The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron
In a Nutshell
Ed Zitron argues generative AI is a fundamentally unprofitable con built on circular funding between Big Tech companies, where OpenAI and Anthropic's massive losses are subsidized by Microsoft, Google, and Amazon while the entire industry hides actual AI revenue figures and runs on hype rather than demand. The trillion-plus dollars being spent on data centers and GPUs creates infrastructure that will cost as much to operate in 2050 as today, with no path to profitability since even power users burn $400-14,000 in tokens monthly while paying just $20-200 subscriptions. Zitron predicts this breaks in 2027 when the unsustainable economics force massive restatements, IPO failures, and a tech depression as the $1.1 trillion in cloud commitments to two unprofitable companies can no longer be maintained.
These notes were generated by AI and may contain inaccuracies.
Ed Zitron describes generative AI as fundamentally a con, arguing that ultra-rich and ultra-powerful people are lying about its capabilities and financials. He states that the technology has been sold from the beginning in terms of magic, but is actually a half-assed autocomplete machine that is misleading the entire world. Zitron notes that he has been in the tech industry for 16 years and loves technology, but does not like being misled, calling this the largest non-consensual push of technology in history.
Zitron disputes the claim that the AI industry is creating enormous economic growth, stating that all of these companies run at horrifying losses. He cites that OpenAI lost $20.9 billion last year, and notes that none of the companies can say they are on the path to making this profitable because they cannot.
Zitron states that AI will replace all human jobs is not happening and there is no economic data to support it.
Zitron questions the need for the United States to spend trillions to beat China in the AI race, asking what the race accomplishes. He notes that people keep saying what if these models fall into the wrong hands, but states they are already in the wrong hands with Mark Zuckerberg, Sam Altman, and Dario Amodei.
Zitron criticizes Mark Zuckerberg's statement that they will continue to invest aggressively in infrastructure to meet the demand, calling it a monstrosity. He references the Shrek line some of you may die, but that's a risk I'm willing to accept, and notes that if only these people cared about poverty or actual problems in the world versus buying enough GPUs.
Zitron explains his controversial opinions about generative AI, stating that it is at its heart a con and is not sold as honest software. He believes the industry overstates what AI can do, what it will do, and the underlying financials to the point that they are misleading the entire world. He argues they are actively exploiting weaknesses in journalism, economies, and responsible parties including sellside analysts and governments.
Zitron addresses the strong use of the word con by explaining that from the very beginning, AI has been sold in terms of magic as something that will replace all jobs and cure cancer. When examined closely, he states it is boring cloud software that is extremely expensive, unprofitable, and unreliable at its core.
Zitron addresses questions about his references and personal experience, noting that he has been in the tech industry for 15-16 years in PR but has had practical experience. He states that people say he does not have finance experience, but he loves technology and is enthusiastic about it. He notes that AI came along as something everyone told him was the best thing since sliced bread, but it cannot even do the basics like search.
Zitron describes how when you ask an AI person about their setup, they describe something like PeeWee's Playhouse with multiple harnesses, right prompts, and specific models for different tasks. He notes that this is meant to be artificial intelligence that is smart, autonomous, and something you set and forget.
Zitron discusses the six leading AI companies: Anthropic, Amazon, Nvidia, Microsoft, OpenAI, and Google. He states that their fundamental business model is a con, noting that their revenues are not really coming from AI. Up until recently, none of their revenues were coming from AI, with only dribbles of revenue currently.
Zitron states that 70% of all AI revenues across Amazon, Microsoft, and Google are from OpenAI and Anthropic, which are unprofitable, unsustainable companies that cannot afford to exist without these companies giving them money. He notes that Amazon sent $50 billion to OpenAI this year and $5 billion to Anthropic, while Google sent $10 billion to Anthropic.
Zitron cites sellside analyst evaluations projecting that OpenAI and Anthropic will generate 400 or more billion dollars of revenue in the next three and a half years, representing 30% of cloud growth from these two unprofitable companies that will need money from somewhere.
Zitron criticizes the companies for having such low respect for investors and analysts that they do not disclose their AI revenues. When they do mention it, they use a run rate or annualized run rate, which he states means nothing and is never defined. He notes that it can mean months 12, month 13, or last 4 weeks time 13, and is different every time.
Zitron notes that these are public companies, or at least the ones that are not Anthropic and OpenAI. He states that when they have good news, they will tell you, and when they do not tell you something, that actually speaks volumes.
Zitron defines capital expenditures as long-term investments that are theoretically one-off, such as data centers or the GPUs put inside AI data centers. He contrasts this with operating expenses like electricity, which come off immediately.
Zitron explains that AI GPUs are much bigger, much more power intensive, and take a bunch of high bandwidth memory. He notes that because of how many GPUs are needed, thousands, tens of thousands, or hundreds of thousands in some cases, a bunch of power is required.
Zitron provides the example of OpenAI and Oracle building a data center in Abilene, Texas called Stargate Abilene with 1.2 GW capacity. He notes that within each of the eight buildings, there will be 50,000 Nvidia GB200 GPUs. He compares this to the city of Bristol, which takes about 7800 megawatt of power a year, noting that Stargate Abilene is condensing more power than that into a space 1,172 times smaller.
Zitron states that all of these data centers cost billions of dollars, and all of these companies other than Microsoft are now taking out debt. He notes that they have spent over a trillion dollars so far and want to spend another trillion dollars next year, all to make tens of billions of dollars, most of which comes from two unprofitable companies, Anthropic and OpenAI.
Zitron questions whether adoption is honest when people are forced to use generative AI when they load Google, with Gemini screaming in their ear, or when loading Word with Copilot bugging them. He states that when using Amazon, whatever Rufus AI is wants has opinions on what socks people are buying, calling this the largest non-consensual push of technology in history.
Zitron notes that ChatGPT for example has had every single media outlet screaming about it for 3 years, saying it will take jobs and that people must use it or fall behind. He states that people are using it because they have been told to use it constantly, and they are using it like search predominantly because Google fell behind search and because it is better at ingesting queries sometimes.
Zitron states that even when generative search works, it is not worth a trillion dollars. He notes that the amount of money being sunk into this is incomparable to anything, including railways, because there is no post-bubble story even for this. He states that AI GPUs are not useful for other things either.
Zitron describes the situation as a directionless egregore of capitalism, a headless beast that lumbers around desperate to seek out growth everywhere in the hopes that if it harasses people, scares people, and demonizes labor enough, people will be forced to use it.
Zitron acknowledges that 88% of organizations regularly use AI at least once for one particular business function, and that 95% of people in the interviewer's company use AI tools like Anthropic, ChatGPT, or Gemini every day. He notes that this exists on a spectrum from super users to those using it less.
Zitron addresses the issue of AI slop on the internet, noting that before AI slop, there was SEO slop because Google incentivized doing the lowest common denominator that would rank well in search. He references a story about how Google pulled back spam guards thanks to Prabhakar Raghavan, which made the internet worse by allowing worse content to rank higher.
Zitron explains the cost structure of AI services, noting that when using AI services, you burn tokens and it is per million tokens. He defines a token as around 3/4 of a word, or characters. He compares it to a taxi in New York having a meter, with the AI companies having a currency in which they charge you called tokens.
Zitron explains that you are charged per million input tokens, which is the stuff you feed into it like a document or a bunch of code base. Output tokens are both the stuff it spits out at the end but also when it thinks. He gives the example of asking for the best restaurants in an area of New York, noting that all of the thinking process counts as output tokens.
Zitron notes that when paying for a monthly service, you do not see any of the token costs, as they just have rate limits. He cites a SemiAnalysis finding that on a $200 a month ChatGPT subscription, you can burn $14,000 worth of tokens, and on Anthropic you can burn $8,000 for $200. He notes that even on the $20 a month service, you can burn $400.
Zitron states that most people do not realize what AI costs and just think it is $20 a month. He notes that all of these companies run at horrifying losses, with OpenAI losing $20.9 billion last year because people can burn as many tokens as they want.
Zitron describes how when companies tried to move everybody on the enterprise side onto actually paying the cost of AI around March of 2026, Sam Altman said people have a big problem with it and it is a huge issue. He notes that enterprises immediately started freaking out, with Uber burning through their entire annual token budget in three months.
Zitron explains that a power user could be costing Anthropic or OpenAI $1,000 but only being charged $100, meaning they are subsidizing $900 of usage because of electricity costs and data center costs. He asserts that this is unsustainable.
Zitron explains that when standing up inference, which creates the output within data centers, you are not just turning on an inference machine. You are standing up the GPUs necessary to take in the demand, and if you buy too much, you have wasted the money because you have to pay for the hourly GPU use regardless.
Zitron states that the simplest way to explain it is that these services would be profitable if they actually believed they were worthwhile and worthy of the cost, they would charge it. He notes that regular people would not be able to get a monthly subscription and would just be paying what it is worth.
Zitron notes that you pay when you use an LLM regardless of whether you get what you want. When these things hallucinate, such as when coding something and they break a bunch of stuff, you are paying for that whether it works or not, unless using one of the subscriptions.
Zitron references the book The Innovator's Dilemma, noting that it talks about how innovation that ends up taking out or transforming an industry often starts worse, does not make economic sense, and none of your customers are asking for it. He gives the example of cars in the 1800s, which broke down all the time, were more expensive, and required someone to walk in front with a red flag.
Zitron states that if they were bringing the cost down, they would have brought the cost down, but it seems to be getting more expensive. He notes that even inference providers do not seem to be profitable, and even companies renting out GPUs do not seem to be profitable.
Zitron explains that the companies kept going because it helped number go up so much, helping stock values pump, helping Nvidia pump, Microsoft, and everyone, but not from the revenues. He notes that people have been saying their AI bets have paid off, but these companies refused to say how much they are making from AI.
Zitron states that existing businesses continued to grow through price increases, changes to how Google and Meta did advertising, and Amazon bumping up prices and starting a remarkable ad business, nothing to do with AI. But because number go up and revenue go up, everyone said it is AI.
Zitron cites that in fiscal year 2026, which just ended for Microsoft, they made total according to Bloomberg about $34.33 billion, with $24.1 billion of that from OpenAI, leaving them with about $10 billion in a year when they spent $115 billion on capital expenditures and intend to spend $175 billion next year.
Zitron states that the math does not make sense. He notes that large language models need a bunch of money to train them, need constant data flow, and need customized data, making it a big expensive monster.
Zitron describes how when you tell people about Nvidia selling $215.9 billion worth of GPUs in the last fiscal year mostly to support $22 billion of revenue total in the entire world outside of these two companies, they go that companies just lose money. He states that there is a cult-like worship of the wealthy where people think someone would not spend all this money for no reason.
Zitron argues that reconciling with the idea that the ultra wealthy and ultra powerful did not get there through big brains but through luck and opportunism and getting an MBA perhaps with the right people, and that they just happened to be in the right place at the right time, is kind of grim. He states that it is easy to say they are not making a mistake and you must be missing something, which is what they want.
Zitron addresses the argument about AI improvement rates, noting that if you imagine any rate of improvement in these AI tools at some point, and then imagine a 5% reduction in cost, which is what happened with the internet and cars, but Moore's Law is not with GPUs.
Zitron explains that Nvidia invented CUDA in the 2000s, which is the underlying software library and the way to run software on GPUs. He notes that it took them a solid decade or more to make it something where they could do data analytics, one of the early things being MapReduce and such. He states that when AI came along, they had lots of experience with it.
Nvidia makes the chips that power AI systems, and their CUDA platform enabled generative AI to grow. These specific chips are required for both training runs and inference operations in data centers.
Unlike the car industry, the AI sector has seen nearly every mathematician and scientist enter the field, with world governments continuously promoting the technology since 2023. The claim that AI represents inevitable progress is presented as fact, despite challenges in measuring actual productivity improvements.
Measuring AI productivity gains through simple metrics like code shipping ignores the contextual learning that human engineers gain from real-world problem-solving experiences. While LLMs have uses for encoding, the 5% improvement claims lack clear definitions—whether this means better performance in specific languages like Rust or C++ remains unspecified.
The standard of software across major tech companies including Google, Microsoft, Amazon, and Meta has deteriorated as LLM usage increases. GitHub reliability issues have become so common that users joke about needing notifications when the platform is up rather than when it's down. Microsoft's management of GitHub exemplifies broader quality problems across the industry.
Bloomberg Terminal's Ask B feature demonstrates how hallucinations manifest in financial contexts. When querying growth rates for Microsoft, Google, Meta, and Amazon stocks over five years, the system generated a false Microsoft stock price of $575. While minor errors like incorrect stock prices cause limited harm, hallucinations become dangerous in medical transcription tools or hedge fund financial models.
When refactoring codebases, AI systems can introduce security vulnerabilities or break existing functionality. The risk compounds when developers rely heavily on AI assistance without maintaining their own coding skills, creating multiplicative problems that become increasingly difficult to identify and resolve.
AI improvement is measured through benchmarks specifically adjusted for large language models rather than real-world task performance. The hallucination leaderboard shows rates dropping from 21.8% to 0.7% on simple summarization tasks over four years, but these metrics focus on basic tasks rather than complex applications. Task completion rates hover around 50% even in controlled demonstrations.
Early internet users experienced dial-up modem limitations, including inability to use phones simultaneously with internet connections. These technical difficulties mirror current AI limitations, suggesting that continued improvement trajectories could make current problems less prevalent over time.
AI should be compared to actual human alternatives rather than perfection. A human intern might hallucinate or have knowledge gaps, making AI's 0.7% hallucination rate potentially acceptable if the system demonstrates superior knowledge and speed. The comparison should focus on net value rather than absolute accuracy.
Human editors provide value through lived experience, emotional intelligence, and contextual understanding that AI systems cannot replicate. The relationship between humans involves mutual learning and shared experiences that go beyond simple output generation.
AI systems access stored files rather than possessing actual memory. While they can retrieve previously provided information like a dog's name, this differs fundamentally from human memory processes that integrate emotion, experience, and contextual understanding.
The distinction between human and AI memory processes matters less than output value. When asking for specific information like revenue numbers, the concern is whether the answer is correct rather than how the system arrived at it. However, this perspective overlooks the learning and understanding that humans gain through the process itself.
Working with human editors involves shared discovery and mutual learning that creates value beyond the final output. The collaborative process itself generates insights and understanding that cannot be replicated through AI systems that produce averages of existing documents.
Trust in human collaborators stems from shared learning experiences and demonstrated reliability over time. AI systems require extensive verification processes even when trusted, as the complexity of reports increases the likelihood of errors. The ability to have back-and-forth discussions with humans provides additional value in identifying missed information.
When AI industry representatives cannot provide direct answers about capabilities, they offer tangential responses. This pattern reflects broader issues with how the industry presents its achievements through benchmarks without clear definitions of what those metrics actually measure.
The original iPhone launch provided immediate, obvious value that required no explanation beyond basic demonstration. Unlike AI systems that require complex setups and explanations, the iPhone's benefits were immediately apparent to both technical and non-technical users.
ChatGPT reached 100 million active users within 60 days of launch, compared to TikTok's 9 months, Instagram's 2.5 years, and the World Wide Web's 7 years to reach equivalent scale. Over 60% of US adults integrated AI tools into daily routines within 3 years, compared to 5 years for the internet and 12 years for personal computers.
Some users report transformative business impacts from AI tools, particularly non-native English speakers who use the technology for copywriting and image generation. The question of honest pricing remains relevant—whether users would pay full per-million-token API costs if the technology were priced at actual operational expenses.
Unlike the development path to smartphones, which had clear roadmaps showing how smaller radios would enable mobile devices, AI lacks a defined trajectory toward promised capabilities. The technology was promoted as inevitable progress rather than presented honestly as expensive, experimental cloud software with uncertain outcomes.
Since 2022, AI companies have promoted their technology as revolutionary, claiming it would change everything, take jobs, and handle complex tasks. The gap between promised capabilities and actual performance creates unrealistic expectations and potential backlash when systems fail to deliver.
AI data centers require massive energy consumption, including gas turbines that impact local communities. The technology raises power bills and creates inflation in consumer electronics due to increased RAM demand.
The current AI hype cycle mirrors the dotcom bubble, where excessive investment and overselling preceded market correction. While 90% of dotcom companies failed, the period also produced lasting companies that transformed the economy.
Krueger, the award-winning economist, predicted that "the growth of the internet will slow drastically as it becomes apparent most people have nothing to say to each other." Clifford Stall's analysis correctly identified that an overwhelming amount of bad information online harms society and that online education would not replace regular education effectively.
The dot-com bubble consisted of two distinct components. The website bubble featured absurd valuations, such as Excite at Home purchasing a company for a billion dollars. The more significant element was dark fiber - the massive overbuild of fiber optic cable infrastructure based on analyst estimates that internet demand was doubling every 90 days, when actual demand grew every 6-12 months or longer. This created overbuilt transmission stations and connections that never achieved the anticipated utilization.
The critical difference between post-dot-com recovery and current generative AI infrastructure lies in demand characteristics. Current generative AI demand is predominantly subsidized, with most users not paying the real cost. The industry benefits from the largest, most disingenuous marketing campaign in history, driven by Microsoft as the apex predator of cloud software.
Microsoft generates only single-digit billions from AI software sales, while outside of OpenAI and Anthropic, the total reaches barely $22 billion - insignificant compared to the trillion-plus dollars invested. OpenAI and Anthropic have $1.1 trillion in cloud commitments. Data centers built today will cost as much to operate in 2050 without breakthroughs in electricity or GPU technology, despite efforts from Broadcom, Nvidia, AMD, and ARM.
Nvidia's Vera Rubin system promises "10x more efficient" performance but delivers "more dollars per megawatt," with companies remaining coy about actual cost reductions that would be the easiest claims to make if true.
The term "artificial intelligence" deliberately conflates unrelated technologies - protein folding, robotics, and autonomous weapons have nothing to do with LLMs. This linguistic strategy allows proponents to claim AI achievements in cancer research when discussing non-LLM systems. All data centers currently under construction serve generative AI exclusively, not the pre-generative AI applications like Google's search systems.
Traditional AI systems like Matic cleaning robots require minimal ongoing compute once trained, unlike generative AI systems that continuously burn resources. Sightline Climate reported 190 gigawatts of data centers in planning, representing $1.6-3 trillion in annual electricity demand, while current annual demand stands at only $130 billion.
The demand driving data center construction comes primarily from Amazon, Microsoft, and Google feeding money to OpenAI and Anthropic. These are the only substantial customers, creating a circular funding mechanism where the same three companies fund two unprofitable entities. Companies have raised $217 billion in 2026 alone based on speculative revenue potential from these two primary customers.
Google's decline stems from deliberate strategic choices rather than AI coding tools. In 2019, Google initiated a "code yellow" due to material weakness in query numbers. Pragar Ragavan, then head of ads, pushed to increase queries even if it meant providing worse answers, recognizing that faster answers would reduce search frequency and advertising revenue. Engineers including Ben Gomes and Shashi Tako objected that this would degrade the product.
After Pragar Ragavan took over Google search in early 2020, the company reduced suppression of spammy sites to increase user engagement time. When generative AI emerged, Google placed AI responses at the top of search results to prevent users from clicking away to external websites.
Google's instability extends across products: Google Docs and Google Sheets have become bugfests. Similar problems affect Microsoft, Amazon, and the broader tech industry. Amazon Web Services experienced multiple outages due to AI coding tools. GitHub downtime represents the clearest example of AI-assisted coding creating systemic problems.
The mechanism involves increased code volume from AI tools, with well-meaning developers using LLMs to contribute to open source projects despite limited understanding of the code they're shipping. This creates a Dunning-Kruger effect where partial code comprehension leads to deployment of inadequately understood systems.
OpenAI leadership has promised autonomous AI systems that will replace software engineers, while Anthropic's Dario Amodei claimed 50% of white collar labor would disappear within years. These predictions contrast sharply with OpenAI's own study showing no correlation between AI token spending and revenue per employee.
Uber CEO Dara Khosrowshahi predicts autonomous vehicles will eliminate drivers within years. However, Waymo's limited, highly controlled deployments reveal the gap between 95% functionality and handling edge cases - rain in San Francisco, children wearing high-visibility clothing, unexpected obstacles.
The appropriate comparison is autonomous vehicles versus human drivers, not versus perfection. Real-world observations include multiple Zoox vehicles blocking hotel exits in Las Vegas and Waymo vehicles stopping unexpectedly in San Francisco, creating cascading blockages.
Statistics showing 68% lower crash involvement, 55% reduction in police-reported crashes, and 80-81% reduction in injury crashes for autonomous vehicles must account for the vastly larger sample size of human drivers over many more years. Autonomous vehicle deployment requires extremely controlled environments with human oversight, which responsible companies are implementing.
Law firm associates - the actual workers doing research, drafting motions, and daily legal work - are not reporting productivity gains from AI tools. The enthusiasm comes from well-paid partners who have experimented with ChatGPT, not from those performing the substantive work. OpenAI research demonstrates no correlation between AI spending and revenue per employee, indicating the promised productivity revolution has not materialized.
Anthropic's second report acknowledges that hallucinations are mathematically guaranteed. The people experiencing work disruption are art directors, transcribers, and translators whose bosses don't care about output quality and view it as cheap work. These bosses would have automated jobs anyway by seeking the cheapest global option. AI is facilitating this cost-cutting, though these companies are only paying subscription fees rather than covering AI's actual costs.
There is no evidence of productivity gains in the white-collar labor force. If such gains existed, companies would be publicizing them. An Oxford Economics study claimed young people are finding fewer jobs because of AI, but multiple journalists failed to read beyond a single line stating "we saw some correlation" without providing any actual numbers.
Society is conditioned to believe the rich and powerful know what they're doing, leading people to internalize narratives about previous booms losing money and technology taking time to mature. AI companies intentionally exploit these mythologies knowing journalists, analysts, and investors will accept them. Stock price increases create the perception that AI must be working.
The fundamental deception is about how possible AI advancement actually is and what evidence companies possess. They lack signs of getting cheaper, autonomously doing work without complex setups, or reliably making customers more money in measurable ways. The most enthusiastic supporters are often "psychopaths on Twitter" with religious attachments to companies like Anthropic.
Bloomberg reported OpenAI on track to hit $40 billion in annualized revenue, though this calculation method remains undefined. Multiple people claimed it was actually $60 billion. The AI industry features unprecedented parasocial relationships with both large language models and companies, creating cult-like dynamics around growth and backing the "right horse."
AI doomers including Jeffrey Hinton warned that AI development is highly dangerous and fundamentally disruptive to society. AI CEOs previously echoed this danger narrative but have pivoted to claiming AI "isn't going to change anything" and "isn't dangerous at all" as they face criticism and regulation concerns.
Sam Altman and Demis Hassabis are described as deeply corrupt and cynical. Their early 2023 claims of being "a little bit scared" about their creations are dismissed as investment tactics designed to create urgency. Every scam begins with rushing people, and the advice given is to slow down unless facing literal mortal danger.
The pivot away from danger narratives stems from desperation to avoid regulation rather than public relations concerns. America remains trapped in a Milton Friedman, Margaret Thatcher, and Ronald Reagan neoliberal framework prioritizing growth at all costs. The appropriate response would have been breaking up these massive companies rather than allowing their current size.
While AI models pose genuine dangers through cyber attacks and vulnerability exploitation at scale, these risks differ from the extinction scenarios warned about by figures like Jeffrey Hinton. The companies themselves represent the "wrong hands" controlling these models, not external threats like Chinese acquisition.
The solution involves regulating compute usage rather than allowing unrestricted access. Companies have already let the "genie out of the bottle" through lack of oversight, with enablers permitting unlimited compute burning despite dire safety warnings.
The claim that the AI industry creates enormous economic growth lacks supporting data. Economics observed are primarily circular: Nvidia feeding money to companies like CoreWeave, or major AI companies feeding money to each other. Beyond semiconductor spending and data center infrastructure, AI-related spending barely reaches $100 billion, mostly covering operational services rather than genuine consumer demand.
The premise that the United States needs to spend trillions to beat China in an "AI race" is questioned. China already possesses Nvidia GPUs despite restrictions, and both countries have developed large language models. The race appears designed to justify spending rather than achieve meaningful technological superiority.
AI will not replace all human jobs, with no economic data supporting such claims. While some contract labor has been replaced through what amounts to digital globalization, this doesn't extend to most jobs. Robotics discussions are separate from generative AI capabilities.
Current robotics demonstrations, including Tesla's Optimus, require human control rather than autonomous operation. While robotics combined with AI could eventually replace certain jobs, the timeline extends to decades rather than years. The shift from software to robotics startups in San Francisco incubators reflects hardware costs dropping while intelligence costs decrease, though data remains expensive.
Agentic AI represents large language models communicating with other LLMs through additional frameworks, not autonomous systems. The terminology creates false impressions of independent action. Email triaging and similar tasks represent basic automation rather than revolutionary AI capabilities.
Modern business culture features leaders disconnected from actual work who pressure employees to use AI tools. These managers, sitting with AI systems that validate their ideas, demand others use the "genius machine" for increased productivity without understanding the actual work involved.
Unlike internet adoption, which required physical infrastructure and faced measured skepticism about information overload and social consequences, AI hype operates differently. Professional consequences exist for not claiming AI productivity gains, creating an environment where employees must "AI wash" their work regardless of actual usage.
In a world where everyone has access to AI tools, the capabilities that these tools can provide become commoditized. What the tools cannot do—human taste, judgment, and skills—becomes the scarce and valuable resource. Through history, the scarce and hard always becomes most valuable while the commoditized becomes least valuable. The very act of commoditizing content generation and code means that's not where value will accrue for users.
If an AI can create something objectively great, then that great thing is not valuable. The commodity tools produce commodity outcomes. Everyone having access to these tools hasn't fundamentally changed anything—it just creates a half-assery machine that amplifies existing poor practices.
People who were previously half-assing their jobs now have a machine to pump out low-quality work at scale. Carl Brown from Internet Bucks noted that AI makes the easy things easy but the hard things harder. While AI can generate small scripts effectively, complex tasks still require significant human intervention.
An example demonstrates the issue: using ChatGPT or Claude to create LinkedIn posts results in generic content because everyone else uses the same tools. A truly great LinkedIn post now comes from someone who doesn't use AI—creating something irreplaceably human, deeper, and more personal based on N-of-1 lived experience.
Ed Zitron uses AI for limited purposes. He uses AskB within Bloomberg Terminal to request consensus analyst estimates for Nvidia, but otherwise doesn't use generative AI. When he has tested AI for financial models, he found errors and wasn't particularly impressed.
The one use case he defends is technical support troubleshooting. He has successfully used AI to diagnose problems with Synergy software (which allows using the same mouse and keyboard across MacBook and PC laptop) by inputting troubleshooting logs. However, this capability doesn't justify trillion-dollar valuations or $2 trillion companies.
Google search has deteriorated significantly, requiring users to push through AI-generated content to find actual links and source material. Zitron finds himself using Bing sometimes despite hating to admit it. While AI can provide links, it doesn't do a particularly good job of this task.
AI companies are not selling their products as useful little tools—they're selling them as software that will change everything, replace all jobs, and transform society. Critics like Geoffrey Hinton and former OpenAI safety team members warn of future impacts, but they have vested interests in AI succeeding. Daniel Kokotajlo's "AI 2027" written with a former Codex engineer was described as badly written science fiction that he had to walk back.
These critics focus on hypothetical future dangers rather than addressing current harms: environmental damage, training on stolen work, unreliable answers, and the social problems of automated slop generation filling feeds with low-quality content.
AI has improved on tests specifically designed for the models and tests that companies can train for. The models show better performance on metrics where you can train directly for the test. Outside of OpenAI and Anthropic, and excluding coding startups, there are essentially no successful AI startup companies.
Agentic workflows involve automated processes like analyzing CEO backend data, summarizing information, searching the internet for context, reviewing past interviews, and generating reports. However, these aren't new capabilities—they've existed in various forms for years through web search and report generation tools.
While video generation has improved dramatically with tools like Seed Dance creating movie-like content, the practical application remains limited. Creating actual films requires complex coordination of first ADs, assistant directors, gaffers, and lighting specialists. Simulating light is particularly difficult, and the gap between creating a one-minute convincing clip and producing a complete movie remains vast.
Gary Marcus predicted diminishing returns in 2022. The current trajectory shows linear improvements in existing capabilities rather than breakthrough new features. Training AI to be more autonomous requires building structures around the models rather than emerging from training data alone—neurosymbolic approaches are needed.
The rate of improvement has already plateaued. While people claim their workflows have been transformed, the question remains whether they've actually paid for the tokens burned and whether the improvements justify the costs.
Exponential improvement in software performance results from direct hardware improvements. Despite nine to ten generations of Google TPUs and Broadcom building chips with OpenAI, companies cannot demonstrate paths to profitability. The environmental problems and profitability challenges remain unsolved.
Sam Altman recently claimed AI would become like a genie granting wishes within six months, but promises consistently fail to align with actual capabilities or improvement trajectories.
The massive data center investments are not being built to make consumer electronics smarter—they're constructed to speculate on capturing demand for generative AI services. Meta's recent earnings call revealed that running all social media posts through AI for context resulted in only 15 basis points (0.15%) increased retention on Instagram.
Zuckerberg referenced being able to train AI to serve specific users (like someone wearing a blue shirt drinking coffee) to improve retention, but spending $10+ billion for 0.15% improvement raises questions about ROI. Meta has spent $14 billion on Scale AI and Alexander Wang without disclosing specific revenue generated from these investments.
The current AI investment frenzy resembles previous speculative bubbles like the metaverse, NFTs, and crypto. The difference is that AI speculation involves major public companies with stocks to trade, fueled by post-COVID easy money and zero-interest rate policies.
Big Tech companies ran out of hypergrowth ideas, so they began purchasing GPUs. When they bought GPUs from Nvidia, markets rewarded them with hundreds of percent stock growth despite zero proof of AI revenue generation. The media attributed revenue growth to AI without companies disclosing actual AI revenues.
While there may be practical uses for the underlying technology, the speculation appears driven more by desperation than actual demand. Private credit is investing hundreds of billions not because of proven demand, but because they observed the largest companies building data centers and making money from feeding the ecosystem.
The three major companies—Microsoft, Google, and Amazon—are projected to spend over $400 billion on OpenAI and Anthropic alone in the next three and a half years, representing a large portion of their future growth expectations. If this spending doesn't materialize, their growth rates will slow dramatically.
The current AI progress is entirely a result of a circular funding system where venture capitalists, private credit firms, and Nvidia keep circulating money to AI companies. Without this continuous flow of capital, progress in large language models stops. The majority of funding that OpenAI received in the last six months came from SoftBank, Nvidia, and Amazon.
Comparisons to companies like Spotify, Uber, and Amazon Web Services that lost money for years miss the scale difference. Spotify didn't lose $20.9 billion in one year. Uber required $33 billion since inception before becoming marginally profitable. AWS between 2003 and 2015 cost $29.7 billion in capital expenditures normalized for inflation across the entire Amazon operation. AI companies like Google, Amazon, and Meta have added more than $700 billion in new property, plants, and equipment in the last four years.
AI proponents argue that the potential of intelligence permeates everything and the total addressable market is enormous compared to Amazon selling books. However, the argument that AI companies need to spend tens of billions of dollars annually on training to keep progressing requires perpetual capital infusion that cannot be sustained indefinitely.
OpenAI plans to spend $750 billion on compute through 2030 according to Wall Street Journal and Isaagi reporting. A large chunk of that is training costs. Training runs sometimes fail - GPT-5 had at least one training run that cost half a billion dollars and produced nothing. Oracle's future depends on OpenAI spending $300 billion over 5 years.
Sundar Pichai of Google stated the risk of underinvesting is dramatically greater than the risk of overinvesting. Andy Jassy of Amazon said they're not investing approximately $200 billion in capex in 2026 on a hunch and they're not being conservative. Mark Zuckerberg of Meta stated they'll continue investing aggressively in infrastructure, preferring to risk building capacity before it's needed rather than being late.
For Ed Zitron to change his position, there would need to be a hardware breakthrough that reduced costs by a factor of a thousand. It would also require dramatic improvements in data center construction practices. Current data center development is reckless and damaging to communities, with examples like violent protests in New Jersey where residents oppose projects but planning boards approve them. The use of gas turbines and behind-the-meter power is disgraceful.
Generative AI demonstrates how unfair the world is regarding capital access. Regular people trying to get business loans or mortgages face intense scrutiny, while AI companies building data centers receive backing from Jensen Huang with 25% residual value guarantees. Coreweave, a company building data centers and renting GPUs, received a $1.3 billion contract from Nvidia in 2023 to rent back GPUs from Coreweave, enabling them to secure bank financing based on that contract.
Claims that AI will be conscious or achieve superintelligence and artificial general intelligence are theories without proof. Stories about AI systems blackmailing and escaping control have been misrepresented. OpenAI's GPT-3.5 system card incident involved a user prompting the model to generate things to say to a TaskRabbit worker - it wasn't autonomous blackmail. Anthropic trained a model specifically to blackmail and then prompted it to do so. Media outlets covered these as examples of AI escaping control.
The AI fear narrative has backfired. Eric Schmidt was booed at a commencement speech by 8,000 people whenever he mentioned AI. Tech companies expected to receive floods of money and public submission through scary messaging, but instead face pushback. Sam Altman drives a $5 million car around San Francisco at 9 miles per hour. These executives are disconnected from regular people's problems.
Dario Amodei of Anthropic previously used scare tactics at OpenAI when GPT-2 came out, claiming it was too scary to release. He also stated on television that 50% of jobs would go away because of AI. While he appears more consistent than others and is now being attacked by Silicon Valley, his approach remains problematic.
For Zitron to admit being wrong by 2026, AI would need to achieve dramatically reduced production costs and demonstrate truly autonomous capabilities. It would need to continue capability improvements and become indistinguishable from magic. AI boosters should be regulated to prevent speaking in the future tense and be limited to discussing current capabilities.
The AI industry is in a bubble with downstream consequences for the broader economy. Companies are changing business models - Anthropic's Claude model now requires enterprises over 150 people to pay by the token or per million tokens. Uber's COO Andrew McDonald stated it's getting hard to justify spending on tokens because it's difficult to connect that spending to actual useful outcomes.
OpenAI was meant to go public this year but delayed to next year after releasing their financials. CFO Sarah Frier indicated they'll go public earlier than 2027. OpenAI's last funding round valued them at $865 billion, and they wanted a $1 trillion valuation for going public. Their advisor reportedly said not to pursue that valuation. OpenAI raised $122 billion this year and needs at least $100 billion annually to survive. Amazon sent them $35 billion contingent on going public early.
Anthropic is likely to go public before OpenAI. Anthropic has a better business growing faster than OpenAI's, though both face eventual limitations. The only way these models improve is through feeding tens of billions of dollars into training runs.
Private credit firms including Blackstone and BlackRock are investing in data centers because they understand the companies driving most of the data center demand.
Companies may run out of steam in 2027. If they attempt rushed IPOs, they will face worse economics than Anthropic and receive harsh market treatment. WeWork served as a cautionary example of another SoftBank investment failure.
OpenAI could collapse through multiple pathways. SoftBank, one of Japan's largest stock market companies, holds approximately $100 billion in OpenAI stock on paper. Without an IPO, SoftBank cannot liquidate these investments or use them as loan collateral, which would severely impact their ability to continue operations and pay stakeholders.
Amazon, Google, and Microsoft will need to restate guidance, admitting they won't grow as fast as previously projected.
A tech depression would occur because companies lack hypergrowth ideas despite market expectations. The manic AI GPU spending serves to maintain the narrative of continued innovation while their core businesses eventually plateau, with limited options for price increases, ad optimizations, or merchant fee adjustments.
The stock market's heavy dependence on these companies means any pullback would cause investors to stop investing and panic. Regular retail investors hold significant positions in these companies, with the Magnificent 7 representing 7-8% of the S&P 500.
Nvidia's circular financing model feeds companies money to raise debt for GPU purchases. Revenue could decline 50-70%, potentially returning Nvidia to single-digit billion dollar annual revenue levels seen in 2022.
Individual investors' retirements will contract severely, with losses unlikely to recover. The S&P 500 and Russell 1000 index values depend heavily on four companies plus the remaining Magnificent 7 members including Apple, Tesla, and Meta.
Over half of venture capital investments in the previous year went into AI. Most AI venture investments will reach zero value since LLM-based companies remain unprofitable. LLM companies have rarely been acquired, with notable exceptions including Cursor acquired by Elon Musk.
Cognition raised a $26 billion valuation, requiring an eventual public offering since few entities could acquire a company at that price point beyond Elon Musk.
The scenario describes equity values dropping without alternative homes for that capital. While theoretical bailouts could occur for OpenAI and Anthropic, the combined $1.1 trillion in commitments makes this unsustainable.
Oracle is building 7.1 gigawatts of data centers worth over $400 billion specifically for OpenAI. Oracle's revenue has remained flat for 15 years when adjusted for inflation, making OpenAI essential to their survival.
Tens of thousands of tech sector layoffs would occur alongside the venture capital implications. Venture capital has experienced historically poor returns since 2018, with total value to paid-in capital ratios between 0.8 and 1.2.
Venture capital celebrates paper gains rather than actual returns. Internal rate of return metrics remain unsatisfactory. Google boosted net profits by $99 billion on paper through increased SpaceX and Anthropic holdings.
Amazon, Google, Microsoft, and Meta valuations assume eternal growth. Without this assumption, markets would value them like airlines - established businesses without new growth prospects.
Cash positions are recommended over market investments. The market resembles a media-pumped casino. Investors should act conservatively during volatility, taking gains when available while remaining suspicious of tech promises and their manipulative metrics like annualized run rates.
Microsoft's $37-38 billion AI annualized run rate lacks definition and appears designed to manipulate perceptions. The absence of functional SEC oversight and skeptical media environments enables such practices.
Unlike the dot-com bubble where Amazon and Oracle survived, current companies are larger with bigger promises. Oracle could potentially fail entirely.
The reluctance to criticize wealthy, powerful companies stems from frustration with their ability to mislead without consequences. Average individuals cannot engage in equivalent levels of deception without repercussions, creating real economic and human costs.
Companies no longer produce quality products and treat customers with contempt rather than care.
Success should lift everyone around you rather than being purely economic. Community support during critical moments provides essential perspective and reminds individuals they are human. Simple gestures like expressing love and appreciation to others are undervalued but crucial.
Listeners and readers most commonly express feeling they have a voice and someone supporting them. Direct expressions of appreciation to artists, writers, and creators remain insufficient in current culture.
Keep The Diary Of A CEO in your library
Save the videos and channels worth coming back to, and find them again in one place.
