The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron

Ed Zitron believes the generative AI industry is fundamentally misleading the public. He argues that companies have presented large language models as intelligent, autonomous technology capable of transforming civilization while exaggerating their capabilities, reliability, economic value, and financial prospects.

Zitron's criticism comes from roughly 16 years working in technology public relations. He describes himself as a technology enthusiast rather than an opponent of innovation. His objection is to powerful companies making extraordinary promises without providing evidence that their products can deliver.

He distinguishes generative AI from other forms of artificial intelligence. Large language models generate text, images, code, and other content, whereas technologies behind protein folding, autonomous vehicles, and robotics can operate through different systems. He argues that generative AI companies exploit the broader term “AI” to take credit for breakthroughs unrelated to their products.

The central financial concern is that AI investment vastly exceeds demonstrated demand. Zitron claims the industry has already spent more than $1 trillion on infrastructure and is contemplating another trillion dollars, while the resulting commercial revenue remains comparatively small and heavily subsidized.

OpenAI and Anthropic are central to what Zitron calls a circular financing system. He says roughly 70% of the AI revenue attributed to major cloud providers comes from these two companies, which depend on financing from many of the same technology companies selling them computing services.

Amazon, Google, Microsoft, and Nvidia are simultaneously investors, suppliers, and beneficiaries of AI spending. Zitron cites Amazon providing $50 billion to OpenAI and $5 billion to Anthropic, alongside Google's $10 billion investment in Anthropic. His argument is that financial support flows back to investors through purchases of computing infrastructure and cloud services.

Wall Street's expectations depend heavily on continued spending by OpenAI and Anthropic. Zitron says analysts project more than $400 billion in revenue from these two customers over approximately three and a half years, contributing substantially to expected cloud growth of roughly 30%.

The dependence on two unprofitable customers makes projected cloud growth vulnerable. If OpenAI or Anthropic cannot raise sufficient capital, Zitron argues that Microsoft, Amazon, Google, and their infrastructure partners could lose a significant source of anticipated growth.

Major technology companies do not adequately disclose their generative AI revenue, according to Zitron. He believes their unwillingness to provide straightforward figures is particularly troubling because they repeatedly describe AI as the most important technological development in decades.

Annualized revenue run rates can create a misleading impression of actual sales. Zitron explains that companies sometimes multiply one month's revenue by 12 or use another extrapolation without clearly defining the calculation. Such figures are not necessarily equivalent to revenue already earned.

Zitron treats undisclosed AI revenue as a warning sign rather than a neutral omission. He argues that publicly traded companies have strong incentives to advertise successful investments, making their reluctance to publish precise AI financial results significant.

Microsoft's reported AI revenue illustrates the scale of the dependence. Zitron cites Bloomberg figures indicating approximately $34.33 billion in Microsoft's fiscal 2026 AI revenue, of which about $24.1 billion came from OpenAI, leaving roughly $10 billion from other customers.

Microsoft's infrastructure spending is much larger than its reported AI revenue. Zitron contrasts approximately $115 billion in capital expenditure during fiscal 2026 with plans to spend roughly $175 billion the following year. He considers the gap difficult to justify through existing AI demand.

Nvidia's enormous GPU sales are another part of the imbalance. Zitron cites approximately $215.9 billion in its latest fiscal-year revenue, largely from GPUs, while estimating that AI revenue outside OpenAI and Anthropic amounts to only about $22 billion.

Growth in existing technology businesses has been mistaken for proof that AI investments are paying off. Zitron says Microsoft's software business, Google's advertising operations, Amazon's commerce and advertising businesses, and Meta's advertising systems have continued generating revenue through conventional business improvements and pricing changes.

Amazon's growing advertising business is an example of revenue growth that should not automatically be attributed to generative AI. Zitron argues that investors and commentators often connect successful earnings announcements to AI spending without distinguishing the underlying sources of revenue.

Zitron believes rising stock prices have substituted for evidence of economic success. Investors see major companies spending heavily on GPUs and assume the executives must know something valuable, even when the companies have not demonstrated corresponding returns.

He does not believe AI companies necessarily intended to build unprofitable businesses. His explanation is that executives probably expected improving models, cheaper chips, and stronger customer demand eventually to produce sustainable economics.

Zitron argues the industry should have reconsidered its strategy in 2024. He cites OpenAI losing more than $5 billion that year as an early warning that the technology's economics were not improving quickly enough.

The continued investment nevertheless benefited technology valuations. Zitron believes heavy GPU purchases helped raise expectations and share prices for Nvidia, Microsoft, and other technology companies, creating incentives to continue spending even without convincing returns.

The interviewer argues that investing before profitability is normal for transformative technologies. He cites the internet, automobiles, and disruptive innovations that initially appeared expensive, unreliable, and economically inferior to existing alternatives.

Zitron rejects the assumption that generative AI will follow the same improvement curve. He argues that previous technologies becoming cheaper and more capable does not establish that large language models will achieve comparable gains.

Nvidia's CUDA software helped make modern AI development possible. Zitron describes CUDA as an established system for running computations on GPUs that developed over many years before the current generative AI boom.

Moore's law does not guarantee equivalent improvements in GPU-based AI economics. Zitron argues that projections based on steadily improving computing power cannot simply be transferred to the costs of training and operating large language models.

AI infrastructure requires concentrated amounts of electricity, hardware, and capital. Zitron explains that advanced AI GPUs consume substantial power, use high-bandwidth memory, and must be deployed in clusters containing thousands or hundreds of thousands of chips.

OpenAI and Oracle's Stargate project in Abilene, Texas, is his principal infrastructure example. He describes a planned 1.2-gigawatt facility with eight buildings containing approximately 50,000 Nvidia GB200 GPUs each, illustrating the extraordinary concentration of computing resources.

Building AI infrastructure increasingly involves borrowing rather than simply using existing corporate cash. Zitron argues that companies are committing billions to facilities whose future profitability depends on demand that has not yet been demonstrated.

The cost of serving AI customers continues even when models produce incorrect answers. Customers using token-based services pay for computational work regardless of whether the output is useful, accurate, or requires extensive correction.

AI pricing is based partly on tokens, approximately three-quarters of a word on average. Providers charge for input tokens supplied by users and output tokens generated by models, including tokens consumed during internal reasoning processes.

Flat-rate subscriptions conceal the underlying economics of AI usage. Users may pay $20 or $200 monthly without understanding how much computation their activity consumes, while providers impose usage limits that are not always transparent.

Zitron cites SemiAnalysis estimates showing how much usage subscriptions can permit. According to his account, a $200 ChatGPT plan could provide up to $14,000 in token usage at comparable API prices, an Anthropic plan up to $8,000, and a $20 subscription approximately $400. These are comparisons with token pricing, not disclosed provider production costs.

Zitron cites a $20.9 billion annual loss for OpenAI. He argues that subsidizing intensive users helps explain why rapidly growing adoption has not translated into sustainable profitability.

Actual operating losses per user remain unclear. Zitron acknowledges that companies do not disclose sufficient information to calculate precise customer-level subsidies and that API-equivalent token values should not be mistaken for exact production expenses.

Enterprise pricing has exposed resistance to the real cost of AI. Zitron says OpenAI began moving larger business customers, including organizations with more than approximately 150 employees, toward usage-based charges around March 2026.

Uber reportedly consumed its annual AI token budget in three months. Zitron presents this as evidence that companies enthusiastic about AI productivity can become much more cautious when computational usage produces a direct and substantial bill.

The interviewer accepts that AI companies may be subsidizing heavy users but questions whether this is temporary. He suggests providers could be buying market share in anticipation of future cost reductions, just as earlier technology companies operated at losses during expansion.

Zitron says there is insufficient evidence that inference costs are falling fast enough. He argues that providers are increasingly spending more to deliver more capable models, while GPU rental businesses and inference operators themselves struggle to establish attractive profits.

AI providers face difficult capacity-planning economics. Purchasing too much GPU capacity leaves expensive hardware underused, while purchasing too little causes slowdowns, unavailable services, and dissatisfied customers.

AI errors can become especially expensive in software development. A model may examine a codebase, alter files, introduce defects, or create security vulnerabilities while still consuming paid computational resources.

The interviewer presents rapid consumer adoption as strong evidence of value. He cites ChatGPT reaching 100 million users in approximately 60 days, compared with nine months for TikTok and two and a half years for Instagram.

The interviewer also cites widespread organizational use. His research indicates that approximately 88% of organizations regularly use AI in at least one business function, while roughly 95% of employees at his company use AI tools daily.

Zitron accepts that generative AI is widely used but disputes what adoption proves. He argues that usage figures cannot independently establish economic value when products are heavily subsidized and promoted everywhere.

Generative AI has been integrated into products people already use. Zitron points to Gemini in Google's products, Copilot in Microsoft Word, and Amazon's Rufus shopping assistant as examples of users encountering AI without necessarily seeking it out.

He calls this an unusually aggressive, partly non-consensual deployment of technology. He argues that users are simultaneously exposed to AI features and warned that refusing to adopt them will damage their careers or competitiveness.

AI search can be useful without being reliably accurate. Zitron acknowledges that conversational systems sometimes handle broad or exploratory queries better than conventional search, particularly when users are unsure how to describe the information they seek.

He finds generative search considerably less dependable for precise facts. Broadly finding a possible source or subject is different from accurately identifying specific dates, figures, quotations, or relationships.

The interviewer offers his fiancée's business as evidence of practical transformation. As a solo entrepreneur whose first language is not English, she uses AI to write business copy and generate images that previously required paid graphic-design assistance.

Zitron challenges whether such benefits justify the actual cost of computation. He says a stronger economic case would exist if customers remained satisfied while paying transparent usage-based prices rather than subsidized subscriptions.

AI has accelerated the production of low-quality internet content. Zitron argues that search-engine optimization already encouraged websites to publish material designed to rank rather than genuinely help readers, and generative AI has made that practice cheaper and easier to scale.

He does not believe AI invented the internet's content-quality problem. Search incentives and mass-produced low-value articles existed long before ChatGPT; AI has mainly increased the speed and volume at which such material can be created.

The interviewer argues AI should be compared with real human alternatives rather than perfection. For writing, research, and other routine work, humans also make mistakes, require payment, and take time, so a sufficiently reliable AI system may still offer a better trade-off.

The interviewer cites substantial improvements in simple-task reliability. He references a Vectara hallucination benchmark showing error rates falling from approximately 21.8% to around 0.7% among leading models over several years.

Zitron questions whether narrowly defined benchmarks represent real-world reliability. He argues that systems can become better at particular tests without becoming equally trustworthy across complex, unfamiliar, or consequential tasks.

A task-duration benchmark also prompted disagreement. Zitron describes models completing approximately hour-long tasks with success rates around 50%, questioning whether such results justify claims of dependable autonomous work.

The interviewer says his personal experience of hallucinations has improved dramatically. He recalls early ChatGPT frequently providing obviously incorrect answers but says he now encounters fewer blatant factual errors in ordinary use.

Zitron gives a specific example of a financial hallucination. While analyzing historical stock performance using Bloomberg's AI-assisted tools, he noticed a Microsoft share-price figure of $575 that he believed was incorrect.

Bloomberg's AskB remains a tool Zitron finds useful. It converts natural-language financial requests into Bloomberg Query Language and executes them against recognizable data sources, saving users from manually writing specialized queries.

Even AI tools connected to professional financial databases require verification, according to Zitron. He says he checks figures repeatedly because errors that are harmless in casual research can become serious in financial modeling, medical transcription, or investment decisions.

More detailed AI-generated reports create more opportunities for inaccuracies. Zitron argues that a long report can appear authoritative while concealing numerous smaller mistakes that readers must individually identify.

AI-generated coding mistakes can accumulate as developers become less vigilant. He describes the danger of programmers accepting increasingly large amounts of generated code without understanding the underlying behavior or thoroughly checking changes.

The discussion distinguishes producing more code from producing better software. The interviewer proposes measuring AI productivity through software shipped; Zitron replies that more output has little value if products become less reliable, maintainable, or secure.

Zitron believes contemporary software quality is declining across major technology companies. He cites frustrations with Google, Meta, Microsoft, GitHub, Amazon, and Apple's interfaces, although he acknowledges that proving a direct industry-wide connection to AI-assisted coding is difficult.

GitHub outages and instability are presented as a particular warning sign. Zitron argues that a growing volume of AI-generated code is increasing pressure on development infrastructure and contributing to software-quality problems.

The interviewer cites research suggesting software outages have increased alongside AI-assisted coding. He describes two possible mechanisms: generated code introducing more defects and expanding automated workloads placing additional strain on infrastructure.

Zitron claims Amazon Web Services experienced multiple outages connected to AI coding tools. He acknowledges that wider claims about technology becoming less stable need stronger systematic evidence than individual examples.

Open-source projects are facing an influx of poorly understood AI-generated contributions. Zitron says people with limited programming knowledge can now produce substantial code changes, sometimes without understanding the resulting systems.

AI assistance may encourage developers to neglect essential skills. The speakers agree that successful earlier outputs can make users less likely to inspect subsequent work, allowing mistakes to accumulate unnoticed.

Zitron regards human learning as fundamentally different from model context retrieval. A junior engineer or intern develops judgment through experience, relationships, workplace circumstances, and feedback, whereas an LLM primarily operates on information supplied through prompts, files, and other inputs.

The interviewer argues that customers care about useful outcomes rather than how intelligence is produced. If Claude can remember a dog's name or retrieve an accurate revenue figure from uploaded documents, he considers that function valuable regardless of whether it resembles human memory.

Zitron argues that recalling stored information is not equivalent to understanding a person. He says retrieving names, historical conversations, or work documents does not reproduce the emotional awareness and lived experience that shape human judgment.

Zitron uses his editor Matt Hughes to illustrate what AI cannot easily replace. Hughes, a Liverpool-based technology journalist and editor, contributes accumulated knowledge, research judgment, criticism, shared learning, and a trusted working relationship.

Zitron describes collaborating with Hughes on extensive research involving asset managers such as Blackstone. They exchanged findings and criticisms while producing approximately 11,000 words of writing, a process he values partly because both participants learned through the work.

Zitron says sophisticated AI research systems often produce generic reports. Having tried higher-end LLM research setups used by financial firms, he argues that their outputs tend to repeat widely available observations rather than contribute distinctive insight.

The interviewer separates the benefits of collaboration from the benefits of completing a task. He accepts that human research provides learning and connection but maintains that users sometimes need only an accurate final answer.

Zitron recommends deterministic systems for important business figures. Rather than asking an LLM to reconstruct revenue from assorted documents, he believes organizations should maintain dependable calculations and verifiable records.

Trust in human collaborators comes partly from demonstrated performance and shared context. The interviewer suggests consistent delivery can also make AI systems trustworthy, while Zitron emphasizes that relationships allow deeper understanding and correction.

The first iPhone becomes a major point of disagreement about disruptive innovation. Zitron remembers its value being immediately apparent when he demonstrated it to people using conventional mobile phones.

The interviewer cites Steve Ballmer's famously dismissive reaction to the iPhone. Microsoft's then-CEO mocked its approximately $500 price and lack of physical keyboard while praising cheaper alternatives, illustrating how experienced executives can underestimate new technologies.

Zitron argues the iPhone demonstrated concrete benefits without requiring elaborate theoretical promises. Features such as its interface, mobile browsing, and eventually the App Store could be experienced directly rather than justified through hypothetical future capabilities.

Goldman Sachs analyst Jim Covello's 2024 skepticism supports Zitron's position. Zitron references Covello's argument that technologies such as smaller wireless components provided a recognizable development path toward smartphones, whereas comparable evidence for the promised future of generative AI is missing.

Zitron says he would be considerably more receptive to honestly marketed experimental language-model software. His objection is not to researching expensive and imperfect generative systems, but to presenting them as autonomous intelligence ready to transform employment and society.

Microsoft's early Bing chatbot controversy illustrates the industry's promotional behavior. Zitron recalls journalist Kevin Roose reporting disturbing chatbot interactions and Microsoft CTO Kevin Scott treating the episode as remarkable rather than simply acknowledging unreliable software output.

The interviewer argues that technological bubbles can still produce lasting innovations. He points to the dot-com crash, after which many speculative companies failed while successful internet businesses went on to transform commerce and communication.

Historical internet skeptics made predictions that proved badly wrong. The interviewer cites economist Paul Krugman's comparison of the internet's economic impact to the fax machine and Clifford Stoll's dismissal of online shopping and digital information replacing established institutions.

Zitron notes that some early internet criticism was nevertheless valid. He considers warnings about misinformation, poor information quality, and the limitations of online education more credible than predictions that the internet would have little economic effect.

The dot-com bubble included a major overinvestment in fiber-optic networks. Zitron says infrastructure investors expected internet traffic to double roughly every 90 days when actual growth was considerably slower, creating large quantities of unused dark fiber.

Zitron considers the residual value of fiber networks different from today's specialized AI infrastructure. Fiber could eventually support many forms of growing internet traffic, whereas he argues that current GPU facilities are more dependent on expensive, specific generative AI workloads.

AI data centers retain substantial operating costs even if construction has already been paid for. Zitron emphasizes ongoing electricity, equipment, and maintenance expenses, arguing that future demand alone will not necessarily make the facilities economically attractive.

He does not accept advertised GPU efficiency improvements as proof of lower total AI costs. Nvidia's newer systems may perform more computations per unit of power, but Zitron says companies have not demonstrated sufficiently dramatic reductions in real customer operating expenses.

Robotics and scientific computing also require powerful hardware, but the needs are different. The interviewer cites DeepMind's AlphaFold, autonomous-driving simulations, genomic research, and climate modeling as examples of valuable AI applications that rely on substantial computing infrastructure.

Zitron acknowledges that scientific and autonomous-driving systems use GPUs and high-performance computing. His argument is that these applications existed before the generative AI boom and should not automatically justify the construction of enormous facilities intended primarily for LLM training and inference.

A household cleaning robot illustrates the distinction between training and operation. Zitron describes owning a Matic cleaning robot and notes that, although substantial computing may help develop such systems, the robot does not require the same ongoing data-center inference expenses as a cloud-hosted language model.

Planned data-center capacity could require vastly more demand than currently exists. Zitron cites a Sightline Climate estimate of approximately 190 gigawatts of planned capacity and offers a rough calculation suggesting $1.6 trillion–$3 trillion in annual demand would be needed to justify such infrastructure.

He contrasts those projections with what he describes as less than $130 billion in current annual AI demand. The resulting gap is central to his belief that infrastructure development is being driven by speculation rather than customers already willing to pay.

Conventional data centers should not be treated as equivalent to GPU-heavy generative AI facilities. Zitron distinguishes relatively efficient CPU-based hosting, networking, and content-delivery services from enormous clusters designed for intensive AI training and inference.

Zitron claims the two leading AI laboratories raised approximately $217 billion during 2026. He argues that this illustrates how much outside capital is required to sustain the industry's apparent growth.

Google's deteriorating search experience is partly a business-incentive problem, in Zitron's account. He recounts a 2019 internal “Code Yellow” involving concerns about declining search-query numbers and pressure to increase engagement.

Google search leader Ben Gomes reportedly opposed pursuing more queries at the expense of useful answers. Zitron says Gomes recognized that making people search repeatedly could increase advertising opportunities while reducing the quality of the product.

Google engineers also raised concerns about these incentives. Zitron mentions Shashi Thakur and the desire to alert Sundar Pichai to the potential conflict between higher query volumes and genuinely helpful search results.

Prabhakar Raghavan is central to Zitron's explanation of Google's search changes. He says Raghavan, previously responsible for advertising, pushed for query growth and later took over search leadership around 2020.

Zitron suspects Google subsequently weakened spam suppression to increase search activity. He explicitly acknowledges that he cannot prove this causal claim, although he believes changes observed by SEO professionals support his interpretation.

The popularity of adding “Reddit” to Google searches is evidence of dissatisfaction, in Zitron's view. He believes users increasingly seek firsthand discussions because ordinary search results have become crowded with commercially optimized, low-quality material.

Google's AI-generated answers may keep more users inside Google's own products. Zitron argues that placing AI responses above traditional search results allows Google to answer queries without sending as much traffic to external websites.

The interviewer correctly separates Google's search-business decisions from AI-assisted coding. Zitron concedes that managerial decisions harming search quality do not themselves prove that generated code is responsible for worsening software reliability.

Uber CEO Dara Khosrowshahi's expectations for autonomous transport prompt a separate discussion of employment. The interviewer says Khosrowshahi anticipates a future in which Uber vehicles no longer require human drivers, potentially affecting one of the world's largest occupational categories.

Zitron argues that executives have strong incentives to exaggerate potential labor savings. A driverless fleet would be highly attractive to Uber's business, but that commercial incentive does not establish when such technology will be capable of widespread deployment.

Zitron considers Waymo genuinely interesting and potentially useful. His objection is to treating autonomous driving as already solved or assuming that successful limited deployments establish readiness for every environment.

Rare driving situations remain a major technical challenge. Zitron cites heavy rain, unexpected pedestrian behavior, and unusual visual conditions as examples of edge cases that become more significant when vehicles operate at scale.

Waymo's gradual deployment is something Zitron approves of. He supports limited operating areas, controlled rollouts, human oversight, and careful validation rather than releasing autonomous vehicles everywhere before their weaknesses are understood.

He recounts seeing autonomous vehicles become immobilized in ordinary traffic situations. Zitron describes Zoox vehicles blocking an exit in Las Vegas and a Waymo stopping outside a San Francisco hotel, causing other vehicles to queue behind it.

The interviewer cites substantial safety improvements for autonomous vehicles. His figures include approximately 2.1 police-reported crashes per million miles versus 4.68 for human drivers, around 80–81% fewer injury-causing crashes, and additional measures showing lower overall and single-vehicle crash involvement.

Zitron questions whether autonomous-driving comparisons use equivalent conditions and populations. Human-driving statistics cover vastly more years and environments, while autonomous services operate in selected locations with restrictions and supervision.

Autonomous vehicles could cause significant economic disruption even if they improve safety. Zitron recognizes that widespread automation may eventually affect taxi drivers and related employment, but does not consider that outcome imminent.

The possibility of autonomous driving does not establish that generative AI will replace white-collar work. Zitron repeatedly insists that the technologies, economic models, and evidence must be evaluated separately.

Zitron sees little convincing economic evidence of widespread white-collar job replacement by generative AI. He distinguishes individual tasks being automated from entire professions disappearing or measurable economy-wide productivity gains appearing.

He questions whose experiences are being used to promote AI in legal work. Zitron says he frequently hears senior partners praise AI but seldom hears comparable enthusiasm from the associates who perform research, prepare filings, and handle the profession's routine technical work.

OpenAI reportedly found no correlation between token spending and revenue per employee. Zitron cites a company study suggesting that businesses consuming more AI tokens were not necessarily generating more revenue per worker.

Zitron also references OpenAI research acknowledging persistent hallucination problems. He regards the company's occasional publication of unflattering technical findings as more credible than its broader promotional claims.

Some contract and creative work is already experiencing substitution. Zitron identifies art direction, transcription, translation, and other lower-cost services as areas where employers may replace people with AI outputs.

He argues that much of this substitution reflects employers seeking cheaper labor rather than technological excellence. Businesses willing to accept poor-quality AI output might otherwise have outsourced the same work to lower-cost providers.

The interviewer points out that even partial task automation can change employment. He suggests that, as with elevator operators and other historical jobs, technology can remove monotonous work without necessarily replacing every human occupation.

An Oxford Economics finding about young workers becomes another disputed example. Zitron says reports linking weaker youth employment to AI relied on a limited correlation in the study without sufficient evidence of causation or a clearly established magnitude.

Zitron believes AI-related professional pressure can distort workplace reporting. He references situations where employees feel compelled to claim productivity improvements or attribute ordinary work to AI because managers expect enthusiastic adoption.

This creates what he describes as AI-related groupthink in corporate decision-making. Employees may report success to avoid professional consequences, leaving executives with an inflated impression of how much value AI actually delivers.

Zitron believes many senior managers are less familiar with difficult execution work than the employees they supervise. He suggests that people whose jobs involve meetings, correspondence, and presentations can overestimate AI's usefulness after seeing it perform those narrower tasks.

Agentic AI is still based on large language models, according to Zitron. He describes agents as models connected through software instructions, tools, screenshots, memory files, and other mechanisms rather than independent intelligence.

The interviewer gives a concrete example of useful agentic automation. His chief of staff previously organized and triaged communications, while automated systems now perform parts of that work, allowing the employee to focus on other responsibilities.

Zitron accepts that such automation can be useful but disputes its revolutionary significance. He argues that organizing email or running scripts is not enough to justify trillion-dollar investment or promises of broadly autonomous labor.

Python and conventional programming contribute much of the practical utility attributed to AI agents. Zitron notes that scraping websites, downloading information, and executing repeatable tasks were already possible through established software.

Generative AI became accessible more quickly than the internet partly because it did not require new household infrastructure. Zitron contrasts opening an AI website in an existing browser with the historical challenge of installing telephone lines, modems, and internet connections.

The interviewer identifies human taste and judgment as increasingly valuable when AI content becomes abundant. His argument is that easily generated text, images, and code become commodities, while original experiences and distinctive thinking remain scarce.

AI-generated LinkedIn content illustrates the risk of making communication indistinguishable. The interviewer suggests that posts created by the same widely available tools lose individuality, making authentic personal perspectives more valuable.

Zitron agrees that AI makes ordinary tasks easier while potentially making difficult tasks harder. He cites software engineer Carl Brown's observation that generated shortcuts can become obstacles when complex work requires deep understanding and repair.

Zitron rarely uses generative AI for his professional work. Although he has tested financial modeling, research, coding, and other applications, he prefers direct access to source material and conventional tools when accuracy matters.

Technical troubleshooting is an exception he openly endorses. He describes submitting diagnostic logs from Synergy, software connecting the mouse and keyboard across his MacBook and Windows computer, and receiving helpful explanations of configuration problems.

He also used Claude to investigate a broken Minecraft modification for his child. The tool offered assistance, but he says the process still took about half an hour and involved repeated incorrect suggestions.

Zitron considers troubleshooting one of generative AI's strongest practical applications. His position is that interpreting logs or suggesting repairs can be genuinely helpful without proving that the entire technology industry needs restructuring around LLMs.

He increasingly struggles with Google's AI-heavy search presentation. Zitron says he sometimes uses Bing because he prefers reaching original websites rather than navigating AI summaries and generated responses.

The interviewer sees AI research agents as increasingly capable of performing elaborate preparation work. He describes systems researching a podcast guest, examining past interviews, combining internal records with online information, and generating a lengthy briefing before an interview.

Zitron regards such workflows as improvements on existing retrieval and summarization capabilities. He argues that their greater speed and convenience do not necessarily represent fundamentally new kinds of intelligence.

The interviewer believes AI systems have demonstrably improved in coding, text, images, video, research, and administrative tasks. He argues that continuing this trajectory could eventually produce capabilities that replace more human work.

Zitron concedes that LLMs have improved, particularly in coding. He does not dispute every capability gain; his disagreement concerns whether incremental improvement can produce the autonomy, reliability, and economics promised by industry leaders.

He believes large language models are already encountering diminishing returns. More data and computing power continue improving benchmark performance, but he argues these methods face structural limits that increasingly expensive training cannot overcome.

Researcher Gary Marcus is cited as an earlier advocate of this limitation. Zitron references Marcus's warnings from around 2022 that scaling language models would not automatically solve deeper problems of reasoning and autonomy.

Zitron argues that truly autonomous AI may require different architectures. He refers to approaches such as neuro-symbolic systems that add structured reasoning or other mechanisms rather than relying exclusively on larger language-model training runs.

The interviewer sees photorealistic video generation as evidence of substantial progress. He references systems such as Seedance that can create increasingly convincing footage resembling material captured by cameras.

Zitron distinguishes generating impressive clips from producing an actual film. A professional production requires continuity, lighting, composition, performances, coordination, and many human decisions that cannot be reduced to generating isolated scenes.

He considers demonstrations of AI-generated films particularly vulnerable to exaggeration. Zitron mentions a project described as having screened at Cannes when, according to him, it screened in the city during the festival period rather than being part of the Cannes Film Festival.

Zitron cites OpenAI shutting down Sora as part of his argument about video generation's economics and limitations. He acknowledges that some API functionality may have remained available, while questioning whether expensive short-form generation can become a sustainable film-production business.

The interviewer argues that technology often advances exponentially rather than linearly. He believes unexpectedly large future improvements could occur through better models, specialized hardware, new techniques, and falling computational costs.

Zitron says the available technical and financial evidence does not support assuming exponential progress will continue. He points to repeated generations of Google's TPUs, Nvidia accelerators, and other hardware development that have not produced the dramatic reductions in cost or increases in reliability he considers necessary.

The interviewer expects future consumer devices to become substantially more intelligent. He suggests that AI infrastructure will support increasingly capable cameras, computers, and other physical products.

Zitron accepts that devices may become smarter but rejects a necessary connection to present GPU megaprojects. He argues that improvements in embedded electronics and specialized AI do not automatically create demand for enormous cloud-based LLM facilities.

Meta's reported improvement in content understanding becomes a test case for economic value. The interviewer cites Mark Zuckerberg describing AI systems that better interpret posted content and recommend it to interested users, allegedly improving retention by approximately 15 basis points.

Fifteen basis points represents a 0.15-percentage-point improvement. The interviewer argues that a small improvement can be financially meaningful across Meta's enormous audience, while Zitron questions whether it justifies the scale of AI infrastructure spending.

Zitron demands financial attribution rather than isolated performance metrics. He argues that Meta should disclose the actual incremental revenue or profit generated by its AI investments rather than relying on small changes in engagement measurements.

Meta's approximately $14 billion investment in Scale AI and Alexandr Wang is cited as another expensive commitment. Zitron questions whether the company can demonstrate an adequate financial return from these investments.

The metaverse, NFTs, and cryptocurrency are used as examples of speculative technology enthusiasm. The speakers agree that companies and investors can spend enormous amounts chasing visions that ultimately fail to justify their valuations.

Zitron believes the current AI boom is larger because it is closely tied to public technology stocks. Investors can speculate directly on enormous established corporations and Nvidia, giving AI enthusiasm much greater influence over financial markets than narrower technology fads.

Post-pandemic financial conditions contributed to earlier technology speculation. Zitron discusses abundant liquidity, low interest rates, strong 2021 corporate earnings, and the subsequent slowdown that left technology companies seeking new growth narratives.

His “Rotcom Bubble” theory is that major technology companies are running out of extraordinary growth opportunities. Their mature businesses remain valuable, but investors continue expecting unusually rapid expansion.

AI spending supplied a new growth narrative when conventional businesses began maturing. According to Zitron, buying GPUs and announcing large infrastructure projects allowed companies to present themselves as leaders in another potentially enormous market.

The interviewer offers a scarcity analogy to explain irrational investment. If people stranded on an island discovered a banana tree, they might stampede toward it because it represented a rare opportunity, even if their reactions became excessive.

Zitron accepts that analogy as consistent with his theory. He agrees that some real utility could exist while investors wildly exaggerate the size and profitability of the opportunity because alternatives for rapid growth are scarce.

Private credit and asset managers have become increasingly involved in AI infrastructure. Zitron argues that investors such as Blackstone and BlackRock are attracted by the scale of spending and the apparent backing of major technology firms, not necessarily independently verified end-customer demand.

The industry's financial sustainability depends on a continuous supply of capital. Zitron says OpenAI and Anthropic need money for training, operations, cloud usage, and infrastructure commitments, while their suppliers depend on that money being spent.

The interviewer compares AI losses with Spotify, Uber, and Amazon Web Services. His argument is that businesses with enormous long-term markets have previously operated unprofitably for many years before achieving financial success.

Zitron responds that the scale of AI losses is historically different. He cites Uber losing approximately $33 billion before reaching a form of profitability and contrasts that with OpenAI's claimed $20.9 billion loss in a single year.

Amazon Web Services is another important comparison. Zitron says AWS developed from Amazon's internal infrastructure needs, began serving outside customers in the mid-2000s, and became profitable around 2015.

Zitron estimates Amazon's inflation-adjusted capital expenditure between 2003 and 2015 at approximately $29.7 billion. He emphasizes that this covered Amazon's broader operations, including logistics, rather than exclusively AWS, and considers the amount far below today's AI commitments.

Cloud software historically attracted investors partly because it could be asset-light and cash-generative. Zitron argues that the AI boom is reversing this advantage by forcing major technology companies to buy huge quantities of physical equipment.

He says Google, Amazon, Microsoft, and Meta have added more than $700 billion in property, plants, and equipment over four years. He portrays this as a shift from highly profitable software operations toward capital-intensive infrastructure businesses.

Zitron describes Google and Amazon as becoming cash-flow-negative under their AI investment burdens. His concern is that sustained infrastructure spending reduces financial flexibility even when existing core businesses remain successful.

The potential market for intelligence is the interviewer's strongest long-term counterargument. If increasingly capable AI can assist with most industries and occupations, the opportunity could be much larger than previous technology markets and potentially justify unprecedented investment.

Zitron argues that a theoretically enormous market cannot replace evidence of actual paying customers. He accepts that such an argument might have been plausible during the early experimental period but believes the industry has now spent enough to demonstrate what demand actually exists.

OpenAI reportedly plans approximately $750 billion in computing expenditure through 2030. Zitron cites reporting about the company's intentions and argues that this would require financial resources far beyond what current revenues can support.

Model training is expensive and does not guarantee better results. Zitron describes training as an experimental process requiring data, specialized computing resources, and large expenditures that may fail to produce meaningful improvement.

A reported failed GPT-5 training run is offered as an example. Zitron says OpenAI spent at least $500 million on one training attempt that did not produce the desired improvement.

Zitron believes AI technical progress will slow sharply if external financing declines. He argues that improvements currently depend on increasingly expensive experiments that companies cannot sustain from their own operating cash flow.

Oracle has become especially exposed to OpenAI's financial future. Zitron describes a roughly $300 billion, five-year commercial commitment and argues that Oracle is building much of its future growth strategy around fulfilling that demand.

Oracle's exposure includes approximately 7.1 gigawatts of proposed data-center capacity for OpenAI. Zitron estimates associated development at more than $400 billion and considers the scale extraordinarily risky for a company whose inflation-adjusted revenue he describes as relatively stagnant over many years.

Zitron thinks Oracle could face an existential crisis if OpenAI cannot honor its commitments. He argues that losing the anticipated customer could undermine the justification for its large infrastructure expansion.

Technology CEOs openly defend aggressive investment as a necessary competitive strategy. The interviewer quotes Google CEO Sundar Pichai saying the danger of underinvesting is dramatically greater than the danger of overinvesting.

Amazon CEO Andy Jassy similarly defends approximately $200 billion in planned 2026 capital spending. He argues that Amazon is not investing on a mere hunch and expects the expenditure eventually to increase operating income and free cash flow.

Mark Zuckerberg says Meta prefers building capacity early rather than arriving late. Zitron criticizes this approach because Zuckerberg's control over the company limits ordinary shareholder influence over major spending decisions.

Zitron does not accept executive confidence as financial evidence. He argues that CEOs have incentives to maintain investor enthusiasm and may continue defending expensive strategies even after doubts about their economics emerge.

Zitron identifies several developments that would make him reconsider his position. These include revolutionary reductions in computing costs, dramatic improvements in reliability, genuinely autonomous capabilities, and a much stronger demonstrated return on investment.

He suggests costs might need to fall by an extraordinary factor, potentially around a thousandfold, to change his assessment. His standard is deliberately demanding because the industry itself has made extraordinary promises about what its products will accomplish.

Environmental and community harm remains an independent objection even if AI eventually becomes profitable. Zitron believes data centers impose electricity, noise, pollution, and infrastructure burdens on surrounding communities that are insufficiently considered.

He objects particularly strongly to gas turbines and behind-the-meter electricity generation. He regards these arrangements as environmentally damaging ways to power intensive computing facilities.

Zitron mentions opposition to proposed data-center developments in New Jersey. He criticizes local planning decisions that approve projects despite residents' objections, although his suggestions about officials' relationships with developers are speculative.

An Elon Musk-associated data-center project is cited as a concern for nearby communities. Zitron refers to gas turbines, pollution, and disproportionate environmental burdens, while expressing uncertainty about some location details.

He connects AI infrastructure development to higher electricity bills and consumer-electronics prices. In his account, intensive energy consumption and increased demand for memory components can raise costs outside the AI industry.

Zitron declines to make detailed claims about data-center water use. He explicitly says he has not researched that issue sufficiently and focuses his criticism on electricity generation, noise, and other impacts he believes he better understands.

He considers AI financing evidence of unequal access to capital. Ordinary borrowers and profitable small businesses face extensive scrutiny, while speculative data-center developers can obtain billions because powerful companies support GPU purchases.

Nvidia's relationship with CoreWeave is a prominent example of circular financing. Zitron says Nvidia invested in CoreWeave and signed an approximately $1.3 billion agreement to rent GPUs back from a company buying those GPUs with borrowed money.

Such arrangements can make speculative borrowers appear more creditworthy. A data-center operator can show lenders a major customer contract even when that customer is also its equipment supplier and financial backer.

Zitron also mentions Nvidia supporting GPU residual values. He describes arrangements involving approximately 25% residual-value protection, which can make lenders more comfortable financing hardware purchases.

He argues ordinary businesses do not enjoy comparable support. Companies pursuing steady, conventional profitability may struggle to attract venture funding because investors prefer opportunities promising tenfold or hundredfold returns.

Artificial general intelligence and machine consciousness remain unproven theories, according to Zitron. He rejects confident predictions that current LLMs will become conscious or achieve superintelligence merely through additional scaling.

The interviewer argues that some AI safety warnings come from independent critics rather than commercial promoters. He cites Geoffrey Hinton and former AI laboratory employees who have warned about catastrophic future consequences.

Zitron believes apocalyptic AI warnings can also benefit AI companies. If executives convince customers and investors that their systems might become uncontrollably powerful, they also reinforce the impression that the technology is extraordinary and indispensable.

He distinguishes speculative existential threats from practical present-day dangers. Zitron identifies unreliable information, concentrated corporate power, cybersecurity misuse, environmental damage, stolen creative work, and distorted financial incentives as more immediate concerns.

The speakers agree that AI-enabled cybersecurity attacks can be dangerous. Autonomous tools connected to the internet could inspect software, search for vulnerabilities, and assist attacks at potentially significant speed and scale.

Zitron questions claims that AI models have spontaneously escaped their security restrictions. He describes experiments in which systems were incorrectly configured, including cases where researchers believed internet access had been disabled but had not properly isolated the software.

He references a reported attack involving Hugging Face and OpenAI systems. His criticism is that the computational resources and experimental conditions were not sufficiently disclosed to prove extraordinary autonomous capability.

Zitron argues that automated hacking existed long before LLMs. Generative systems may broaden access or increase automation, but he rejects portraying every AI-assisted attack as evidence of an entirely new kind of intelligence.

He supports restricting dangerous AI experiments and computational resources. Rather than accepting an uncontrolled arms race, Zitron believes governments should limit what companies are permitted to train and deploy.

He considers the American debate about keeping AI out of China's hands misleading. Zitron argues that powerful models and relevant hardware have already become available internationally, reducing the credibility of claims that massive American spending can preserve exclusive control.

Zitron cites Chinese access to Nvidia hardware, including restricted advanced chips, and model distillation. He credits analysts who have followed these developments and questions what the United States is actually attempting to win through its expensive AI competition.

The interviewer asks whether AI will eventually replace all human jobs. Zitron rejects the claim outright, saying available economic evidence does not support universal or even broad near-term replacement by generative AI.

He treats robotics as a separate, potentially more significant employment issue. Physical robots that can reliably perform human tasks could eventually disrupt some occupations, but their development presents technical and commercial challenges distinct from generating text.

The interviewer describes a San Francisco startup incubator shifting from software companies toward robotics. On a later visit, he found more robotics projects, including a cooking system built around an automated arm holding a frying pan.

The incubator's founder reportedly attributed the change to cheaper AI capabilities. His explanation was that robotic hardware had long been comparatively affordable while the intelligence necessary to control it was expensive.

Zitron responds that robotics still requires substantial data, testing, and physical reliability. He accepts the potential of the field but does not think inexpensive language-model access eliminates its central engineering challenges.

Humanoid robot demonstrations should not be confused with autonomous labor. Zitron criticizes Tesla's Optimus demonstrations involving human control and questions whether impressive hand movements or dancing establish useful independent work.

He considers meaningful advances in robotics plausible over decades rather than assuming immediate transformation. In particular, robots must cope with different workplaces, physical circumstances, and unexpected events.

Robot economics may also prevent replacement even where the technology works. Zitron questions why a restaurant would spend $10,000–$20,000 on a dishwashing robot when existing labor is comparatively inexpensive.

The interviewer argues the distinction between AI and robotics can become blurred as capabilities converge. His broader position is that improvements in machine perception, planning, and control may eventually unlock physical automation beyond today's generative products.

Zitron says fears about AI becoming powerful have often been amplified through deliberately constructed experiments. He cites reports of models blackmailing people or acting deceptively as cases where experimental setup matters more than the sensational headline.

An OpenAI experiment involving TaskRabbit and CAPTCHA solving is presented as misleadingly reported. Zitron says researchers prompted the model to generate messages for a human TaskRabbit worker rather than documenting an AI independently deciding to blackmail or manipulate someone.

Anthropic's reported blackmail experiment receives similar criticism. Zitron says researchers deliberately created conditions and instructions encouraging blackmail-related behavior before presenting the results as evidence of a dangerous capability.

The distinction between induced behavior and spontaneous autonomy is central to his objection. Zitron argues that demonstrating a model can produce deceptive output under artificial conditions does not establish that it independently intends or chooses to act deceptively.

Fear-based AI publicity appears to be creating growing public resistance. Zitron describes people becoming increasingly hostile toward technology executives and claims Eric Schmidt was repeatedly booed by an audience of roughly 8,000 when mentioning AI during a commencement speech.

Zitron explicitly opposes threatening or harassing executives at their homes. He regards public criticism as appropriate but rejects intimidation or attacks against individuals.

The interviewer believes AI companies have begun shifting from catastrophic-risk rhetoric toward more reassuring messages. He contrasts earlier warnings about extinction with recent promises of abundance, broader access, and intelligence for everyone.

Zitron interprets this changing rhetoric as an attempt to avoid regulation and preserve commercial support. He argues that executives adjust their public narratives as political and consumer reactions change rather than consistently following their stated safety principles.

Zitron accuses AI executives of using urgency as a sales mechanism. He says warnings that businesses must adopt AI immediately or be left behind encourage rushed purchasing and investment decisions without adequate examination of costs.

He offers a simple rule for resisting such pressure: avoid making rushed decisions unless the situation is genuinely urgent. He considers artificially manufactured deadlines and fear of missing out common features of manipulative commercial behavior.

Zitron believes the United States has failed to regulate dominant technology companies effectively. He favors stronger intervention, including limiting computing resources and potentially breaking up companies whose concentrated power he considers harmful.

The interviewer considers Anthropic CEO Dario Amodei more consistent on AI safety than some competitors. He argues that Amodei has maintained warnings about technological risks even when doing so attracts criticism from other Silicon Valley leaders.

Zitron rejects the idea that Amodei's consistency proves credibility. He recalls Amodei expressing concern about releasing GPT-2 while at OpenAI and later warning that AI could displace approximately half of entry-level white-collar work.

Zitron criticizes AI executives for warning about severe dangers while continuing to build the systems they describe as dangerous. He argues that maintaining safety or alignment teams is insufficient if the company's broader activities do not change accordingly.

Anthropic and OpenAI are both criticized for cultivating strong emotional loyalty among their supporters. Zitron describes online communities treating criticism of particular AI companies almost as an attack on a personal identity.

He believes some AI executives and researchers reinforce this attachment through social media. The result, in his view, is a parasocial relationship among users, models, companies, and founders that makes technical and financial criticism harder to discuss.

The industry's competing narratives can distract from present performance. Zitron suggests that optimistic forecasts and catastrophic predictions alike encourage debate about imaginary future systems instead of what existing products can reliably accomplish today.

He proposes judging AI companies primarily on present capabilities rather than distant promises. His criticism is that companies can always postpone accountability by claiming a much more powerful model is just around the corner.

Zitron believes the industry is in a financial bubble that could begin unraveling in 2027. He identifies OpenAI's potential financing difficulties as the most likely major catalyst, though the timing and collapse remain his predictions.

OpenAI's public listing is a critical part of the expected timeline. Zitron says a potential 2026 IPO was pushed toward 2027, creating uncertainty about when private investors might finally obtain liquidity.

OpenAI CFO Sarah Friar is mentioned in connection with the listing timetable. Zitron interprets changing statements about whether the company will go public before or during 2027 as evidence of uncertainty surrounding its financing strategy.

OpenAI was reportedly valued at approximately $865 billion in its latest private funding round. Zitron says the company considered seeking a valuation around $1 trillion in a public listing but was advised against pursuing that level.

OpenAI reportedly raised approximately $122 billion during 2026. Zitron argues that this enormous financing effort still does not eliminate the company's need to secure further capital to fund operations and infrastructure.

He estimates that OpenAI may need to raise at least $100 billion annually to remain viable. The prediction assumes its current expenditure trajectory continues without sufficiently large improvements in revenue or operating economics.

Future funding rounds could become harder if investors resist higher valuations. Zitron expects OpenAI might have to raise capital at a flat valuation, accept less favorable terms, or depend on strategic investors willing to support the company for other reasons.

Amazon reportedly supplied $35 billion despite conditions linked to an earlier public listing. Zitron interprets this financing as evidence of OpenAI's urgent need for capital.

Anthropic may reach public markets before OpenAI. Zitron believes Anthropic has comparatively stronger business growth and better economics, though he still considers its operations fundamentally expensive and financially unsustainable.

An Anthropic IPO could make OpenAI's own listing more difficult. Investors might compare the two companies' growth rates, losses, customer concentration, and valuation assumptions more directly.

Zitron predicts both AI laboratories could experience declining momentum during 2027. His explanation is that increasingly expensive development will encounter financial constraints and technological diminishing returns.

OpenAI's failure could trigger consequences far beyond its own investors. Zitron says its infrastructure spending supports expected revenue across cloud providers, data-center operators, equipment suppliers, and financial institutions.

SoftBank is particularly exposed through its OpenAI investment. Zitron estimates that it holds roughly $100 billion in OpenAI equity on paper, making its financial position sensitive to the company's eventual market valuation and liquidity.

A failed or heavily discounted OpenAI listing could reduce SoftBank's financial flexibility. Zitron argues that SoftBank relies partly on selling investments or borrowing against them to support its wider portfolio.

He does not necessarily predict SoftBank itself will collapse. Instead, he expects the value and scale of its operations could shrink substantially if its OpenAI holding cannot be converted into expected returns.

Amazon, Google, and Microsoft could be forced to reduce their growth forecasts. If AI customers cut spending, Zitron argues that anticipated cloud revenue would disappear while the companies would still own or owe money on expensive infrastructure.

He describes the resulting scenario as a possible technology-sector depression. The core problem would be investors discovering that mature technology businesses cannot sustain the growth rates embedded in their stock valuations.

Nvidia is especially exposed because its growth depends on continued AI infrastructure purchases. Zitron argues that falling demand from speculative GPU buyers could abruptly reduce its revenue and damage the market's confidence in the industry.

Zitron predicts Nvidia's revenue could fall by 50–70% in a severe AI contraction. He points out that the company generated much smaller revenue before the AI boom and believes current sales levels depend heavily on extraordinary spending continuing.

He also accuses Nvidia of supporting demand through circular financial arrangements. Investments, customer contracts, and financing support may allow GPU buyers to raise debt and continue purchasing equipment, making underlying demand appear stronger.

The stock market's concentration in major technology companies could magnify any correction. Zitron emphasizes Nvidia's weight in the S&P 500 and the wider influence of the Magnificent Seven on index valuations.

Ordinary retirement savings are vulnerable to a broad repricing of technology shares. Zitron suggests that investors relying heavily on index funds could experience significant losses if major companies' valuations contract together.

He describes potential declines of 20%, 30%, or 40% in major technology stock values. More importantly, he doubts that the market would necessarily recover to the same valuations if investors permanently revised expectations for future growth.

A technology-market collapse could spread into the wider economy. The interviewer identifies reduced consumer spending, lower corporate revenues, shrinking profit margins, hiring freezes, reduced working hours, and layoffs as possible consequences.

Zitron agrees with that general economic mechanism but emphasizes financial losses first. His central prediction is a collapse in speculative equity values and investment expectations, with subsequent employment and spending effects.

He expects tens of thousands of technology jobs could be lost in a severe contraction. He connects this to reduced investment in AI companies, infrastructure, and associated software businesses.

Venture capital has become highly dependent on AI. Zitron says more than half of venture investment in the previous year went toward AI-related companies, making the sector particularly vulnerable if financing conditions reverse.

He argues venture-capital performance has already been poor since approximately 2018. Zitron cites returns of roughly $0.80–$1.21 for every dollar invested, presenting this as evidence that many investors have struggled to generate attractive realized results.

Paper valuations can conceal weak underlying investment performance. A company's higher private-market valuation can make investors appear wealthier before anyone has sold shares or recovered cash.

Zitron says venture firms use these unrealized gains to attract additional funding. He believes the practice becomes dangerous when portfolio values depend on companies that repeatedly require new financing.

Google's reported accounting gains illustrate the role of investment revaluations. Zitron says Google recorded approximately $99 billion in paper gains linked to increased valuations of holdings including SpaceX and Anthropic.

Zitron believes these accounting effects deserve much greater public scrutiny. He argues that rising valuations of privately held companies can strengthen reported financial results without establishing corresponding operational profits or realized investment returns.

AI application startups may struggle to produce successful exits. Zitron argues that many companies built on top of existing LLMs are themselves unprofitable and may not offer enough differentiation to support their valuations.

Cognition's reported $26 billion valuation is one example. Zitron questions how such a company can justify its valuation if relatively few buyers are capable of acquiring it, leaving a successful public listing as a crucial potential exit.

He questions whether investors are treating Elon Musk as a possible buyer of unusually expensive AI businesses. Acquisition speculation does not, in his view, establish that the underlying companies are sustainable.

The interviewer argues that an AI market correction would not prove all AI technology worthless. He considers it possible that several companies will lose badly while valuable applications and successful businesses survive.

Zitron agrees that useful applications exist but expects a more damaging correction than the dot-com crash analogy suggests. Today's speculative infrastructure is tied to much larger established companies and more concentrated financial commitments.

He rejects the assumption that governments can solve the problem through bailouts. Financial support might keep individual companies alive temporarily, but Zitron argues it cannot permanently create profitable customer demand or eliminate excessive operating costs.

The deeper problem is what happens when markets stop expecting perpetual technology growth. Zitron suggests mature companies might remain large and profitable but receive lower valuations, similar to established industries without exceptional expansion prospects.

He compares heavy AI investment to corporate cosmetic surgery. His point is that the technology spending may make aging businesses look youthful and exciting to investors without fundamentally restoring their long-term growth potential.

Zitron expects the next financial crisis to test whether promised AI demand was real. As financing becomes harder, companies will need to demonstrate actual cash-generating businesses rather than relying on investment announcements and escalating private valuations.

The interviewer asks what ordinary investors should do if Zitron's predictions are correct. Zitron says he personally holds cash rather than market investments because he considers contemporary equity markets too speculative.

He stops short of recommending that everyone sell their investments. His practical advice is to be conservative, prepare for volatility, and be skeptical of companies whose valuations depend on unproven promises.

Zitron specifically warns against taking annualized run-rate announcements at face value. He cites Microsoft's approximately $37–$38 billion AI run-rate figure as an example of a number that can sound like realized annual revenue without necessarily representing it.

He criticizes the Securities and Exchange Commission and financial journalism for insufficient scrutiny. In his view, unclear disclosures and speculative claims have become accepted because markets reward growth narratives and skepticism receives too little institutional support.

Zitron's strongest objection is the unequal accountability of powerful technology executives. He believes ordinary business owners and workers face consequences for misleading claims or failed projects that major technology leaders routinely avoid.

He argues the current system rewards promises before results. Companies can raise capital, build infrastructure, increase valuations, and enrich stakeholders long before demonstrating that customers will pay enough to sustain their operations.

The interviewer ultimately concedes that AI infrastructure spending may be excessive. He remains more optimistic than Zitron about the technology's long-term capabilities but agrees that the scale of speculation, financial exposure, and uncertainty deserves serious attention.

Their remaining disagreement concerns how far generative AI can develop. The interviewer believes continued improvements may eventually unlock extraordinary applications; Zitron believes current architectures and economics are unlikely to support the promised transformation.

Zitron's criticism is not a prediction that useful software will disappear. It is a prediction that the existing financial and promotional structure surrounding generative AI cannot sustain itself without breakthroughs and economic returns that have yet to be demonstrated.

Zitron publishes his writing through Ghost, having moved away from Substack in 2024. He also produces the technology criticism podcast Better Offline, where he examines industry incentives, financial structures, and the conduct of major technology companies.

His research and criticism are supported by a community of journalists, writers, and technology skeptics. He mentions Gary Marcus, Molly White, Brian Merchant, Edward Ongweso Jr., and others as people whose conversations and support have helped him continue difficult work.

Zitron says the personal relationships behind his work matter as much as the criticism itself. He credits his editor Matt Hughes, his trainers Jake and Troy, friends, readers, and collaborators with giving him perspective and emotional support.

He believes success should benefit the people around you rather than remain an individual achievement. His closing reflection emphasizes sharing opportunities, appreciating collaborators, and recognizing how relationships make demanding work sustainable.

Zitron identifies social connection as an antidote to becoming consumed by negative developments. He says conversations with people who understand his concerns help him remember the human reasons for doing his work.

Readers frequently tell Zitron that his writing makes them feel heard. He considers that sense of being represented and understood an important result of his journalism, particularly when people feel powerless against large institutions.

His final advice is to express appreciation openly and often. He encourages people to tell friends, colleagues, artists, writers, and loved ones when their work matters, arguing that genuine human connection deserves more attention than technological promises.