic coding, document summaries, preliminary research, image generation and certain routine tasks. His argument is that these systems are not genuinely intelligent; they are simply large language models that generate outputs by identifying patterns and predicting likely sequences from the data on which they were trained.

His central argument is that these useful but limited products could have been developed and delivered with a fraction of the money poured into the industry.

According to Zitron, approximately $30 billion (30,000 multiplied by $1 million) in properly directed investment could have produced essentially the same level of generative-AI products and services available today. Instead, technology companies have spent more than $1 trillion (one million multiplied by $1 million) in capital expenditure on specialized chips, enormous data centers, electrical infrastructure and the computing systems required to train and operate generative-AI models.

That is the first problem: massive overinvestment without a corresponding improvement in the product.

The second problem is financial: after more than $1 trillion in capital expenditure, the largest companies involved are producing only tens of billions of dollars in directly attributable AI revenue.

Based on these figures, the industry has spent $33 for every $1 in AI-related revenue generated—and revenue is not profit.

In Zitron’s blunt formulation, the math is not mathing.

TWO DIFFERENT FINANCIAL FAILURES

Zitron identifies two separate problems that should not be confused.

First, he believes the industry could have provided today’s level of generative-AI service with approximately $30 billion in investment. Spending well over thirty times more has not produced artificial general intelligence, autonomous digital employees or systems capable of consistently reasoning without human supervision.

Second, the industry has spent more than $1 trillion while producing only tens of billions of dollars in identifiable AI revenue—and revenue is not the same as profit.

OpenAI and Anthropic remain enormously expensive to operate, while the major cloud companies like Microsoft, Amazon and Google, continue committing greater sums to data centers, specialized chips and electrical capacity.

Zitron points to Microsoft as one example. The company reportedly generated approximately $34.3 billion in AI-related revenue during its 2026 fiscal year, with roughly $24.1 billion connected to OpenAI. That would leave approximately $10 billion revenue attributed to its remaining AI business during a year in which Microsoft recorded approximately $115 billion in total capital expenditure.

Zitron argues that, after nearly four years of intensive generative-AI development and commercialization, the figures reveal an extraordinary imbalance between the scale of the infrastructure buildout and the independent commercial demand demonstrated so far.

POPULARITY DOES NOT NECESSARILY MEAN PROFITABILITY

Hundreds of millions of people use generative-AI tools, but many use them free of charge or through $20-per-month subscriptions that may not cover the full computing cost of providing the service.

Every prompt requires a process called inference, during which expensive computer hardware calculates and generates a response. Companies must also pay for model training, electricity, cooling, specialized chips, data centers, engineers, licensing, data acquisition and continual model development.

The problem, according to Zitron, is that increased use also means increased expense.

A conventional software company may develop a program once and distribute additional copies at relatively little cost. Generative AI must perform substantial computation every time a user asks it to generate an answer.

If revenue rises while computing expenses rise along with it, enormous user numbers do not automatically create a path to profitability.

THE BUSINESS MODEL IS BASED ON HOW MUCH THE AI “TALKS”

Unlike conventional subscription software, generative AI has a continuing cost every time it communicates with a user.

The industry measures this activity in tokens—small number of of words processed when a user submits a request and when the model generates its response. A longer question requires more input tokens. A longer answer requires more output tokens. Uploading documents, analyzing large amounts of text, generating computer code or conducting research require vastly more processing.

In practical terms, the business model is based upon the volume of words exchanged between the customer and the AI.

This creates a fundamental problem with the standard $20-per-month consumer subscription.

A light user who asks a few short questions may cost the ai provider relatively little. A heavy user who uploads documents, generates thousands of lines of code, creates images or conducts repeated research can consume far more computing capacity than the same $20 monthly payment covers.

The provider receives a fixed subscription fee while its expense changes with every prompt, response and research task.

That is the opposite of the highly profitable economics traditionally associated with consumer software. Once conventional software has been developed, serving another customer may cost very little. Generative AI must continue building immensely costly computing power to produce every new answer.

The longer and more complicated the request, the greater the expense.

Zitron argues that the $20 subscription price does not economically match the service being offered—particularly when companies encourage customers to use AI constantly for writing, coding, research, images, video, financial analysis and other processing-intensive work. AI features are also being pushed across email services, search engines, office software, online shopping platforms, social media and other everyday digital services, repeatedly prompting users to generate summaries, compose messages, compare products or ask AI questions—whether they requested those features or not.

Providers attempt to control this mismatch through:

Message and usage limits
Slower service after customers reach certain thresholds
Restricted access to the most capable models
Separate charges for advanced research, images, video and large files
More expensive professional and enterprise plans
Token-based charges for application programming interfaces
Shortened research processes and limits on the number of sources examined

These restrictions are not merely product-design choices. They help prevent individual customers from consuming more computing resources than their subscriptions can support. The same cost controls may also limit AI research: instead of crawling millions of potentially relevant websites, a typical research task may examine only a selected group of approximately 30 to 40 webpages before generating what appears to be a comprehensive answer—ironically marketed as “deep research.”

The economics become even more strained when companies offer free access to attract hundreds of millions of users. Those users may produce impressive adoption statistics, but they still generate processing expenses without providing corresponding subscription revenue.

The industry is therefore caught between two conflicting objectives. It must encourage enormous use to justify its valuations and infrastructure investments, but every additional interaction costs money. Charging customers the full cost could discourage adoption, while maintaining artificially low prices requires the companies or their investors to absorb the remaining costs—and the resulting financial losses.

Future advertising may be presented as a solution, but advertising introduces additional concerns about privacy, commercial influence and whether AI answers will be shaped to benefit advertisers. It also remains unclear whether advertising revenue could support the enormous computing expenses involved.

The $20 subscription model makes generative AI appear affordable to consumers. It does not prove that the underlying service is economically sustainable.

THE PROBLEM OF CIRCULAR REVENUE

OpenAI and Anthropic have become some of the largest customers of the AI infrastructure being built by Microsoft, Amazon, Google, Oracle and other companies.

At the same time, many of those corporations are also investing directly in OpenAI and Anthropic.

This creates a potentially circular system: a technology company like Microsoft, Amazon, Google, Oracle or others invests billions in an AI developer like OpenAI, and the developer then spends much of that investment purchasing computing services, data-center capacity or chips from the investor or its partners.

The transactions produce recorded revenue. But Zitron questions whether that revenue represents broad, independent and sustainable customer demand—or money circulating among companies that depend upon one another to preserve the appearance of continued growth.

The apparent AI boom may therefore be supported partly by technology companies financing their own largest customers.

UNRELIABLE ANSWERS ARE NOT A MINOR DEFECT

The economic problem is only part of Zitron’s criticism.

Large language models (LLMs—computer systems trained on enormous amounts of text to identify patterns and predict which words are most likely to come next) do not retrieve truth in the way a conventional database retrieves a stored fact. They generate responses by predicting statistically likely sequences of words. This allows them to produce fluent, confident and persuasive answers even when the underlying information is incomplete, unsupported or often entirely false.

These fabricated responses are commonly called hallucinations.

In casual situations, an incorrect answer may be merely frustrating. In higher-risk settings, it can be dangerous.

A hallucinated restaurant recommendation or incorrect movie title is an inconvenience. Fabricated medical guidance, an invented legal precedent, an incorrect tax recommendation, a faulty financial calculation or unsafe instructions for handling chemicals can cause serious harm.

In an interview with the Carnegie Endowment for International Peace, Zitron described tests in which an AI system reportedly recommended methods of extinguishing certain chemical fires that could instead have caused explosions. His broader point was that an answer sounding professional is not evidence that it is correct.

Medical, financial, legal and public-safety information is particularly dangerous because people may act upon it before consulting a qualified professional. Even when an AI system includes citations, those citations may not support the claim, may be taken out of context or may not even exist.

This creates another mismatch between the promise and the product. Companies are presenting generative AI as a dependable assistant and possible replacement for professional labor while the technology still requires people to verify its important conclusions. The platforms themselves typically display disclaimers warning that AI can make mistakes and that users should verify important information—effectively acknowledging that their answers cannot be trusted without human review.

If every consequential answer must be independently researched and verified by a human, the claim that AI delivers major productivity gains becomes fundamentally questionable: uncovering and correcting a confidently presented hallucination—a fabricated or made-up answer—requires considerably more time, than completing the work correctly in the first place.

“DEEP RESEARCH” IS NOT A DEEP SEARCH OF THE INTERNET

Zitron also challenges the way AI companies market their so-called “deep research” systems.

In many consumer-facing research runs, an AI system may crawl or consult only select 40 to 50 webpages before producing what appears to be a comprehensive report.

That number may sound substantial until it is compared with the millions of active websites, publications, government records, court filings, institutional archives, research papers and specialized databases that may contain relevant information within any major field.

The result is not an exhaustive investigation of the available evidence. It is a limited sample selected and summarized through a search process that may not be visible to the user.

This limitation is largely driven by the economics of processing costs.

Searching, retrieving, reading, comparing and processing substantially more material requires additional computing power. Additional computing power costs money. To keep response times and operating expenses manageable, AI systems restrict how many sources they examine, compress the material they retrieve and take computational shortcuts when constructing an answer.

Those shortcuts can cause a system to overlook decisive evidence buried deeper in search results, local government records, court databases, academic archives, older publications or poorly indexed websites.

The system may also cite several webpages that all repeat the same original claim. That creates the appearance of confirmation without providing genuinely independent evidence.

The polished final report can therefore appear much more comprehensive than the underlying research actually was.

A reader sees citations, organized headings and confident explanations and reasonably assumes that the system conducted a wide investigation. In reality, it may have reviewed only a few dozen webpages from an information universe containing millions of potentially relevant sources.

These systems generally do not clearly disclose:

The total number of sources searched
How those sources were selected
Which potentially relevant sources were excluded
Whether primary records were reviewed
Whether several cited articles originated from the same underlying source
How contradictory evidence was evaluated
Where uncertainty or missing information remains

In practice, AI research frequently relies upon a limited selection of prominently indexed sources—often major mainstream media outlets—while millions of smaller publications, specialized websites, local news organizations, independent researchers, public records and historical archives may be ignored or excluded. Processing a much wider body of material would require additional computing power, time and expense, so potentially valuable sources are often disregarded in the interest of processing economy.

AI-assisted research can still help identify leads, organize known information and provide a preliminary starting point. But calling the result “deep research” risks overstating what was actually performed.

A report based upon 40 or 50 webpages is not equivalent to a comprehensive review of an entire subject. It should never be presented as though the system searched everything available.

This limitation reinforces Zitron’s larger criticism: after more than $1 trillion in capital expenditure, the industry is offering research systems that examine a remarkably narrow portion of the available internet, take shortcuts to control processing costs and can still produce confident but incomplete or false conclusions.

ARTIFICIAL INTELLIGENCE WITHOUT ACTUAL INTELLIGENCE

Zitron argues that the term artificial intelligence gives the public a fundamentally misleading impression of what large language models have achieved.

Despite the enormous investment, these systems have not developed anything approaching human intelligence, independent understanding or genuine reasoning. They do not comprehend the meaning of their answers in the way a person does. They recognize patterns in enormous quantities of data and calculate which words, images or pieces of code are statistically likely to come next.

That process can produce remarkably convincing results, but convincing language is not proof of intelligence.

An AI system can explain a complicated subject clearly in one response and then make an elementary factual or logical error in the next. It can summarize a document without understanding its wider significance, write functional code without knowing why the program works and generate authoritative medical or financial guidance without recognizing the consequences if its answer is wrong.

Zitron argues that the industry has deliberately blurred the distinction between generating an intelligent-sounding answer and possessing intelligence.

Terms such as “reasoning,” “thinking,” “learning” and “understanding” encourage consumers and investors to imagine a digital mind developing behind the screen. Zitron’s position is that no such intelligence has emerged.

After more than $1 trillion in capital expenditure, the industry has produced increasingly sophisticated prediction systems—not artificial minds.

That distinction matters because the promise of eventual “true intelligence,” artificial general intelligence or superintelligence is being used to justify current spending. Investors are not simply financing a useful writing or troubleshooting tool. They are being asked to finance the anticipated creation of systems that could replace enormous amounts of human labor, discover new medicines, operate businesses and transform the global economy.

Zitron contends that the industry is nowhere near achieving those results. More importantly, he sees no convincing evidence that building more data centers and feeding existing language-model systems more computing power will turn statistical prediction into genuine intelligence.

The industry’s most important product may therefore be the promise of what AI will someday become—not what it can reliably accomplish today.

AI IS NOT REPLACING THE HUMAN WORKFORCE ON A MASSIVE SCALE

For several years, technology executives have predicted that artificial intelligence will eliminate enormous portions of the human workforce—particularly accountants, attorneys, programmers, administrators, analysts and other white-collar professionals.

Zitron argues that this promised mass replacement is not happening, and will not happen.

Generative AI can assist workers with individual tasks. It can produce a first draft, summarize documents, generate basic computer code or reorganize existing information. But completing portions of a job is not the same as independently assuming the responsibilities of an entire occupation.

Most professional work requires judgment, accountability, institutional knowledge, communication, verification and an understanding of consequences. Large language models cannot reliably provide those qualities without continuing human supervision.

Zitron specifically disputes predictions that AI will soon eliminate a substantial portion of white-collar employment. Anthropic CEO Dario Amodei has warned that AI could eliminate approximately half of entry-level white-collar jobs within several years. OpenAI CEO Sam Altman and other technology leaders have made similarly dramatic predictions about the replacement of programmers and other professionals.

Yet the world has not yet experienced and will not experience anything approaching that level of workforce replacement.

Businesses may announce layoffs while mentioning AI, but that does not prove AI successfully performed the work of the eliminated employees. Companies also reduce staff because of declining revenue, economic uncertainty, restructuring, outsourcing, excessive pandemic-era hiring or pressure to increase profits.

In some cases, executives may describe conventional cost-cutting as “AI transformation” because investors reward companies for appearing technologically advanced.

Reduced hiring is also not necessarily proof of improved productivity. A company may simply require its remaining employees to carry heavier workloads, repair AI-generated mistakes or produce the same work with fewer resources.

Zitron points out that the people most enthusiastic about replacing professional workers are often senior executives rather than the employees performing the detailed work.

A law-firm partner may praise AI’s ability to conduct legal research, while associates must verify whether it found the controlling precedent, misinterpreted a decision or invented a case. An executive may celebrate AI-generated software while engineers spend additional time identifying faulty or insecure code.

The technology may make certain easy tasks faster while making difficult work harder—particularly when employees must locate errors hidden inside polished, convincing output.

Zitron acknowledges that some occupations have already experienced disruption, including transcription, translation, illustration and other forms of creative or contract work. But he argues that these examples do not demonstrate successful replacement across the wider professional workforce. They often show employers accepting cheaper, lower-quality output or shifting the burden of correcting mistakes onto remaining workers.

The distinction is critical:

Eliminating a job is not the same as successfully automating its work.

If AI were producing dramatic, economy-wide productivity gains, companies should be able to demonstrate clear increases in revenue per employee, lower operating expenses, improved output and reliable performance. Zitron argues that convincing evidence of those gains has not appeared at anything close to the scale promised.

Instead, the industry continues offering predictions about what AI will accomplish several years from now.

The anticipated transformation is always approaching, yet never fully arriving: the next model will reason reliably, the next generation will become autonomous, the next data center will reduce costs and the next software agent will finally replace entire categories of workers.

Promises upon promises—but not mass workforce replacement.

AI may continue changing how people perform particular tasks. Some occupations may shrink, new roles may develop and workers may be expected to incorporate AI tools into their existing responsibilities.

None of that establishes that generative AI is close to replacing the human workforce on a massive scale—especially in white-collar professions where accuracy, responsibility and human judgment remain essential.

The human workforce is not disappearing. The promise that it soon will remains one of the AI industry’s most effective—and economically valuable—marketing claims.

WHY SPENDING MORE HAS NOT ELIMINATED THE PROBLEMS

The extraordinary infrastructure expansion has produced faster models, longer context windows, more polished interfaces and new image, video and coding capabilities.

However, it has not eliminated the fundamental problems of hallucinations, incomplete research, unpredictable outputs and high processing costs.

The industry continues to promise that larger models, additional data centers and more advanced chips will solve these shortcomings. Zitron questions how long the public should continue accepting that explanation when each new generation requires greater investment without delivering dependable reasoning or a sustainable business model.

If approximately $30 billion could have supported the present level of service, the remaining expenditure did not buy a proportionately better product. It financed a race to build capacity based upon expectations of future demand that has not yet been proven.

ZITRON’S MID-2027 PREDICTION

Zitron predicts that the AI boom could begin seriously unraveling by approximately mid-2027.

This is not a prediction that every AI tool will suddenly stop operating or that useful generative-AI software will disappear. His prediction concerns the financial system constructed around the technology.

By that point, he expects the gap between spending, revenue and profitability to become increasingly difficult to conceal. OpenAI, Anthropic and other AI companies will require additional financing to pay their enormous computing commitments. Data-center operators, cloud providers and investors will want evidence that the promised demand and revenue are materializing.

If the companies cannot provide that evidence, Zitron believes lenders and investors may finally refuse to provide another round of funding at ever-higher valuations.

That is when, in his view, the gig may be up.

The collapse could begin when one major participant admits that projected demand has not appeared, reduces its capital spending, fails to raise sufficient money or cannot fulfill its cloud and data-center commitments.

Such an event could force investors to reevaluate the entire chain of AI companies, infrastructure providers and financial arrangements built upon the same expectations.

His mid-2027 timeline remains a prediction, not an established fact. Financial bubbles can continue longer than critics expect, particularly when the companies supporting them possess enormous cash reserves and political influence.

Nevertheless,

Zitron argues that money cannot conceal the underlying economics indefinitely.

At some point, the industry must demonstrate that its products can generate enough independent revenue to pay their operating expenses, finance their infrastructure and provide a return on the extraordinary capital invested.

WHY A COLLAPSE COULD REACH BEYOND SILICON VALLEY

Zitron warns that the AI boom is no longer confined to a collection of speculative startups.

AI investment is now connected to some of the world’s largest public corporations, the stock market, retirement accounts, semiconductor manufacturers, utilities, private-equity firms, venture-capital funds, commercial real estate and a growing amount of corporate and private debt.

The valuations of Nvidia, Microsoft, Amazon, Google, Meta, Oracle and numerous infrastructure companies increasingly reflect expectations of enormous future AI growth. Data centers are being financed and constructed based on the assumption that demand will continue rising for years.

If OpenAI, Anthropic or another major customer cannot fulfill its financial commitments, projects built around that projected demand could be delayed, abandoned or written down. Investors could begin questioning whether the broader infrastructure boom was justified.

Zitron believes OpenAI is especially important because so many companies have tied their investments, cloud revenue and growth projections to its continued expansion. He warns that its failure could produce a chain reaction across technology stocks, data-center developers, lenders, venture-capital portfolios and retirement funds heavily exposed to the largest technology companies.

That does not necessarily mean the entire economy would disappear. It means an AI crash could contribute to falling markets, layoffs, construction cancellations, credit losses and a broader economic contraction.

Ironically, in Zitron’s analysis, AI may threaten far more jobs through the economic damage caused by a collapsing investment bubble than it has successfully replaced through automation.

A WARNING—NOT AN ESTABLISHED OUTCOME

Zitron’s conclusions remain his analysis, not a proven forecast.

Supporters of the industry argue that AI is still in an early infrastructure-building phase, similar to the early development of railroads, electrical systems, telecommunications or the internet. They believe revenue will eventually catch up as adoption increases, models become more efficient and businesses discover more valuable applications.

Zitron rejects that comparison.

He argues that fiber-optic networks and traditional internet infrastructure retained broad usefulness after the dot-com crash, while specialized AI chips and energy-intensive GPU data centers have more limited alternative uses and remain costly to operate.

His position is not simply that the industry has spent too much. It is that the industry has spent too much on products that remain unreliable, expensive to operate and incapable of delivering many of the transformational promises used to justify the investment.

A CRITIC WORTH HEARING

Ed Zitron hosts the Better Offline podcast, which reportedly reaches more than one million monthly downloads, and publishes the technology newsletter Where’s Your Ed At. He is the founder and CEO of EZPR and is writing The Hater’s Guide to Silicon Valley, expected in the second quarter of 2027.

His language is deliberately confrontational, and his predictions are considerably more pessimistic than those of mainstream technology executives and many financial analysts.

Nevertheless, the questions he raises deserve serious consideration:

Why did the industry spend more than $1 trillion if approximately $30 billion could have produced the services available today?
Why is directly attributable AI revenue still so small compared with the infrastructure investment?
How much apparent demand is independent, and how much is subsidized or created through circular financial relationships?
Can a fixed $20 consumer subscription support a service whose cost rises with every word processed and generated?
Can AI companies produce reliable profits when increased use requires additional expensive computation?
Why do “deep research” systems inspect only a few dozen webpages when millions of potentially relevant sources exist?
How can systems with persistent hallucination problems be trusted in medicine, finance, law and other consequential fields?
After more than $1 trillion in investment, why has the technology not developed anything approaching true intelligence?
Where is the promised mass replacement of the white-collar workforce?
Who will ultimately absorb the losses if the predicted demand never materializes?

Whether Zitron proves correct or overly pessimistic, the burden of proof should not rest only upon the critics.

The companies spending more than $1 trillion—and asking investors, governments, utilities and communities to support an even larger expansion—must demonstrate that the promised economic and social benefits are real.

After investment on this scale, it is reasonable to ask why the products remain unreliable, the companies remain unprofitable, the research remains severely limited, the human workforce remains indispensable and the transformative future always appears to be just over the horizon.