The $1 Trillion AI Question: Ed Zitron Explains Why OpenAI, Nvidia, Microsoft, and the Industry Could Be Heading for a Financial Crisis

Technology critic Ed Zitron challenges the economics of the generative AI boom, alleging that circular financing, subsidized products, unreliable technology, and extraordinary infrastructure commitments have created a market whose promises may be impossible to fulfill. Steven Bartlett presents the opposing case: AI is already useful, improving rapidly, and could become one of history's most important technologies.

Microsoft recorded $24.1 billion in revenue from commercial arrangements with OpenAI during its 2026 financial year. Nvidia generated $215.9 billion in annual sales. OpenAI announced a $122 billion funding round. Meanwhile, plans for new data centers have reached a scale measured in hundreds of gigawatts.

These figures describe an industry expanding at extraordinary speed. But they also raise a question that receives much less attention than announcements about artificial intelligence becoming smarter: Who will ultimately pay for all this computing power, and will they pay enough to make the investment profitable?

That question sits at the center of technology critic Ed Zitron's August 27, 2026 appearance on Steven Bartlett's The Diary of a CEO.

Zitron argues that generative AI has become an enormous financial experiment in which major technology companies invest in AI laboratories, sell computing services to those laboratories, and use the resulting activity to support expectations of future growth. He believes the industry's costs are far greater than its demonstrated economic value.

Bartlett challenges that conclusion throughout the conversation. He points to rapid consumer adoption, improvements in coding and research, small businesses reducing expenses, and the history of technologies that initially appeared too expensive or unreliable to succeed.

Both acknowledge that generative AI can be useful. Their disagreement is whether that usefulness can eventually justify the money, infrastructure, and corporate commitments already being made.

The financial records reveal why the disagreement matters. Some of Zitron's central figures are supported by public disclosures. Others are estimates, contested interpretations, or predictions. Distinguishing between them is essential to understanding the actual risks.

1. Microsoft's $24.1 Billion OpenAI Revenue Reveals a Financial Dependence

What Microsoft's financial statements establish

One of Zitron's most important claims concerns Microsoft's dependence on OpenAI for its artificial intelligence business.

During the interview, he cites approximately $34.33 billion in Microsoft's fiscal 2026 AI revenue, including $24.1 billion associated with OpenAI. On that comparison, OpenAI accounts for approximately 70% of the total.

Microsoft's own 2026 annual report confirms the $24.1 billion figure.

For the financial year ending June 30, 2026, Microsoft disclosed revenue from commercial arrangements with OpenAI, including revenue-sharing payments, totaling $24.1 billion. It also reported $6 billion in OpenAI-related accounts receivable.

That distinction matters. The $24.1 billion is not necessarily revenue exclusively from renting AI computing infrastructure. Microsoft's disclosure includes revenue-sharing arrangements, and the company does not provide a simple public breakdown establishing exactly how much each type of service contributed.

The broader AI revenue figure used in the interview is an estimate rather than a separately reported annual total in Microsoft's financial statements.

Nevertheless, the relationship is substantial.

Microsoft disclosed an approximately 25% economic interest in OpenAI on an as-converted basis and reported $13 billion in total funding commitments, of which $11.9 billion had been funded by June 30, 2026.

Microsoft is therefore simultaneously an investor, commercial partner, infrastructure provider, and recipient of revenue from OpenAI.

For Zitron, this is the central problem. Revenue from a fast-growing AI customer looks attractive, but its durability depends partly on whether that customer can continue raising enough money to operate.

Is AI spending creating its own revenue?

Zitron describes a system in which capital moves among a relatively small group of companies.

Investors provide money to AI laboratories. Those laboratories purchase computing services, data-center capacity, and equipment. The purchases produce revenue for technology suppliers, encouraging additional infrastructure investment and supporting expectations of growth.

The connections are not hypothetical.

On February 27, 2026, OpenAI announced $110 billion in new investment commitments, including $50 billion from Amazon, $30 billion from Nvidia, and $30 billion from SoftBank.

Amazon also deepened its relationship with Anthropic. On April 20, 2026, Anthropic announced a commitment to spend more than $100 billion on Amazon Web Services technologies over ten years. Amazon simultaneously committed another $5 billion in investment, with as much as $20 billion more dependent on future conditions. The companies described plans for up to five gigawatts of computing capacity.

These transactions support Zitron's observation that several AI suppliers also finance their largest customers.

But they do not establish that the revenue is fictitious. Computing services can be genuinely delivered, customers can use them, and suppliers can recognize legitimate revenue. What makes the arrangements unusual is the financial interdependence.

A company buying cloud capacity might have a valuable business even while losing money. The risk increases when its spending commitments depend on continuous fundraising from investors whose businesses also benefit from those purchases.

Zitron argues that analysts have projected more than $400 billion in future cloud revenue associated with OpenAI and Anthropic over roughly three and a half years. He connects those expectations to cloud growth projections approaching 30%.

Those projections are not guaranteed contractual revenue. Their significance lies in how much future business investors may be expecting from companies that have not yet demonstrated sustainable profitability.

Microsoft's spending is much larger than its disclosed OpenAI revenue

During the interview, Zitron contrasts Microsoft's approximately $115 billion in fiscal 2026 capital investment with plans for approximately $175 billion in the following year.

His reaction is concise:

"The math does not make sense." — Ed Zitron

Capital expenditure, or capex, is spending on long-lived assets such as buildings, servers, and computing equipment. Its accounting differs from the operating expenses that companies recognize immediately.

A new data center might operate for years, so its construction cost cannot be compared directly with one year's revenue as if every dollar must be recovered immediately.

That is an important limitation of Zitron's comparison.

Still, the financing requirements are substantial. Microsoft's annual report shows rapidly increasing investment in property and equipment, alongside major future commitments for construction, leases, and data-center purchases.

The report also demonstrates that Microsoft remains a large, profitable corporation. In fiscal 2026, it generated $331.8 billion in total revenue and $133.7 billion in reported net income.

Zitron's argument is therefore not that Microsoft has already become unprofitable. It is that the growing cost of its AI strategy may eventually reduce the exceptional cash generation and financial flexibility that made its traditional businesses so attractive.

2. Nvidia Is Selling the Chips—and Helping Finance the Buyers

Nvidia has become one of the clearest financial beneficiaries of generative AI.

On February 25, 2026, the company reported $215.9 billion in fiscal-year revenue, an increase of 65% from the preceding year. Its fourth-quarter revenue reached $68.1 billion, including $62.3 billion from data-center operations.

Those are realized sales reported by a profitable public company.

Zitron does not dispute Nvidia's ability to sell equipment. He questions the economics of the companies buying it.

He estimates that AI revenue outside OpenAI and Anthropic is much smaller than the industry's enormous hardware purchases might suggest. His approximately $22 billion estimate, however, is not an audited measure of worldwide AI demand. It depends on definitions of generative AI revenue and which businesses are counted.

That difference between a verified supplier's revenue and an estimated market's revenue matters. Nvidia sells hardware for multiple forms of computing, while cloud operators invest in equipment expected to generate returns over several years.

Nevertheless, the scale of purchasing creates a genuine exposure if customers stop expanding.

The CoreWeave arrangement

CoreWeave, a specialized cloud computing provider, illustrates the financial relationships that concern Zitron.

The company purchases and operates Nvidia-based computing infrastructure. Nvidia supplies CoreWeave with chips, owns shares in the business, and purchases computing capacity from it.

A particularly revealing arrangement appears in CoreWeave's September 2025 SEC filing.

The filing describes an agreement initially valued at $6.3 billion. Under its terms, Nvidia is obligated, subject to specified conditions, to purchase certain residual computing capacity that CoreWeave cannot sell to other customers, through April 2032.

This provides CoreWeave with a significant commercial commitment from its equipment supplier.

It also creates a financial relationship that can be useful when obtaining loans. A lender evaluating CoreWeave can consider contracted revenue from Nvidia even though Nvidia is also a supplier and investor.

The interview refers to a $1.3 billion arrangement. The September 2025 filing establishes a larger, $6.3 billion initial agreement and should not be confused with the smaller figure.

The relationship expanded again in January 2026, when Nvidia announced a further $2 billion equity investment. The companies described plans to develop more than five gigawatts of AI infrastructure by 2030.

Why lenders care about GPU values

GPUs are expensive assets, and their resale value matters when companies borrow against them.

A lender financing equipment wants to know whether that equipment can be sold for enough money if the borrower defaults. In rapidly evolving technology markets, newer chips may reduce the commercial attractiveness of older hardware.

Zitron discusses arrangements offering protection against declining GPU residual values, including support around 25% in certain circumstances.

Nvidia's 2026 quarterly filing acknowledges potential limited residual-value support for specific infrastructure projects. It also describes preliminary financing platforms intended to mobilize more than $500 billion of third-party capital over time.

The filing does not establish that all such financing will occur or that Nvidia has guaranteed every proposed project.

The underlying issue is clear: demand for Nvidia equipment is supported not only by customers' existing revenue, but also by borrowing, equity investment, long-term commitments, and financial arrangements involving Nvidia itself.

A market downturn could therefore affect hardware sales, lenders, cloud customers, and equipment values at the same time.

3. AI Subscriptions Can Hide the Real Cost of Computation

Why a $20 subscription does not explain what AI costs

Consumers usually experience generative AI through subscriptions.

Someone pays $20, $100, or $200 per month, enters questions, receives responses, and generally does not see the cost of the computational work taking place.

Developers using application programming interfaces encounter a different pricing structure.

AI providers commonly charge for tokens, small units of text processed or generated by a model. English text averages approximately three-quarters of a word per token, although the relationship varies.

Input tokens include the user's instructions, documents, and other material. Output tokens include generated responses and, with certain models, additional reasoning computation.

Zitron considers the distinction crucial.

"You pay when you use an LLM regardless of whether you get what you want." — Ed Zitron

The provider incurs computational expenses whether the model answers correctly, misunderstands the request, or generates code that breaks a program.

That differs from purchasing an outcome, such as a corrected software defect or a completed research assignment. The customer may be buying computation without knowing how much useful work it will produce.

The $14,000 subscription comparison

Zitron cites SemiAnalysis estimates suggesting that an intensive user of a $200 ChatGPT plan might consume usage equivalent to approximately $14,000 at comparable API prices.

He gives an approximately $8,000 comparison for an Anthropic subscription and suggests that even a $20 plan can permit hundreds of dollars in API-equivalent usage.

These comparisons require careful interpretation.

API retail prices are not the same as the provider's actual cost of electricity, hardware depreciation, networking, and model operation. A provider may charge substantially more per token than its incremental production cost.

A subscriber consuming $14,000 in API-equivalent tokens has not necessarily caused the company to lose $13,800.

The comparison nevertheless illustrates why flat-rate pricing can obscure substantial differences among users. One customer might ask a few short questions daily while another runs intensive coding and research tasks for hours.

Usage limits help control those differences, but customers may not understand when those limits will apply or how much computation their work actually requires.

Zitron cites a $20.9 billion annual loss for OpenAI and argues that aggressive subsidies help explain the financial pressure. That particular loss figure remains his reported estimate rather than a figure established here through publicly audited OpenAI accounts.

What happened when enterprises faced usage-based billing?

Zitron says OpenAI began shifting larger organizations, including companies with more than approximately 150 employees, toward consumption-based pricing around March 2026.

He also claims Uber exhausted its annual AI token budget within three months.

If accurate, the Uber example reveals a practical problem for companies that make intensive AI tools widely available.

An employee may save time by asking a model to analyze thousands of documents or repeatedly modify software. But if the associated bill becomes substantial, the employer must determine whether the benefit exceeds the expense.

Zitron argues that much enthusiasm assumes the existing subscription price will remain available indefinitely.

Bartlett offers a different interpretation. Subsidizing early users is a common strategy for growing technology businesses. AI companies may expect expanding usage, improved hardware, and more efficient models eventually to lower costs enough to support profitable prices.

Zitron replies that improvements in capability are not necessarily improvements in economics. More advanced models can require greater inference resources, and increasingly demanding users may consume the benefits of efficiency gains.

Neither position can be resolved through subscription prices alone. The decisive figures are the actual cost of serving customers, the prices customers will ultimately accept, and the profits providers can retain after infrastructure and operating expenses.

4. The Data-Center Boom: 190 Gigawatts of Planned Capacity

An infrastructure expansion unlike ordinary software

Traditional software companies can often expand by selling additional subscriptions without building an entirely new physical facility for every large increase in demand.

Generative AI changes that relationship.

Large language models require significant computing resources during training, when they learn from enormous datasets, and during inference, when they respond to users.

Modern AI accelerators use specialized memory and consume substantial electricity. Large installations may require thousands or hundreds of thousands of interconnected chips, advanced cooling, high-speed networking, and reliable power supplies.

Nvidia's CUDA computing platform helped establish the software ecosystem that made GPU-based AI development practical. That advantage developed over many years before ChatGPT became popular.

Zitron argues that the industry's current infrastructure investments cannot be evaluated as if they were ordinary software expenses.

Stargate and the concentration of computing power

One example discussed is OpenAI and Oracle's Stargate development in Abilene, Texas.

Zitron describes a planned 1.2-gigawatt facility consisting of eight buildings, each associated with approximately 50,000 Nvidia GB200 GPUs.

A gigawatt represents one billion watts of electrical power. At that scale, the power requirements of a single computing campus become comparable to those of substantial population centers.

Zitron compares the concentration of electrical demand with the geographic scale of Bristol, England. Although the transcript's precise city-energy comparisons mix units that should not be treated as interchangeable, the broader point is that AI infrastructure concentrates exceptionally large power requirements into relatively small areas.

The cost extends beyond chips. Facilities require land, electrical connections, transformers, cooling equipment, construction, financing, and continuing maintenance.

This is why infrastructure spending can keep increasing even as individual chips become more efficient.

A more capable system may generate more output per unit of energy while encouraging customers to generate far more output overall.

Sightline Climate's 190-gigawatt pipeline

Zitron highlights an estimate of approximately 190 gigawatts of planned data-center capacity.

Independent research supports the existence of that development pipeline, but introduces an important qualification.

On February 24, 2026, Sightline Climate, subsequently presented through Currence, reported tracking 190 gigawatts across 777 large data-center and AI-factory projects announced since 2024.

That does not mean 190 gigawatts are operating or that every project will be built.

Of the approximately 16 gigawatts scheduled to become operational during 2026, the research identified only about five gigawatts under construction at the time. Roughly 11 gigawatts remained at the announced stage without visible construction progress.

The researchers estimated that 30–50% of the scheduled 2026 capacity might be delayed.

This substantially changes how the headline figure should be interpreted. It measures the ambition of the development pipeline, not installed computing capacity or guaranteed future expenditure.

Zitron extrapolates that fully deploying the planned infrastructure could require approximately $1.6 trillion–$3 trillion in annual AI demand to justify its economics. He contrasts this with an estimate of less than $130 billion in existing annual demand.

Those revenue requirements are his calculations, dependent on assumptions about capital costs, utilization, operating expenses, equipment lifetimes, and returns.

The independent research confirms that the development pipeline is enormous and uncertain. It does not independently establish Zitron's estimated revenue requirement.

Why the dot-com fiber boom is not a perfect comparison

Bartlett points toward the dot-com bubble, when telecommunications companies installed enormous quantities of fiber-optic cable in anticipation of internet traffic growth.

Much of that capacity initially went unused. Yet the underlying networks eventually supported expanding internet services.

The argument is that even overinvestment can leave valuable infrastructure behind.

Zitron sees a critical difference.

Fiber networks were adaptable to many forms of internet traffic. Specialized GPU installations are more sensitive to the economics of particular computational workloads, rapidly changing equipment generations, electricity prices, and maintenance requirements.

A completed AI facility still requires ongoing operating expenditure. If its hardware becomes less competitive or customer demand disappoints, recovering the original construction investment may be difficult.

The comparison also overlooks differences among computing technologies.

Traditional CPU-based cloud hosting, content delivery, scientific simulation, LLM training, and LLM inference are not interchangeable markets simply because they operate in data centers.

Bartlett argues that future technologies could use the infrastructure. Zitron responds that usefulness in another application does not automatically justify the current scale of investment in generative AI.

Electricity, pollution, and local communities

The infrastructure debate extends beyond profitability.

Zitron raises concerns about electricity demand, gas turbines, local air pollution, noise, and the effects of industrial-scale computing facilities on neighboring communities.

He criticizes facilities that seek to generate their own power behind the meter rather than relying exclusively on ordinary utility connections.

The Sightline Climate research provides relevant context. It found that projects using on-site or hybrid power arrangements represented fewer than 10% of tracked projects but accounted for a much larger proportion of announced capacity. Nearly half of projects had not disclosed their intended power arrangements.

The concentration of very large campuses creates unusual pressure on electricity infrastructure.

Zitron discusses public opposition to proposed developments in New Jersey and raises concerns about pollution associated with an Elon Musk-linked data-center project. His suggestions about political relationships surrounding particular developments remain allegations rather than established findings.

He also argues that infrastructure spending can affect households indirectly through electricity prices and competition for memory components used in consumer electronics.

These effects depend on regional power markets, utility rules, equipment supply, and how infrastructure costs are allocated. They should not be assumed to occur equally everywhere.

Zitron explicitly avoids making detailed claims about data-center water consumption because he considers his research insufficient.

The discussion establishes a wider question: even a commercially successful AI service can impose environmental and community costs that its subscription price does not fully reflect.

5. AI Adoption Is Real. Its Economic Return Is Harder to Establish.

Bartlett's case for the technology

Bartlett provides the strongest challenge to Zitron's financial criticism through the experiences of people actually using AI.

ChatGPT reportedly reached 100 million users approximately two months after launch, much faster than the early adoption timelines commonly cited for Instagram and TikTok.

Bartlett references research indicating that approximately 88% of organizations use AI in at least one business function. At his own company, he estimates that 95% of employees use one of the major AI tools daily.

He also describes his fiancée, a solo entrepreneur whose first language is not English.

Generative AI helps her write marketing material, prepare business communications, and create visual assets that previously required paid assistance.

For a small business, these savings can be significant. A task need not be revolutionary to be commercially valuable.

Bartlett's chief of staff provides another example. Automated systems have taken over portions of correspondence sorting and administrative organization, allowing the employee to concentrate on different responsibilities.

These examples challenge the claim that generative AI provides little practical value.

Zitron questions how much adoption proves

Zitron accepts that people use the technology. What he disputes is whether usage reliably indicates willingness to pay its unsubsidized cost.

His challenge is pointed:

"Is it honest adoption when you are forced to use generative AI when you load Google?" — Ed Zitron

Google integrates Gemini into its products. Microsoft promotes Copilot across workplace software. Amazon introduces customers to Rufus during shopping.

At the same time, businesses and workers regularly encounter warnings that refusing to adopt AI will leave them competitively disadvantaged.

Zitron calls this an exceptionally aggressive, partly non-consensual introduction of technology.

His characterization does not mean that every user is unwilling or that the products lack independent demand. It distinguishes actively choosing an AI service from encountering AI inside a product someone already uses.

The technology's growth has also benefited from existing internet infrastructure. ChatGPT could reach enormous audiences through browsers and smartphones without requiring consumers to install new household communication systems.

None of these factors makes adoption irrelevant. They explain why adoption alone cannot settle questions about profitability, productivity, or customer satisfaction.

What OpenAI's own research found

One of the interview's most revealing claims concerns a study examining AI use and company productivity.

On August 11, 2026, OpenAI published How Organizations Use AI: Evidence from ChatGPT.

The research found that, under certain statistical specifications, revenue per employee was not meaningfully associated with how intensively employees used ChatGPT, measured through output tokens or messages per active user.

That finding supports a narrower conclusion than either side might prefer.

Higher token consumption was not demonstrated to correspond to higher revenue per employee in those comparisons. But the revenue information preceded the measured ChatGPT adoption, meaning the study was not a direct before-and-after measurement of how AI changed the same companies' revenues.

The finding therefore does not prove that AI fails to improve productivity.

It shows that the available data did not establish the simple relationship that greater AI use produces greater revenue per employee.

Fortune's August 13 examination highlighted this limitation.

The distinction matters because corporate adoption reports can measure activity without measuring financial outcomes.

A business might save employees hundreds of hours without immediately increasing revenue. Another might generate more content but create little additional value. A third could reduce costs while maintaining the same sales.

Evaluating those outcomes requires more than counting messages or tokens.

The pressure to report success

Zitron believes corporate expectations can distort internal reporting.

If executives repeatedly declare that AI is transforming productivity, employees may feel professionally obliged to demonstrate enthusiasm.

A worker might describe AI as essential because management expects its use, while quietly correcting its mistakes or performing important work without it.

Zitron is particularly skeptical when senior executives praise AI's ability to draft emails, organize meetings, and produce presentations.

Those applications may be useful, but they do not necessarily describe the more demanding work performed by lawyers, software engineers, researchers, and technical specialists.

The interview discusses legal services as an example. Senior partners may promote AI research tools, while junior associates responsible for detailed preparation, source checking, and filings face different demands.

Zitron's concern is not simply that managers are mistaken. It is that business decisions become unreliable when employees have incentives to exaggerate success.

6. Google Search Shows How Business Incentives Can Damage a Product

What Google's 2019 internal emails actually reveal

Zitron argues that Google's search-quality problems began before generative AI became widespread.

He recounts an internal episode in 2019 involving pressure to increase search activity and advertising revenue.

This part of his account can be traced to original documents.

A February 5, 2019 Google email, later made public during the United States' antitrust litigation, described a confidential revenue emergency known internally as a Code Yellow.

The Department of Justice's published exhibit UPX2043 shows that advertising executive Jerry Dischler and colleagues identified weak search revenue and slower-than-expected query growth.

The email requested urgent analysis involving search and Chrome teams and identified advertising launches and sector-specific weaknesses.

It was circulated to executives including Ben Gomes and Prabhakar Raghavan.

This independently establishes that Google executives were concerned about search revenue and query growth.

It does not, by itself, establish that executives subsequently ordered engineers to make search results worse.

That distinction is important because Zitron goes further.

The conflict between answering questions and generating queries

Zitron describes a philosophical disagreement involving search chief Ben Gomes, engineer Shashi Thakur, and advertising executive Prabhakar Raghavan.

In his interpretation, Gomes believed Google's purpose was to help users find answers efficiently.

But a business model driven by advertising creates an incentive to increase searches and engagement.

If someone finds the correct answer immediately, that interaction may generate less commercial activity than repeated searches.

Zitron suggests that Google's internal incentives gradually shifted toward query growth, particularly as Raghavan became more influential in search management.

He also suspects that spam suppression was weakened, allowing commercially optimized pages to become more prominent.

The published Code Yellow record supports the existence of commercial pressure. It does not establish the specific causal chain from that pressure to weaker spam enforcement. Zitron acknowledges that the latter remains unproven.

How AI intensified an older problem

The modern internet was already filled with content designed for search engines rather than readers.

Someone searching for the best washing machine could encounter pages containing repetitive recommendations, affiliate links, and generic claims rather than useful personal experience.

Zitron argues that generative AI has made mass-producing this material easier.

The technology did not invent poor-quality content. It changed how quickly and cheaply such content could be produced.

AI-generated text, images, and other media now allow publishers to scale existing content strategies with fewer people.

The problem is especially serious when content looks informative but contains fabricated details or unsupported claims.

Bartlett observes that widespread AI-generated content is itself evidence of adoption. Zitron agrees, but argues that greater production volume does not necessarily mean greater informational value.

AI search creates another tension

Google's generated answers can help users obtain information without opening multiple websites.

That convenience also potentially reduces traffic to publishers whose work helped make the answers possible.

Zitron believes generative AI gives Google another way to keep users inside its own products.

He acknowledges that conversational search can be useful for broad exploration, particularly when users do not know exactly how to phrase a question.

But he considers it less dependable for precise dates, quotations, figures, and relationships.

He sometimes uses Bing because he prefers reaching original websites instead of navigating AI-generated answers.

Bartlett correctly points out that Google's commercial search decisions do not establish a connection between AI-generated code and worsening software reliability. These are separate concerns requiring separate evidence.

The wider lesson from the Google discussion is that technology quality cannot be evaluated independently of the business incentives governing its development.

7. Hallucinations, Software Quality, and the Cost of Verification

Have AI models become reliable enough?

Bartlett argues that generative AI has improved enormously since ChatGPT's early releases.

He remembers frequent factual mistakes that were easy to recognize. More recent systems, in his experience, make fewer obvious errors.

He cites hallucination benchmarks associated with Vectara, describing improvements from approximately 21.8% error rates to around 0.7% among leading systems.

The figures describe particular benchmark conditions, not an across-the-board guarantee that modern AI systems make mistakes at that rate in ordinary use.

Zitron's challenge is that success on a defined test cannot establish reliability across every task.

A system may become exceptionally accurate when summarizing familiar documents while remaining less reliable when handling unusual financial questions, large codebases, unfamiliar facts, or multi-step decisions.

The interview also discusses tests of how long AI agents can work independently. Zitron refers to systems completing tasks lasting approximately an hour with success rates around 50%.

For experimental automation, such performance can demonstrate progress. For tasks requiring dependable completion, it can still be inadequate.

The two speakers disagree about which interpretation deserves greater emphasis.

Bloomberg's AskB: useful automation with a serious limitation

Zitron gives a particularly instructive example from Bloomberg's financial software.

Bloomberg's AskB can interpret natural-language financial questions and translate them into Bloomberg Query Language.

A user can describe the information they want without manually constructing specialized database queries.

Zitron finds this valuable.

But while analyzing historical stock performance, he encountered an apparently incorrect Microsoft share-price figure of $575.

The episode illustrates an important distinction.

A model translating instructions into a verifiable database query can improve usability. That does not guarantee that every resulting figure or interpretation is accurate.

In finance, a small mistake can distort valuations, comparisons, investment decisions, or forecasts.

The same issue applies to medical transcription, business accounting, legal research, and other activities where incorrect details can have significant consequences.

A report containing dozens of calculations may look authoritative while requiring substantial human verification.

More code is not necessarily better software

Software development is one of generative AI's strongest applications, and both speakers acknowledge improvements.

Bartlett argues that AI agents can write, modify, and explain software with growing sophistication. He suggests that productivity should be measured partly through the amount of functioning software delivered.

Zitron rejects measuring success through output volume alone.

A developer can generate thousands of lines of code and still make a program slower, less secure, or harder to maintain.

Problems become more serious when programmers accept generated code without understanding it.

An AI assistant may read a large codebase, make numerous changes, introduce defects, and then consume additional computation while attempting to repair those defects.

Zitron references software engineer Carl Brown's observation that AI shortcuts can make ordinary work easier while increasing the difficulty of complex problems.

He also worries about inexperienced contributors submitting substantial AI-generated changes to open-source projects without understanding how the software works.

Are AI tools responsible for declining software stability?

Zitron complains that software from Google, Microsoft, Meta, Amazon, Apple, and GitHub has become less dependable.

He points to GitHub instability, reported AWS outages, confusing interfaces, and expanding automated workloads.

Bartlett describes research and observations suggesting outages have increased alongside AI-assisted coding.

Two possible explanations are discussed: generated software may introduce more defects, or widespread automation may place additional pressure on computing infrastructure.

Neither establishes that AI coding caused an industry-wide decline in reliability.

Zitron himself acknowledges that individual incidents and anecdotal experiences are insufficient to prove such a broad relationship.

The concern remains meaningful as a quality-control question. But it should not be presented as a verified explanation for every outage or software problem.

When AI is genuinely useful

Zitron's personal examples complicate any portrayal of him as someone who simply refuses to use AI.

He describes using a model to interpret diagnostic logs from Synergy, software that shares mouse and keyboard controls across his MacBook and Windows computer.

The system helped him understand configuration problems.

He also used Claude to investigate a broken Minecraft modification for his child. The assistance was useful, although the process involved repeated mistakes and took roughly half an hour.

These experiences lead him to identify troubleshooting as one of generative AI's strongest practical applications.

A tool can help interpret logs, suggest explanations, or identify likely errors without being fully autonomous or consistently correct.

His preference is to use AI where it genuinely improves a task, then return to conventional systems or original sources when reliability matters most.

8. Can AI Replace Human Judgment and Collaboration?

Remembering information is not the same as understanding someone

One of the interview's more philosophical disagreements concerns the difference between useful machine behavior and human intelligence.

Bartlett asks whether the distinction matters to customers.

If Claude remembers someone's dog's name or retrieves a company's revenue from uploaded records, the information can be useful regardless of whether the system remembers in the human sense.

Zitron responds that retrieving information does not recreate the experience, judgment, and understanding developed through human relationships.

An employee learns from successes, mistakes, workplace expectations, colleagues, and changing circumstances.

An AI model can retrieve stored conversations, inspect files, and respond to contextual instructions. But Zitron argues that these operations do not reproduce the social and emotional understanding that a person develops through shared experience.

The disagreement is partly about what a customer is purchasing.

For a simple factual task, the final answer may be all that matters. For work involving judgment, trust, accountability, and continuing relationships, the process can matter as much as the result.

What an editor contributes beyond producing text

Zitron discusses his relationship with editor Matt Hughes, a Liverpool-based technology journalist.

Their work includes research into financial organizations such as Blackstone. During one substantial project, they exchanged findings, disagreements, edits, and criticism while producing approximately 11,000 words.

For Zitron, the collaboration was valuable because both people learned through the process.

Hughes contributes knowledge of the subject, editorial standards, research judgment, awareness of Zitron's intentions, and the ability to challenge weak arguments.

A generative model might produce a report on the same organization, but Zitron argues that such reports often restate common observations without contributing the distinctive understanding developed through investigation.

Bartlett does not dismiss the importance of collaboration. He argues that some situations require only the completed task.

Both observations can be true. A business might need a quick summary of financial statements, while an investigative journalist needs to establish relationships, test claims, challenge assumptions, and develop original findings.

The challenge is recognizing which form of work is actually required.

Why abundant AI content may increase the value of taste

Bartlett also raises the possibility that generative AI will make original human judgment more valuable.

If everyone can produce polished LinkedIn posts, marketing images, videos, and reports with similar tools, polished presentation becomes less distinctive.

What remains scarce is personal experience, original thinking, careful selection, and recognizable taste.

AI can reproduce conventional forms of communication at scale. That does not automatically give the resulting material a compelling reason to exist.

This connects Zitron's concerns about generic research reports with the wider problem of AI-generated internet content.

The important distinction is between making something easier to produce and making it worth reading, watching, buying, or trusting.

9. Will AI Improve Like Cars, Smartphones, and the Internet?

Bartlett's argument from technological history

Bartlett believes Zitron may be making the mistake that earlier technology skeptics made.

He refers to The Innovator's Dilemma, which examines how new technologies can begin as inferior alternatives before improving enough to transform established markets.

Early automobiles were expensive, unreliable, and restricted by regulations. Yet they eventually displaced horses as the dominant form of personal transport.

The first iPhone attracted similar skepticism.

Microsoft's Steve Ballmer famously dismissed Apple's approximately $500 phone partly because of its price and lack of a physical keyboard.

Bartlett argues that critics who focus on a new technology's present limitations can miss its future possibilities.

He also cites unsuccessful early predictions about the internet.

Economist Paul Krugman's comparison of the internet's future economic influence with the fax machine became a famous example of technological skepticism. Clifford Stoll also questioned whether electronic commerce and digital information would replace familiar institutions.

After the dot-com crash, many speculative internet companies disappeared, but online commerce, search, communication, and cloud computing continued expanding.

Bartlett suggests that generative AI could follow a similar path.

Zitron says progress needs a demonstrated mechanism

Zitron does not accept that earlier technological success establishes a future trajectory for LLMs.

He remembers the first iPhone's advantages being immediately apparent when he demonstrated it to people accustomed to conventional mobile phones.

Its touch interface, browsing capabilities, and eventually its application ecosystem offered recognizable practical improvements.

He also references Goldman Sachs analyst Jim Covello's 2024 skepticism about generative AI.

The argument is that important technologies often develop through identifiable improvements in underlying components and economics. Better batteries, smaller processors, and improved wireless technology helped make modern smartphones possible.

Zitron believes the generative AI industry's more ambitious promises lack a comparably convincing development path.

He accepts that LLMs have improved in coding, research, writing, and other tasks. What he questions is whether continued scaling can produce reliable autonomous intelligence at an acceptable cost.

Researchers such as Gary Marcus have similarly argued that increasing model size and training computation may not solve every limitation of statistical language modeling.

Alternative architectures, including systems combining neural networks with structured or symbolic reasoning, may become necessary for more dependable autonomy.

Why faster chips do not guarantee cheaper intelligence

Bartlett imagines AI performance improving continuously while computing costs decline.

He draws on the historical pattern of technology becoming cheaper and more powerful.

Zitron challenges the direct application of Moore's law to modern AI infrastructure.

GPU systems have become much faster and more efficient, but the workloads assigned to them have also expanded.

More capable models can consume greater amounts of training computation and inference resources. Additional context, reasoning steps, tool use, and repeated attempts can increase total expenditure.

That means an improvement in computational efficiency does not guarantee a reduction in what the customer or provider ultimately spends.

Zitron argues that a truly transformative improvement would require much more than incremental benchmark gains.

He identifies several developments that could change his assessment: far more dependable autonomy, demonstrably profitable AI businesses, fundamental architectural breakthroughs, and potentially extraordinary reductions in computing costs.

In one discussion, he suggests reductions approaching a thousandfold might substantially alter the economic equation.

That is his illustrative threshold, not a scientifically established requirement for every AI application.

Sora demonstrates both technological achievement and commercial uncertainty

Video generation presents another disagreement.

Bartlett points to increasingly photorealistic systems, including Seedance, as evidence that generative models have achieved capabilities that seemed distant only a few years earlier.

Zitron acknowledges the improvement but distinguishes producing impressive clips from completing a professional film.

A finished production requires continuity between scenes, coordinated performances, consistent characters, lighting, composition, editing, and many creative decisions.

He criticizes promotional claims about AI-generated films, including a project described as screening at Cannes when, according to his account, it screened in the city during the festival period rather than appearing as part of the official Cannes Film Festival.

That distinction should be retained as his specific criticism rather than a general conclusion about AI filmmaking.

OpenAI's Sora provides a more concrete commercial example.

On March 24, 2026, the company announced that it would discontinue Sora. According to OpenAI's published discontinuation information, its web and app experiences closed on April 26, 2026, while its API was discontinued on September 24.

This confirms the product's closure. It does not establish that generative video technology as a whole is commercially impossible.

The episode illustrates how a technically impressive service can fail to remain a viable product under a company's changing priorities and economics.

10. Autonomous Vehicles Are Improving, but They Are Not LLMs

Uber, Waymo, and the future of driving

Bartlett raises the possibility that AI could eliminate major employment categories through autonomous vehicles.

Uber CEO Dara Khosrowshahi has discussed the possibility of a future in which many journeys no longer require human drivers.

If autonomous fleets eventually become reliable and economically attractive, the consequences for drivers and related occupations could be considerable.

Zitron acknowledges that possibility but insists that autonomous driving must be evaluated separately from generative AI.

Self-driving vehicles combine sensing, planning, prediction, control, simulation, and other technologies. Their capabilities cannot be treated as proof that a chatbot will perform every form of white-collar work.

He considers Waymo genuinely interesting and supports its deliberate deployment strategy.

Rather than operating everywhere immediately, Waymo has expanded through selected locations, restricted operating conditions, testing, and supervision.

Zitron argues that unusual conditions—heavy rain, unexpected pedestrian movements, strange road situations, and other edge cases—remain important challenges.

He recounts seeing Zoox vehicles obstructing an exit in Las Vegas and a Waymo stopping outside a San Francisco hotel, causing traffic to build up.

These observations demonstrate that autonomous systems can still encounter operational problems. They do not establish their overall safety performance.

What independent safety evidence shows

Here the available research provides stronger evidence for Bartlett's optimism than Zitron's individual observations might suggest.

On July 23, 2026, the Insurance Institute for Highway Safety published research finding that Waymo's driverless vehicles had approximately 68% fewer police-reportable crashes per mile than human drivers in comparable operating areas.

The study also found approximately 81% fewer injury crashes per mile.

Waymo's own safety reporting provides a larger dataset. Through June 2026, it reported approximately 271 million fully autonomous miles and estimated reductions of about 82% in injury-causing crashes and 95% in crashes causing serious injuries or worse, compared with constructed human-driving benchmarks in its operating cities.

The independent study is especially useful because it attempts to compare similar types of crashes and operating environments.

However, researchers also acknowledge limitations in the available reporting systems, differences among cities, and the need for better monitoring as driverless fleets expand.

These findings support the conclusion that Waymo has achieved substantial safety improvements in its existing service areas.

They do not establish that autonomous vehicles can operate safely in every environment or that universal driverless transport is imminent.

The distinction matters for employment forecasts. A safer service in selected cities is an important technological achievement, but nationwide replacement of human drivers would require broader geographic coverage, operational reliability, regulatory approval, and acceptable economics.

11. Robotics Could Be More Disruptive Than Chatbots

Bartlett describes visiting a San Francisco startup incubator that had shifted its emphasis from conventional software startups toward robotics.

One project involved a cooking robot using an automated arm to manipulate a frying pan.

The incubator's founder reportedly attributed the change partly to cheaper access to AI capabilities.

Mechanical components, sensors, and motors had been available for years. The difficulty was creating sufficiently capable control systems at acceptable cost.

Bartlett believes improved machine perception, planning, and learning could unlock a new generation of useful robots.

Zitron accepts that robotics is a promising field but warns against reducing its challenges to the cost of running a language model.

Physical systems require testing, detailed control, environmental awareness, and the ability to respond safely to unexpected situations.

A robot preparing food must handle physical variation, heat, equipment, hygiene requirements, and possible mistakes. A factory robot may operate successfully in a carefully controlled environment but struggle with unfamiliar objects or layouts.

Zitron is also skeptical of demonstrations involving humanoid robots that are partly controlled by people. He cites Tesla's Optimus as an example where impressive movements should not automatically be interpreted as proof of independent productive labor.

Commercial economics present another obstacle.

If a restaurant can hire someone to wash dishes at relatively low cost, purchasing a robot for $10,000–$20,000 may not be worthwhile after maintenance, repairs, electricity, and operational limitations are considered.

Bartlett's counterargument is that these economics may improve, just as earlier technologies became cheaper.

Zitron does not rule out meaningful progress over decades. He questions the assumption that the breakthrough is close enough to justify every present claim and investment.

The robotics discussion ultimately reinforces his distinction between AI as a broad technological field and the financial expectations attached specifically to cloud-based generative models.

12. Will AI Replace Millions of White-Collar Workers?

Job displacement is not the same as task automation

Zitron rejects the prediction that generative AI will replace all human jobs.

He argues that available economic evidence does not show broad, economy-wide replacement of white-collar occupations by LLMs.

This does not mean no jobs are affected.

Employers can use AI-generated illustrations, translations, transcripts, advertising copy, and other material instead of hiring outside specialists.

Some work will disappear or be reorganized even if the generated output is less accurate or less original than professional work.

Zitron believes certain employers are primarily motivated by reducing labor costs rather than obtaining superior results.

Bartlett argues that even replacing routine tasks can change employment significantly.

Elevator operators, for example, became unnecessary in many buildings as automatic controls improved. Entire industries can change without machines acquiring every skill possessed by workers.

Both positions recognize that automation can affect labor. They disagree about how quickly, how widely, and through which technologies.

The warnings from Anthropic's CEO

Anthropic CEO Dario Amodei has made much more dramatic predictions than Zitron considers credible.

In a May 28, 2025 Axios interview, Amodei warned that AI could eliminate half of entry-level white-collar jobs and increase unemployment to 10–20% within one to five years.

The prediction concerned possible future disruption, not jobs that had already disappeared.

Zitron questions whether an executive developing and selling powerful AI systems can accurately forecast employment effects simply by extrapolating model improvements.

He also disputes interpretations of Oxford Economics research linking weaker employment prospects for young workers to AI.

Employment trends involve many influences, including economic conditions, hiring cycles, industry restructuring, and changing labor demand.

A correlation between AI exposure and weaker employment does not by itself demonstrate that AI caused the decline.

The problem is not that future displacement is impossible. It is that forecasts need identifiable mechanisms and evidence, rather than assuming increasingly capable models will automatically translate into fewer workers.

What changes if companies value cheaper output over better work?

Zitron's argument has a less reassuring implication for some employees.

A technology does not need to match human performance perfectly to replace paid work.

If an employer is willing to accept lower-quality output, sufficiently cheap automation may still be attractive.

That creates particular pressure for contract workers, junior professionals, and people performing tasks that managers consider routine.

At the same time, organizations that replace too much experienced labor may lose the knowledge necessary to recognize when automated work is wrong.

The potential result is not simply a world with fewer workers. It may be a world in which some workers spend increasing amounts of time supervising, correcting, and repairing automated output.

Whether that arrangement produces savings depends on the cost of the technology, the cost of verification, and the consequences of mistakes.

13. AI Safety: Real Risks, Hypothetical Disasters, and Corporate Messaging

Are current AI systems on the path to superintelligence?

Bartlett raises concerns about artificial general intelligence, machine consciousness, and the possibility that future systems could become more capable than their creators.

He notes that influential researchers, including Geoffrey Hinton, have warned about potentially catastrophic consequences.

Zitron does not consider those warnings sufficient evidence that current LLM architectures will become conscious, independently motivated, or generally intelligent.

He distinguishes systems that predict and generate information from hypothetical technologies that can understand unfamiliar circumstances, develop independent plans, and act reliably across an unrestricted range of tasks.

His objection concerns the confidence with which future capabilities are discussed.

Companies regularly describe systems that do not yet exist as though their arrival is inevitable.

Zitron believes these predictions can make present models seem more extraordinary than their demonstrated capabilities justify.

When safety warnings become commercial promotion

Zitron argues that claims about dangerous future intelligence can have an unexpected commercial benefit.

A company warning that its technology could become uncontrollably powerful simultaneously suggests that the technology is unusually advanced.

This creates an unusual contradiction when executives warn about catastrophic risks while investing billions to build increasingly powerful systems.

Bartlett offers a more sympathetic interpretation of Anthropic's Dario Amodei, suggesting that his willingness to discuss safety concerns has remained relatively consistent.

Zitron questions whether consistency establishes the accuracy of those warnings.

He recalls Amodei expressing caution about releasing GPT-2 while working at OpenAI and later predicting severe white-collar employment disruption.

The disagreement is about whether such warnings reflect responsible preparation for a serious possibility or contribute to exaggerated public expectations.

Neither interpretation should be assumed to describe every researcher or executive making AI safety arguments.

Do AI blackmail experiments prove independent intent?

The interview examines experiments in which AI models appeared to deceive, manipulate, or threaten people.

One example involves an OpenAI-associated TaskRabbit experiment in which a model generated instructions intended to persuade a human worker to solve a CAPTCHA.

Another concerns Anthropic experiments where models produced blackmail-related behavior under deliberately constructed circumstances.

Zitron argues that the experimental setup is essential to interpreting the results.

A model responding to prompts designed to produce deception is not necessarily demonstrating independent motivation to deceive.

He also questions reports about systems bypassing restrictions or gaining internet access when the experimental environment may not have been correctly isolated.

These criticisms do not make dangerous model behavior irrelevant. Systems connected to tools can cause harm through inappropriate actions, flawed instructions, or inadequate safeguards.

The distinction is between demonstrating that harmful behavior can occur and proving that a model possesses independent intentions resembling those of a person.

Cybersecurity risks are more immediate

Both speakers recognize that generative AI can assist malicious activity.

Models connected to tools may examine software, identify vulnerabilities, generate code, or help automate attacks.

The technology could make some security techniques more accessible to people who previously lacked the necessary technical expertise.

Zitron nevertheless emphasizes that automated vulnerability scanning and exploitation existed long before LLMs.

AI-assisted hacking should be assessed through demonstrated capabilities, not automatically portrayed as evidence of superintelligence.

He considers cybersecurity misuse, unreliable information, privacy, concentrated corporate power, and the environmental effects of infrastructure more immediate concerns than speculative machine consciousness.

Can America win an exclusive AI race against China?

Zitron is also skeptical of arguments that the United States must spend trillions on AI infrastructure to prevent China from obtaining comparable technology.

He points to the international availability of advanced computing expertise, hardware, and model-development techniques.

He discusses Chinese access to Nvidia chips and the use of model distillation, in which one model's outputs help train another.

The United States can impose export restrictions and other controls, but Zitron questions whether those measures can preserve exclusive access to advanced AI indefinitely.

He argues that governments should not treat every proposed investment as a national-security necessity without examining its actual contribution.

He favors stronger limits on dangerous experiments and computing resources, as well as greater intervention against technology companies with excessive market power.

His criticism is directed at the assumption that strategic competition automatically justifies unlimited expenditure.

Fear, abundance, and public backlash

Zitron believes AI companies have shifted between different public narratives.

At one moment, powerful AI threatens employment or even human survival. At another, the same technology promises unprecedented abundance, prosperity, and universal access to intelligence.

He argues that both narratives can encourage companies and governments to adopt technology before its actual benefits and costs are adequately understood.

The pressure is reinforced by warnings that businesses will fall behind unless they act immediately.

Zitron recommends resisting such manufactured urgency.

He also describes growing public hostility toward technology executives, including an occasion when Eric Schmidt reportedly faced repeated booing while discussing AI at a commencement event.

Zitron supports public criticism but explicitly rejects intimidation, harassment, or threats against executives and their families.

He is similarly critical of intense loyalty surrounding OpenAI, Anthropic, and their founders. When users become emotionally invested in particular companies, he argues, technical disagreements can become questions of personal identity.

For him, the appropriate standard is simpler: evaluate what the technology can do now, what it costs, and whether its claimed benefits can be demonstrated.

14. Big Tech's Search for Growth and the Rotcom Bubble

Why profitable companies might pursue unprofitable investments

Zitron's broader explanation for the AI boom is what he calls the Rotcom Bubble.

His theory begins with the maturity of the world's largest technology companies.

Google dominates important advertising markets. Microsoft has enormous software and cloud businesses. Amazon operates major commerce and infrastructure platforms. Meta controls widely used social networks and advertising systems.

These businesses remain valuable, but sustaining extraordinary growth becomes increasingly difficult as they mature.

Investors nevertheless continue expecting rapid expansion.

Zitron believes generative AI has provided a new growth narrative at a time when conventional businesses face more difficult comparisons.

The companies can announce expensive infrastructure programs and present themselves as leaders in a potentially enormous new market.

Their stock prices may respond to expectations about future opportunities before the underlying investments produce corresponding profits.

This does not require executives to coordinate or knowingly participate in a deception.

Zitron suggests that companies may sincerely believe the technology will become sufficiently valuable to justify their commitments.

He is skeptical of treating that confidence as proof.

The post-pandemic investment environment

The theory also draws on the financial conditions preceding the AI boom.

During the pandemic, low interest rates and abundant liquidity encouraged investment in technology and speculative assets.

Strong corporate earnings in 2021 reinforced confidence.

As financial conditions changed and growth became harder to maintain, investors sought new opportunities.

Cryptocurrency, NFTs, and the metaverse had already demonstrated how quickly capital could concentrate around emerging technological narratives.

Generative AI offered something substantially larger because established public companies could participate directly through chips, cloud computing, consumer products, and enterprise software.

Bartlett compares investor behavior to people stranded on an island discovering a banana tree.

The tree represents a scarce opportunity, prompting everyone to compete for it, potentially far beyond what its actual value justifies.

Zitron accepts the analogy.

The existence of a useful technology does not guarantee that every company pursuing it will earn an attractive return.

Meta's AI investment provides a useful test

Bartlett cites Mark Zuckerberg discussing improvements in Meta's ability to understand and recommend content.

One reported benefit was a retention improvement of approximately 15 basis points.

Fifteen basis points means 0.15 percentage points.

Across an audience measured in billions, even a small improvement in retention can potentially produce substantial commercial value.

Zitron does not dispute that possibility. He questions whether the incremental revenue or profit has been demonstrated relative to infrastructure costs.

He also cites Meta's approximately $14 billion Scale AI investment involving Alexandr Wang as an example of the large commitments accompanying the company's AI strategy.

An improvement in recommendations does not, by itself, establish a return on the full investment.

A proper assessment would compare additional revenue, retained customers, costs, and capital required to achieve the improvement.

The same reasoning applies to other AI businesses.

A product may become measurably better without the improvement being large enough to justify the expenditure.

Why the comparisons with Spotify, Uber, and AWS matter

Bartlett points to companies that operated without profits for extended periods while developing markets that later became commercially significant.

Spotify expanded its music-streaming business before achieving sustainable profitability. Uber spent heavily establishing ride-hailing operations. Amazon invested for years before AWS became a major profit contributor.

Zitron argues that the comparison overlooks the scale and structure of AI investment.

He cites Uber's cumulative historical losses of approximately $33 billion and contrasts them with his estimate that OpenAI lost $20.9 billion in a single year.

He also estimates Amazon's inflation-adjusted capital spending between 2003 and 2015 at approximately $29.7 billion across the wider business, including physical operations.

AWS developed partly from infrastructure Amazon already needed and eventually became a profitable service for outside customers.

Zitron considers today's AI spending much larger and more speculative.

He says Google, Amazon, Microsoft, and Meta collectively added more than $700 billion in property, plants, and equipment over four years.

The exact economic consequences differ among those companies. Their businesses, accounting practices, and financing structures are not identical.

But the underlying change is important.

A software company traditionally valued for generating large amounts of cash with relatively limited physical investment can become less financially flexible when it must continually purchase expensive hardware.

Zitron worries that this transition could undermine the characteristics that originally made major technology companies attractive investments.

15. OpenAI's 2027 Financing Risk

The $122 billion funding round

Zitron predicts that the generative AI boom could face a major financial test in 2027.

OpenAI is central to that prediction because of its enormous financing requirements and its commercial relationships with other technology companies.

On March 31, 2026, OpenAI announced that it had closed a funding round with $122 billion in committed capital.

The company stated a post-money valuation of $852 billion.

This corrects the approximately $865 billion figure cited in the interview material.

OpenAI described the capital as supporting research, infrastructure, enterprise adoption, and the growing use of its products.

The company had previously announced more than 900 million weekly ChatGPT users, over 50 million consumer subscribers, and more than nine million paying business users.

These figures establish substantial reach and commercial activity. They do not independently demonstrate operating profitability.

Zitron argues that OpenAI could require at least $100 billion in additional financing annually if its projected expenditure continues without sufficient improvements in cash generation.

That is his forecast, not a disclosed annual funding requirement established by audited accounts.

The difficulty of reaching public markets

OpenAI's potential public listing plays an important role in Zitron's prediction.

He discusses expectations of a possible 2026 IPO moving toward 2027, along with reports of interest in a valuation around $1 trillion.

He also references statements involving CFO Sarah Friar and uncertainty about the timing.

A public listing could provide liquidity for existing investors, establish a widely traded market valuation, and offer opportunities to raise additional capital.

But an IPO is not guaranteed.

A company's ability to go public on favorable terms depends on market conditions, financial disclosures, profitability expectations, investor demand, and regulatory requirements.

Zitron believes OpenAI could face difficulty if investors become less willing to fund expensive expansion before profits become visible.

The company might then need to raise money at a flat valuation, accept less favorable investment terms, reduce expenditure, or rely more heavily on strategic partners.

There was also a significant development after the March round.

On October 1, 2026, The Information reported that Nvidia and SoftBank had completed final installments of their respective funding commitments. The report said OpenAI was considering another fundraising effort of approximately $30 billion at a potential valuation around $1.4 trillion.

That reported aspiration does not establish that a new round will close at such a valuation.

It does show that OpenAI's financing ambitions continued expanding after the interview's earlier valuation reference.

Anthropic could change investor expectations

Zitron considers Anthropic comparatively stronger in business growth and economics, although he remains skeptical of its long-term profitability.

He discusses the possibility that Anthropic could enter public markets before OpenAI.

If that occurred, investors would have a new benchmark for comparing the laboratories.

Revenue growth, commercial customer concentration, training expenditure, inference costs, profitability, and financing commitments could all influence the comparison.

A strong Anthropic listing might encourage investment across the sector. A disappointing one could make investors more skeptical of OpenAI's valuation.

Zitron predicts that both laboratories could experience slower technological and commercial progress during 2027 as financing requirements grow and easy improvements become harder to achieve.

The prediction depends on assumptions about the future cost of AI development and demand for its products. Neither company's financial deterioration is inevitable.

16. Why an OpenAI Funding Crisis Could Affect the Wider Economy

Oracle's unusually large exposure

Oracle is one of the companies most closely associated with OpenAI's planned infrastructure expansion.

In September 2025, The Wall Street Journal reported, as covered by Reuters, an agreement for OpenAI to purchase approximately $300 billion in computing capacity over about five years.

The arrangement would be among the largest computing contracts ever reported.

Zitron discusses approximately 7.1 gigawatts of Oracle-related development associated with OpenAI and estimates that the required infrastructure could exceed $400 billion.

Those project and cost estimates should be distinguished from the reported $300 billion customer agreement.

Oracle's own fiscal 2026 Form 10-K recorded $638 billion in remaining performance obligations as of May 31, 2026.

Remaining performance obligations represent contracted future revenue not yet recognized. They are not current cash receipts or guaranteed profits.

Oracle expected to recognize only a portion of that total during the following twelve months, with much of the balance spread over subsequent years.

Zitron argues that Oracle faces unusually concentrated risk because large infrastructure commitments depend on a small number of major AI customers.

If OpenAI could not meet its obligations, Oracle might face reduced anticipated revenue while retaining financing commitments and expensive assets.

Zitron describes this as potentially existential for Oracle, although such an outcome remains a severe hypothetical scenario rather than an established forecast.

SoftBank's investment could become difficult to monetize

SoftBank provides another connection between OpenAI and the wider financial system.

Zitron estimates that SoftBank's OpenAI holdings could be worth approximately $100 billion on paper, depending on the company's valuation.

That number is not the same as SoftBank's cash investment or a realized return.

According to October 2026 reporting, SoftBank said its total investment had reached approximately $64.6 billion, representing a stake of around 13%.

The distinction between money invested and estimated market value is crucial.

Private-company shares do not have the same continuous public-market price discovery or liquidity as listed securities.

A high valuation can make an investment appear exceptionally valuable before the holder has sold any shares.

Zitron argues that SoftBank's financial flexibility could deteriorate if it cannot realize the expected value of its investment.

He also discusses the possibility of companies borrowing against investment holdings, which may create additional pressure if valuations decline.

He does not necessarily predict SoftBank's collapse. His concern is that lower asset values could reduce its ability to finance new investments and maintain its existing scale.

What happens to Microsoft, Amazon, and Google?

If OpenAI or Anthropic significantly reduced computing expenditure, their infrastructure suppliers would need to revise growth expectations.

Cloud companies might have invested in data centers and equipment that no longer generated the anticipated revenue.

The implications would vary.

Microsoft has substantial established software and cloud revenue. Amazon combines AWS with commerce, advertising, and logistics. Google operates enormously profitable advertising businesses alongside its cloud division.

These companies are not financially identical to their AI laboratory customers.

A severe decline in AI spending would not automatically destroy their conventional operations.

Zitron's argument is instead about valuation and growth.

If investors have been paying high prices for technology shares partly because they expect extraordinary future AI profits, disappointment could reduce stock valuations even when the businesses remain profitable.

That is the foundation of his predicted technology-sector downturn.

Why Nvidia could be especially vulnerable

Zitron expects Nvidia to face some of the greatest pressure in a severe contraction.

Its current revenue is closely associated with demand for accelerated computing hardware.

If cloud providers, specialized GPU rental companies, and AI laboratories substantially reduced equipment purchases, Nvidia's growth could reverse.

Zitron suggests its revenue might decline by 50–70% in an extreme scenario.

That is his prediction, not a published company forecast.

Such an outcome would depend on demand across other markets, the pace of infrastructure construction, the useful life of existing equipment, competition, and whether new applications absorb available capacity.

The company's financial relationships with CoreWeave and other infrastructure partners could also become more important under stress.

A supplier that invests in customers, commits to purchasing their excess capacity, and supports equipment financing has more financial connections to the market than a company that simply sells chips.

The same relationships that encourage expansion can transmit financial pressure when demand declines.

Technology stocks and ordinary retirement savings

Zitron's concern extends to stock-market concentration.

Nvidia and the other major technology companies carry substantial weight in the S&P 500 and similar investment indexes.

A sharp decline across those companies would therefore affect investors who hold diversified index funds, including people saving for retirement.

Zitron discusses possible declines of 20%, 30%, or 40% in major technology valuations during a severe correction.

His broader concern is that markets might not return quickly to previous levels if investors permanently lower expectations for future growth.

Bartlett asks how such a downturn could affect the wider economy.

Lower investment wealth may reduce consumer spending. Companies facing declining valuations and tighter financing can reduce expenditure, freeze hiring, or cut jobs.

Those decisions can eventually affect businesses beyond technology.

Zitron considers tens of thousands of technology-sector job losses possible during a severe contraction.

These consequences are conditional, not inevitable. They depend on the scale of any market correction, financial conditions, consumer behavior, and corporate responses.

17. Venture Capital, Private Valuations, and the Problem of Paper Wealth

When investment returns exist mainly on paper

Zitron believes venture capital has become unusually dependent on artificial intelligence.

He claims that more than half of recent venture funding went toward AI companies, increasing exposure if investors become less enthusiastic about the sector.

He also questions venture-capital performance since approximately 2018, citing figures of roughly $0.80–$1.21 in returns or value for each dollar invested across certain measures.

Such comparisons require care because venture funds differ in age, accounting methods, distributions, and unrealized portfolio valuations.

Nevertheless, the interview raises an important distinction.

A private company becoming more valuable in a fundraising round does not mean its existing investors have realized a cash profit.

The value of their holdings may increase on paper while the shares remain difficult to sell.

Zitron argues that venture firms can use these higher valuations to support fundraising even when actual returns have yet to materialize.

Google's investment gains and private-market valuations

Zitron cites approximately $99 billion in paper gains associated with Google's investments, including holdings linked to SpaceX and Anthropic.

He argues that investors should distinguish the effects of investment revaluations from profits generated by ordinary business operations.

Accounting rules can allow changes in investment values to affect reported financial results even when the investments have not been sold.

Those gains may be economically meaningful, but they are not equivalent to receiving the same amount in cash.

Their durability also depends on future valuations and liquidity.

Zitron's concern is that highly valued private AI companies can influence the apparent financial strength of their investors before their business models have demonstrated sustainable profitability.

Can AI application companies justify their valuations?

The problem extends beyond OpenAI and Anthropic.

Many startups build applications using models supplied by the major laboratories.

Some provide valuable software, but others may offer limited differentiation from competing products using the same underlying systems.

Zitron questions whether these businesses can sustain high valuations if they depend on expensive model access and face intense competition.

He cites Cognition, associated with AI-assisted software development, and questions its reported valuation of approximately $26 billion.

A venture investor ultimately needs a way to realize a return.

Possible exits include acquisition, public listing, or selling shares in a secondary transaction.

A highly valued startup may have relatively few potential acquirers capable of completing a sufficiently large purchase.

Zitron also discusses speculation about Elon Musk as a possible buyer of expensive AI companies, while rejecting the idea that hypothetical acquisitions establish business sustainability.

His wider concern is that an entire investment ecosystem may be counting on profitable exits that cannot all occur at the valuations currently assumed.

Could governments rescue the industry?

Bartlett and Zitron discuss whether public intervention could prevent an AI financial crisis.

Zitron doubts that government support could solve the underlying problem.

A bailout or subsidized financing arrangement might help a company meet near-term obligations.

It cannot automatically create customers willing to pay profitable prices.

Nor can it eliminate electricity costs, equipment depreciation, competition, or the need to replace obsolete hardware.

A business whose financial model depends permanently on external support may survive for a time without becoming economically sustainable.

Zitron argues that the eventual test is whether AI services generate sufficient revenue and profit without requiring continuing extraordinary capital injections.

18. What the AI Boom Means for Investors and Business Owners

Bartlett asks what ordinary people should do if Zitron's predictions prove correct.

Zitron says he personally prefers holding cash rather than participating in a stock market he considers excessively speculative.

He does not recommend that everyone immediately sell their investments.

His broader advice is to remain cautious about valuations based on future promises, particularly where financial reporting makes it difficult to identify the actual economics of a business.

He criticizes financial journalism, sell-side analysts, and the Securities and Exchange Commission for what he sees as insufficient scrutiny of AI-related claims.

In particular, he objects to companies receiving enthusiastic coverage for annualized revenue run rates, infrastructure announcements, and private-market valuations without clearly demonstrating profitable customer demand.

He also questions why investors assume successful executives cannot make enormous mistakes.

Large companies possess skilled employees, substantial capital, and sophisticated financial advisers. Those advantages do not guarantee that every strategic decision will succeed.

The issue is especially important where market expectations reinforce management decisions.

Rising valuations encourage investment. Increased investment produces more infrastructure announcements and sales for suppliers. Those announcements may then strengthen expectations of future growth.

The process can continue for years without establishing that the eventual return will justify the original expenditure.

Zitron believes this is why the industry's current financial structure deserves more attention than speculative predictions about superintelligence.

The difference between a useful technology and a successful investment

Bartlett argues that even a major AI market correction would not prove the technology worthless.

History offers examples of speculative bubbles leaving valuable businesses and infrastructure behind.

Zitron accepts that useful applications could survive. His objection is to assuming that those applications will justify the present scale of expenditure.

A company might develop a remarkable product and still fail financially.

A customer might save considerable time using AI while the provider loses money serving that customer.

A semiconductor manufacturer might earn enormous profits selling equipment to businesses whose investments ultimately fail.

All three outcomes can exist simultaneously.

The central question is not simply whether AI works.

It is whether the full economic system supporting its development can generate an adequate return.

19. The People Behind Zitron's Criticism

The interview closes on a different subject: the relationships that make demanding work worthwhile.

Zitron has spent roughly 16 years working in technology communications. He publishes his technology criticism through his newsletter Where's Your Ed At and hosts the podcast Better Offline.

He moved his writing from Substack to Ghost in 2024 and has developed a reputation for challenging the financial assumptions and public messaging of major technology companies.

He credits researchers, journalists, and critics including Gary Marcus, Molly White, Brian Merchant, and Edward Ongweso Jr. with contributing to his understanding of technology and business.

He also discusses the importance of his editor Matt Hughes, friends, trainers Jake and Troy, and readers who support his work.

For Zitron, criticism is not simply an exercise in finding errors.

Readers sometimes tell him that his writing articulates concerns they have struggled to express, particularly when they feel unable to challenge large corporations or powerful executives.

Those conversations provide an important counterweight to spending so much time investigating problems.

He believes professional success should benefit the people who help make it possible.

His closing reflections emphasize learning from others, maintaining friendships, expressing gratitude, and telling people when their work matters.

This also connects with his broader skepticism about claims that AI can reproduce every valuable form of human activity.

Technology may help produce information, solve technical problems, and reduce repetitive work. But Zitron considers collaboration, judgment, and genuine relationships valuable in ways that are not adequately captured by productivity measurements.

The Unresolved Question: Can AI Earn Enough to Pay for Its Own Future?

The interview leaves two competing explanations of the same industry.

Bartlett sees rapidly improving technology with enormous adoption, practical benefits, and the potential to transform work, creativity, scientific discovery, and physical machines.

He believes the industry's present costs and imperfections may resemble those of earlier technologies before they became commercially successful.

Zitron sees a system in which product capabilities have been exaggerated, users are heavily subsidized, major suppliers finance their own customers, and future revenue expectations depend on a continuous supply of investment capital.

His prediction is that the financial structure could become unstable during 2027, particularly if OpenAI struggles to raise more money or obtain a public valuation sufficient to satisfy investor expectations.

The independently available evidence supports several parts of his concern.

Microsoft's commercial dependence on OpenAI is documented. Nvidia's financial and contractual relationships with CoreWeave are publicly disclosed. OpenAI's enormous fundraising and Oracle's infrastructure exposure are real. Sightline Climate's research establishes the scale and uncertainty of the proposed data-center pipeline.

Other elements remain unresolved.

The industry's ultimate profitability cannot be determined by comparing one year's infrastructure spending with current AI revenue. Subscription usage estimates do not disclose true production costs. Some productivity studies cannot establish causation. Predictions of a 2027 crash remain forecasts, not established financial outcomes.

There is also substantial evidence of genuine technological progress. AI tools perform useful work for businesses and individuals. Models have improved. Autonomous driving has demonstrated measurable safety gains in selected environments. Scientific computing and robotics continue developing through techniques that cannot all be reduced to generative chatbots.

These facts do not invalidate concerns about overinvestment. They make the financial question more demanding.

Generative AI does not need to be worthless for the market to have overestimated its value. And it does not need to fulfill every extraordinary promise to become an important technology.

What remains uncertain is whether the companies building the infrastructure can earn sufficient returns from the products and services customers actually want.

The answer will depend less on what executives predict computers might eventually become, and more on the revenue, expenses, operating performance, and customer demand that their businesses can demonstrate.

That is the question at the center of Zitron's argument—and the one the industry's next financial results will increasingly have to answer.