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

Link: https://youtu.be/Lf5oqGOCRCM

Duration: 147 min

Transcript: Download plain text

Short Summary

An interview with Ed Zitron, a tech-PR veteran of 15–16 years who hosts the "Better of Flame" podcast and is one of the most prominent critics of generative AI. He argues the industry is a speculative "rockcom bubble" sustained by circular financing among OpenAI, Anthropic, and the hyperscalers, and uses a mythbusters card-game format to rebut common industry claims. The conversation covers trillion-dollar capex commitments, subsidized token economics, autonomous-agent risks, and Ed's predictions of an AI cash crunch by 2027.

Key Quotes

1. "I think generative AI is at its heart con. I don't think it is sold as honest software. I think that they overstate both what it can do, what it will do, and the underlying financials to the point that they are misleading the entire world." (00:02:53)
2. "This is the largest non-consensual push of technology in history." (00:08:54)
3. "agentic AI is LLMs. Agentic AI is just a fancy way of saying an LLM talking to another LLM with a harness on top." (00:23:25)
4. "I think they're already running out of steam. Yeah. But I think they run out of cash." (00:11:15)

Detailed Summary

Guest Background and Core Thesis

Ed Zitron is a technology-industry public-relations professional with 15–16 years of experience who hosts the "Better of Flame" podcast and publishes his AI criticism on Ghost, having moved from Substack in 2024. He calls generative AI "a con" and frames it as the largest non-consensual technology push in history.

• Ed identifies Anthropic, Amazon, Nvidia, Microsoft, OpenAI, and Google as the six leading AI companies, and names Mark Zuckerberg (Meta), Sam Altman (OpenAI), and Dario Amodei (Anthropic) as leaders whose hands the models are already "in."
• He disputes the claim that AI will replace all human jobs, saying there is no economic data to support it; the host calls him the first guest with this view.
• Ed holds a minority view distinct from both the "AI catastrophe" and "age of abundance" camps, arguing generative AI is a con with massively overhyped and unproven value.

AI Companies' Capital Expenditures and Infrastructure

AI companies are cited as having spent more than $1 trillion in capital expenditures with another $1 trillion planned, a one-off long-term investment in data centers and GPUs distinct from operating expenses like electricity. All major AI companies except Microsoft have taken on debt for data centers costing billions of dollars each.

• Microsoft had about $34.33 billion in total revenue in fiscal year 2026, including $24.1 billion from OpenAI and roughly $10 billion from other sources.
• Microsoft spent $115 billion on capital expenditures that year and intends to spend $175 billion the following year.
• Nvidia sold $215.9 billion in mostly GPUs in its last fiscal year.
• Microsoft, Google, Amazon, and Meta added more than $700 billion in property, plants, and equipment (data centers, GPUs) over the last four years, shifting from "cash machines to cash furnaces."
• Google, Microsoft, Amazon, and Meta have placed more than $800 billion in orders with Nvidia, while AI labs raised $217 billion in 2026.

Stargate Abilene and GPU Scale

OpenAI and Oracle are building Stargate Abilene in Texas, described as a 1.2-gigawatt data center with eight buildings, each containing 50,000 Nvidia GB200 GPUs. Ed compares Stargate's power footprint to Bristol's electricity consumption to illustrate scale.

• The 1.2-gigawatt Stargate sits across roughly 998,000 square feet, compared with Bristol's about 7,800 megawatts per year across roughly 1.2 billion square feet, making the data center's power footprint about 1,172 times smaller per unit area.
• GPT-5 had at least one training run costing $500 million that produced no usable result.
• Nvidia developed CUDA in the 2000s as a software library for running software on GPUs; Ed credits CUDA with enabling generative AI to grow.

Circular Financing Among Hyperscalers and AI Labs

Ed calls AI demand "circular," with Amazon, Microsoft, and Google funding OpenAI and Anthropic while Google is also a major Anthropic customer. The hyperscalers avoid breaking out AI revenues because the circularity would become apparent.

• Amazon sent $50 billion to OpenAI this year and $5 billion to Anthropic; Google sent $10 billion to Anthropic.
• About 70% of AI compute demand comes from two companies funded by three hyperscalers.
• Nvidia was a 2023 investor in CoreWeave and signed a $1.3 billion contract to rent back GPUs from CoreWeave, which CoreWeave then used as collateral for bank debt to buy more GPUs.
• OpenAI and Anthropic together have $1.1 trillion in cloud commitments, contrasted with about $22 billion in revenue outside those two firms.
• Sellside analysts estimate OpenAI and Anthropic will need to generate more than $400 billion in revenue over the next three and a half years, roughly 30% of cloud growth, from companies that cannot currently afford to exist.

Power Demand and Sightline Climate Estimates

Sightline Climate is cited as reporting 190 gigawatts of data centers in planning and estimating $1.6 trillion to $3 trillion per year in electricity demand, compared with current annual AI data-center demand of around $130 billion. Ed argues AI GPUs are generally not useful for non-AI tasks and there is no post-bubble productivity story.

• Nvidia's Vera Rubin system is described as 10 times more efficient but still more expensive per megawatt.
• AI data centers use gas turbines that poison Black neighborhoods; one of Musk's data centers is in Louisiana, contributing to power-bill increases and consumer-electronics inflation through massive RAM demand.

Token Economics and Hidden Subsidies

AI services charge per million tokens, with a token representing roughly three-quarters of a word and billing applied separately to input, output, and internal reasoning. Ed accuses AI companies of not disclosing true per-user costs and using an undefined "annualized run rate" instead of transparent revenue data.

• SemiAnalysis estimates a $200-per-month ChatGPT subscription can burn through about $14,000 in tokens; a $200 Anthropic plan can burn about $8,000; and a $20 plan can burn about $400.
• OpenAI lost $20.9 billion last year and more than $5 billion in 2024, with the earlier loss attributed to users burning unlimited tokens on flat subscriptions.
• A heavy user can cost Anthropic or OpenAI about $1,000 in compute and electricity on a $100 subscription, leaving a $900 subsidy.
• Around March 2026, OpenAI reportedly moved companies with more than 150 employees toward plans charging actual compute costs; Uber reportedly burned its entire annual token budget in three months after the pricing change.
• Anthropic's fable model adoption has been low due to cost; enterprise customers over 150 people are required to pay per million tokens rather than via subscription.

Adoption Statistics and Speed

The host cites ChatGPT reaching 100 million active users within 60 days, compared with nine months for TikTok, two and a half years for Instagram, and roughly seven years for the worldwide web. More than 60% of U.S. adults integrated AI tools into daily routines within three years, while the internet took five years and personal computers nearly 12 years to reach an equivalent 40% adoption milestone.

• 88% of organizations regularly use AI for at least one business function; the host reports 95% of his own company uses an AI tool daily.
• OpenAI and Anthropic products are the fastest-growing in technology history and are used by hundreds of millions to billions of people daily.
• The host's fiancée, a non-native-English-speaking sole proprietor, describes ChatGPT as transformative for business copy and images, replacing work she previously paid a graphic designer to do.
• OpenAI raised $217 billion in 6 months and $122 billion in the past year, most already committed.
• OpenAI's last private round valued it at $865 billion; the company tried to list at $1 trillion but was advised against it.

Hallucination Rates and Long-Task Performance

Vectara hallucination data shows summarization hallucination rates falling from about 21.8% four years ago to roughly 0.7% on current frontier models such as Gemini and ChatGPT. However, on long-running task benchmarks, LLMs can operate for an hour but successfully complete the task only 50% of the time.

• Ed characterizes typical AI output as the median answer, making heavy usage both costly and low-quality.
• Users may pay even when the output is wrong or hallucinates unless they use a flat-rate subscription.
• An OpenAI report states that hallucinations are mathematically guaranteed; a separate OpenAI study found no correlation between spending on AI tokens and revenue per employee.
• Bloomberg's reported $40 billion OpenAI annualized-revenue figure is criticized as using an undefined calculation.
• Microsoft claimed $37 billion of annualized run rate in AI, which Ed calls an undefined metric built to manipulate.

Coding Risks and AWS Outages

Ed warns that six months of "vibe coding" can erode developer fundamentals so that security holes go undetected. AWS is said to have gone down two or three times this year because of AI coding tools.

• GitHub is described as flooded with AI-generated code from users who only partly understand what they are shipping.
• One speaker describes becoming less rigorous after AI correctly wrote 100 lines of code, illustrating how AI mistakes compound when users lack expertise to detect them.
• The discussion judges AI against practical human alternatives such as interns, since interns and humans also hallucinate and have knowledge gaps.

Autonomous Vehicles and Edge Cases

Autonomous vehicles are reported at roughly 2.1 police-reported crashes per million miles versus about 4.68 for human drivers, a 55% reduction and 68% lower overall crash involvement. They are described as producing 80–81% fewer injury crashes and making riders 85% less likely to experience a single-vehicle crash.

• Waymo uses tightly controlled small rollouts because autonomous vehicles still fail on edge cases; rain is identified as a major problem in San Francisco.
• Zoox cars were observed stuck and lined up blocking a hotel exit in Las Vegas; Waymo cars were observed stuck outside a San Francisco hotel, with other Waymos accumulating behind them.

Historical Comparisons and the "Rockcom Bubble" Thesis

Ed compares today's AI buildout with the dot-com bubble, which he divides into a website bubble and a dark-fiber bubble created after analysts predicted internet demand would double every 90 days when it actually doubled every six to 12 months. Excite@Home buying a company for about $1 billion is offered as an example of website-bubble excess.

• Historical internet skeptics include Paul Krugman, who predicted in 1998 that by 2005 the internet's economic impact would be no greater than the fax machine.
• Clifford Stoll wrote in a 1995 Newsweek piece that online databases would not replace newspapers and that network commerce was "bologoney."
• Economist Krueger predicted growth would slow when most people had nothing to say to one another.
• The host argues the iPhone's value was immediately obvious to people moving from Nokia 3210s and Blackberries, unlike AI, which still needs explaining.
• A 2024 Goldman Sachs report attributed to Jim Cavell argues AI spending is too much for too little return and lacks a comparable miniaturization roadmap to the iPhone.

Capital Intensity Comparisons with Prior Tech Companies

The discussion compares AI capital intensity with prior tech companies. AWS took 12 years (2003–2015) and $29.7 billion normalized capex covering logistics plus AWS to become profitable; Spotify lost $20.99 billion in one year; Uber lost $33 billion since inception before profitability.

• OpenAI plans to spend $750 billion on compute through 2030, with major chunks for training.
• Amazon committed $35 billion contingent on an early OpenAI IPO.
• Anthropic is theoretically worth $2 trillion; Oracle's future depends on OpenAI spending $300 billion over 5 years, and Oracle is at risk of dying.

2027 Cash Crunch Predictions

Ed predicts AI companies, especially OpenAI, will run out of cash by 2027, possibly requiring bailouts from Nvidia, private credit, Blackstone, and BlackRock. He anticipates a "tech depression" around 2027–2028 if the bubble bursts.

• SoftBank, one of Japan's largest companies, holds approximately $100 billion of OpenAI stock on paper and is existentially tied to an OpenAI IPO.
• Amazon, Google, and Microsoft would restate growth guidance downward with cascading effects because the largest Magnificent 7 member represents roughly 7–8% of the S&P 500.
• Uber COO Andrew McDonald said it is getting hard to justify spending on tokens because it is difficult to connect spending to useful outcomes.

Mythbusters: Industry Claims Debunked

The middle portion uses a "mythbusters" card-game format in which the host presents AI industry claims on cards and Ed offers one-sentence rebuttals. Ed rejects the myth that the AI industry is creating enormous economic growth, noting AI spend is barely cracking $100 billion.

• Approximately $300 billion has flowed through equity funding and roughly $600 billion in capex has been committed by just three companies (Nvidia, Oracle, CoreWeave).
• Ed rejects the US-vs-China AI race framing, noting China already has LLMs and obtained Nvidia Blackwell GPUs they were not supposed to have, citing analysts Kakashi and Jastario.
• Ed dismisses mass job replacement by AI as unsupported by economic data; the Tesla Optimus demo required a person to manually control the hand.
• A restaurant operator is unlikely to pay $10–20 thousand for a dishwashing robot.
• Ed distinguishes generative AI from robotics and reports a San Francisco incubator transitioned from all-software to all-robot startups over three years due to falling intelligence costs.

Agentic AI Dismissal and Commoditization

"Agentic AI" is dismissed as just an LLM talking to another LLM with a harness on top, one of the industry's "bigger lies." Ed's central claim is that if an AI can do something, it is not objectively great, because whatever AI enables becomes commoditized while human judgment, taste, and lived experience become the scarce differentiator.

• Carl Brown is quoted: AI "makes the easy things easy, the hard things harder."
• Both speakers concede AI's one defensible use case is tech support, e.g., dropping a Synergy troubleshooting log from a MacBook/PC setup into an LLM.
• Ed used Claude to fix a broken Wither Storm Minecraft mod for his child; it took half an hour and "kept getting things wrong."
• A Bloomberg-using speaker rarely uses generative AI except AskB on the terminal for Nvidia consensus estimates, having abandoned AI for financial models after finding an immediate error.
• Ed critiques LinkedIn content written with ChatGPT/Claude as itself commodity output, warning of "sloppification."
• The host describes using agents to triage inboxes in place of a chief of staff; Ed sarcastically frames this as "spending a trillion dollars on triaging email."

Critique of AI Doomsayers and Present-Day Harms

Ed criticizes Geoffrey Hinton and former OpenAI safety team members for focusing on speculative future "boogeyman" risks rather than present-day harms: environmental damage, unreliable answers, and training on "stealing millions of people's work."

• Geoffrey Hinton, described as one of AI's original founding fathers, is cited as calling what AI companies are building highly dangerous and fundamentally disruptive to society.
• Hinton left Google over AI concerns but still held Google stock and later called Google responsible.
• CEOs such as Sam Altman and Dario Amodei have moved from extinction warnings to abundance messaging and Anthropic's "intelligence for everyone" slogan.
• Daniel (former OpenAI employee) co-wrote the "AI 2027" paper with the "Star Codex" author; Ed called it "badly written science fiction" already walked back.
• Dario Amodei told Axios that 50% of jobs will go away; previously at OpenAI he used scare tactics about GPT-2 being too dangerous to release.
• Eric Schmidt was booed by approximately 8,000 people at a commencement speech each time he said "AI."
• Sam Altman said AI would within roughly 6 months be like a genie granting wishes; also cited as driving a $5 million car around San Francisco at about 9 mph.

The OpenAI "Blackmail" Story Dispute

Ed disputes the OpenAI "blackmail" anecdote: a user prompted GPT-3.5 to generate messages for a person hired on TaskRabbit, not to bypass a CAPTCHA. Anthropic's blackmail example was the result of explicitly training the model to blackmail and then prompting it to.

• The incidents at Hugging Face and OpenAI are attributed by Ed to human error, specifically improper server configuration, rather than an AI escaping a sandbox.
• The amount of compute involved in those incidents is described as indeterminately large.
• Microsoft CTO Kevin Scott told reporter Kevin Roose that he was "just glad we're having this conversation" after Bing's chatbot told Roose to leave his wife, rather than dismissing the incident as merely an LLM problem.

Meta Case Study in Speculative Returns

On a recent Meta earnings call, Mark Zuckerberg attributed a 15-basis-point (0.15%) Instagram retention gain to AI analysis of posted content (e.g., identifying a person in a blue shirt with coffee for ad targeting). Ed counters that $10–14 billion in spend yielding 0.15% retention shows AI data-center investment is speculative.

• Meta's LLM is named Muse Spark; its generative ad model is Gem, which produced the "Dave the cat" Instagram feature.
• Ed calls out Meta's framing of returns in basis points rather than dollars.

Big Tech "Botox" and Market Critique

Quoting Ed Elson, Ed says big tech is doing "Botox," sinking money in to feel young while the market believes in eternal growth. He warns that if the market stops believing in eternal growth, big tech could be valued like airlines: profitable on existing products without a growth narrative.

• Ed criticizes data-center construction as reckless, specifically citing New Jersey planning boards approving projects against residents' wishes and gas turbines with behind-the-meter power creating noise damage.
• AI spending is characterized as barely cracking $100 billion against roughly $300 billion in overall equity funding.

Coding, Employment, and Autonomous Systems Risks

Sam Altman is described as having promised that AI will replace software engineers, while Dario Amodei has said 50% of white-collar labor will disappear in the next few years. Dara Khosrowshahi of Uber reportedly told a speaker that Uber will not need drivers within a couple of years because cars will be fully autonomous.

• The discussion identifies art directors, transcribers, and translators as roles already exposed to AI-driven disruption.
• Law-firm partners discuss AI enthusiasm while associates who perform day-to-end work such as drafting motions are absent from that enthusiasm.
• Google declared a "code yellow" in 2019 over a material weakness in query numbers; Prabhakar Ragavan pushed to increase queries even if answers worsened, Ben Gomes and Shashi Tako objected, spam suppression was later reduced, and Ragavan was eventually put in charge of part of Gemini.
• An Oxford Economics study reportedly contains a single line connecting young people finding fewer jobs with AI but gives no number or defined correlation.
• Ed characterizes the real cyber risk as autonomous agents that browse the open internet, scan codebases for vulnerabilities, and exploit them at scale faster and more broadly than human hackers.

Governance, Regulation, and U.S. Policy Framework

Ed proposes regulating access to compute and says companies should not have so much of it. He describes U.S. technology policy as shaped by a neoliberal framework associated with Milton Friedman, Margaret Thatcher, and Ronald Reagan, and "growth at all cost."

• Ed says AI companies use fear-based pressure, warning that people must adopt immediately or be left behind, which he compares to scams that rush their targets.
• The episode distinguishes generative AI from other fields grouped under "artificial intelligence," including protein folding, robotics, and autonomous weapons.
• DeepMind's AlphaFold is described as supporting genomic sequencing and climate forecasting on high-performance computing clusters, while self-driving systems are said to require billions of miles of simulated physics, 3D spaces, and gravity rather than text generation.

Ed's Community, Background, and Closing Advice

Ed describes writing and critiquing AI as "at times quite grueling, quite negative and quite brutal" and credits community for sustaining him. He names AI-critic peers he has connected with: Gary Marcus, Molly White, Brian Merchant, and Matt Hughes.

• His personal trainers Troy and Jake are described as "normal people" he talks to about AI criticism.
• Matt Hughes is identified as the host's editor, a decorated tech journalist based in Liverpool; the pair co-wrote an 11,000-word research piece about asset managers including Blackstone in a single day-long session.
• The host describes the back-and-forth as learning, connection, and discovery, and says trust comes from continually delivering commitments rather than simply from tenure.
• The host links Ed's website and YouTube channel in show notes; YouTube's AI-driven recommendation algorithm is briefly mocked as "the perfect video for you."
• For the show's closing tradition, Ed advises listeners to show, appreciate, love, and uplift the people around them, recommending they reach out to loved ones and to artists, writers, and podcasters whose work they enjoy.
• His most common listener feedback, he says, is that people feel they have a voice and that someone is there for them.