Intro
AI skeptic Ed Zitron makes the full bear case to a show that has spent five years mostly booking bulls, working through Nvidia's first full-year guidance, Microsoft's AI revenue disclosure, the run-rate figures behind OpenAI and Anthropic, the data-center build-out and the private credit funding it.
Guest: Ed Zitron, founder and CEO of EZPR, host of the Better Offline podcast, and writer of the Where's Your Ed at newsletter
Hosts: Josh Brown and Michael Batnick
Published: 28 August 2026 on The Compound and Friends · 1 hr 27 min
Key Takeaways
The hyperscalers are now financially dependent on two unprofitable startups
Analyst expectations quoted by Zitron put "$440 billion of cloud revenue across Google, Amazon, and Microsoft coming just from OpenAI and Anthropic"
Microsoft's AI business is mostly one customer
"$34.33 billion of AI revenue. 24.1 billion of that is OpenAI" in fiscal 26, which he says leaves everything else a single-digit-billion business
AI data center demand is concentrated in one buyer
Zitron puts OpenAI at "like 60% of all AI data center demand" and says there is no second spender that size without venture capital
"ARR run rate" is the tell that nobody is defining the revenue
Token spend annualized is not recurring software revenue, and a customer can spend $1,000 one month and $50 the next
Nvidia's money is real; the source of the money is the question
Michael Batnick's chart put the quarter at "106% revenue growth at 75% gross margins" and trailing net income past Apple
The pale horses are all financing events, not product events
CoreWeave failing to roll debt, a thin hyperscaler bond sale, a major AI startup running out of money, or an IPO that bombs
Josh Brown pushed back from live experience, not theory
His firm runs AI across operations for 90-plus employees and 4,000 client households, and he calls Claude "miraculous"
Zitron is more worried about the debt than the equities
Private credit, insurance annuities and pension money sit behind the data centers, and he says he cannot size the damage
There is no dot-com-style salvage value in this build-out
AI GPUs are specialist, the electricity does not get cheaper, and an unfinished data center costs the same to finish later
He is no longer giving timelines because his 2024 call was wrong
"I was wrong because I was naive about how the world works"
How He Went Looking for AI's Revenue and Couldn't Find It
Zitron ran a PR firm until last year and started writing his newsletter in 2020, first about management theory and remote work, then crypto and Elon Musk, both of which left him "so thoroughly depressed" with the topics
The turn came when Sam Altman was fired from OpenAI and Zitron watched journalists treat him "like a rock star" on Twitter
He kept asking one question about the AI story and could not get an answer: "But I couldn't find any revenues."
Companies were spending billions on capex without disclosing AI revenue, and every earnings season produced articles saying the bet had paid off anyway
He says the economics are self-taught: "everything's self-taught, which to the chagrin of my many haters, is just like stuff I learned"
Zitron corrected Josh Brown on where the newsletter lives — it runs on Ghost, not Substack
Brown framed the booking as a first for the show: they have taped every Thursday since the summer of 2021, the majority of guests have been bullish, and while they have had broad bears like Jeremy Grantham, they had never had a specifically AI bear on
Brown's view of why the piece travels: professional investors will read both sides when they have money on the line, even when they disagree
The Projections He Says Bulls Are Taking at Face Value
Zitron splits the bulls in two. The ones who love the technology but concede there is a bubble, he respects — "at the very least, you're living in reality"
The other group takes OpenAI's own 2030 projection of $284 billion in revenue at face value. "Those people are not living in reality."
His arithmetic on that projection: OpenAI would become bigger than Meta in three and a half years while spending more than twice Meta's opex, and spending $200-something billion or more on compute in 2030
Batnick corrected the number mid-flow when Zitron reached for a larger one, which Zitron accepted
Zitron's account of how the trade got built: after a flat fiscal 23 for Nvidia and a rough year in tech, the big platforms raised prices, changed ad auctions and bought GPUs — and because none of them disclosed AI revenue, the market credited all the growth to AI
"The hyperscalers have now become financially dependent on the growth of OpenAI and Anthropic" — he cites analyst expectations from UBS, Barclays and Wells Fargo for "$440 billion of cloud revenue across Google, Amazon, and Microsoft coming just from OpenAI and Anthropic, two unprofitable startups who need to constantly raise money"
He says the two of them are taking up 90% of AI infrastructure, which he calls a distortion: "This is creating an illusory demand signal" — "There's a big Barry Bonds asterisk at the top of that"
Nvidia's First Full-Year Guidance and Its Handful of Customers
Brown put the bull case first: Nvidia had just given a full year of guidance for the first time rather than going quarter by quarter, pointing to roughly 70% revenue growth, with sales locked in for GPUs and Vera CPUs
Brown's own framing of the risk: about 90% of the business is selling to data centers and a huge chunk of that goes to the same four or five hyperscaler customers, even as Nvidia talks up automation and robotics
Zitron conceded the money: "Nvidia made a bunch of money. Like, I'm not questioning that. I'd be crazy."
His concentration numbers: "16% of their latest quarter revenue was one customer. 44% of their first half of fiscal 2027 was three customers. five customers make up 70% of their accounts."
Brown said those are good customers with millions of their own customers behind them — Fortune 500 businesses, governments, sovereigns. Zitron's reply: "Are they? Because they don't name them."
Zitron thinks the single largest customer was SpaceX, a company he notes has taken on a lot of debt
On the path to consensus estimates for fiscal 28, he says Nvidia has to take three to five customers much further while the price of debt rises, and notes Nvidia has just raised prices 17% with memory costs climbing, plus shortages of talent and electrical-grade steel
Microsoft's AI Revenue Is Mostly One Customer
The figure the rest of his argument hangs on: "fiscal 26, which just ended, $34.33 billion of AI revenue. 24.1 billion of that is OpenAI."
His conclusion from that split is that selling AI to everyone else is a single-digit-billion business, against more than $260 billion of capex
"Microsoft is the apex predator of software sales" with hundreds of thousands of resellers and tens of thousands of salespeople, and he says it still cannot scrape together more than single-digit billions for Copilot and the rest
GitHub Copilot, one of the few AI products he calls successful, moved to token-based billing, which he says killed the version of the business where a customer paid a flat fee
He says OpenAI is now material enough to Microsoft to be a liability rather than an asset — "It's actually a deeply load-bearing company" — and that 7% of Microsoft's fiscal year 26 revenue came from OpenAI
His sourcing for the demand claim is a Bloomberg news article he attributes to Brodie Ford, built on the earnings documents, rather than an opinion piece
"70% of their AI revenue being OpenAI is existentially bad." Zitron says he would rethink his position if AI were 30% of revenue instead
OpenAI's Funding Round and the Contracts Altman Keeps Signing
OpenAI's last round was $122 billion, of which Zitron says only $12 billion came from venture capital and the same asset managers funding data centers
He does not think that mix repeats, and notes Nvidia has said it would not invest again: "If they invest again, it's bad times ahead."
Since the Microsoft exclusivity arrangement ended, he counts the commitments stacking up: 138 billion over eight years with Amazon, 20 billion with Cerebras, 22.4 billion with CoreWeave, 300 billion-plus with Oracle
He adds a UBS note from Steven Drew putting OpenAI's spend at Google at about $12.5 billion a year, which he says has been barely reported
"Sam Altman has one real talent, and that's signing his name. My man loves signing contracts."
Zitron's structural complaint about how this gets analyzed: "the problem that everyone has right now is you can only rethink a quarter, two quarters in the future"
Where He Was Wrong in 2024, and What He Says the Market Still Misses
Batnick's chart showed Nvidia's trailing 12-month net income screaming past Apple's. Zitron's first question was how much of that is equity gains from stakes in Anthropic, OpenAI, CoreWeave, Nebius and IREN
Zitron volunteered the miss: "To be clear, I was wrong back in 2024." His explanation — "I was wrong because I was naive about how the world works" — is that he did not believe the companies would take it this far, or that the debt system would support it
What he thinks the market is not pricing: the centralization of data center demand. "OpenAI, I think, is like 60% of all AI data center demand."
Without venture capital dollars there is no second spender at that scale, and he says the market has not reacted because it has not happened yet
Brown's counter was that revenues are growing, they are real dollars, and both companies are talking about going public within roughly six months
Why He Calls "ARR Run Rate" a Scam
Brown put Anthropic at $60 billion of run-rate revenue. Zitron's answer: "No, they're not."
"This is the biggest scam of them all." ARR used to mean annual recurring revenue — a stack of signed contracts — and he says "run rate" has quietly replaced it without ever being defined
He says even Bloomberg's 65 billion story did not define it, and it could mean four weeks times 12 or four weeks times 13
The substance of the objection: both companies are annualizing token spend, which is not a subscription. A customer can spend $1,000 one month and $50 the next, especially if they move to open-source models
The period itself is unverifiable, so a company can pick a window that flatters a model launch or a feature release
"It is not a trustworthy measure of a business. And the fact it gets accepted is an insult to investors' intelligence."
Brown's rebuttal from earnings calls: every public company that talks to Wall Street about compute says it has already run out of its compute budget and needs more. Zitron's reply is that this is scarcity created by two companies absorbing most of the infrastructure
He put his own estimate of real compute demand at about $22 billion, against a Sightline Climate figure from February of 190 gigawatts in planning — "that's about $1.5 to $3 trillion a year in compute demand"
The Lenders Funding Data Centers Are What Frighten Him
Asked who is not seeing past their nose, Zitron answered lenders: "all of the investment in AI data centers terrifies me to my core"
He compares the scale to a city, then to the specific project: "Stargate Abilene, OpenAI Data Center, 1.2 gigawatts of power in 998,000 square feet" — "So they're condensing a city's power into a thousandth of the space, and that thing is way behind schedule"
Gigawatt data centers are a brand-new idea, and a gigawatt facility is really a campus of smaller buildings being dropped onto a grid he describes as held together with staples and tape
He rejects the argument that the build-out upgrades the grid — he says it strategically places power where it might be rented
On the diligence: he cites a report in the Information that Blue Owl agreed to invest in Stargate Abilene in about ten minutes. "That's how the due diligence is going, boys." Brown flatly refused to believe the real-world version of that even while accepting the article exists
Zitron's read on motive is not fraud but something worse for a lender: "It's nefariously ignorant. I think that their ignorance is so harmful." He calls the justification that the biggest companies would not spend a trillion dollars for no reason "gambler logic"
Nvidia's Numbers Are Real; the Money Behind Them Is the Question
Batnick's frame for the quarter: "I call Jensen Huang the chairman of the AI Federal Reserve" — the man who comes out three weeks after hyperscaler earnings and ratifies everything
The chart: $96 billion in revenue, with roughly 50% of a data center's cost going to chips. "That's 106% revenue growth at 75% gross margins." Zitron's only amendment was that the chip share is higher now because of memory
Batnick pinned down the disagreement — Zitron does not dispute the spending is real, he disputes where the capital funding it comes from — and Zitron agreed
"Every AI startup is subsidized. Every single one loses money." Until recently OpenAI and Anthropic let customers burn thousands of dollars of tokens for $200 a month
"These companies, when their customers are exposed to the real costs, they shrivel away" — his test being that if they could charge the real cost, they would
With everything now debt-funded and both the debt and the things it buys getting more expensive, he says the demands on the system compound
Is Corporate AI Demand Real, or Performance?
Zitron pointed at a blog by Nick Suresh, a tech consultant and software engineer who talks to CEOs, called AI is eviscerating global decision making
His analogy for enterprise AI claims is the opening of The Death of Stalin, with everyone gathered around the corpse agreeing how healthy he looks. "Most AI integrations fail."
Brown's counter came off earnings calls he listens to every season. Airbnb's Brian Chesky called AI the best thing that ever happened to the company and gave specifics — 45% of customer support calls closed by AI with no human involvement, lower expenses, better cash flow
Zitron's response was to note Chesky moved to open source early and is not paying the frontier labs, which is the part he is most bearish about
On AT&T, he says the Wall Street Journal's 80–90% savings headline dissolves on reading: the savings are in some functions, and the company has thousands of things plugged into AI
His broader claim, via Suresh, is that a lot of adoption is defensive — companies afraid of how it looks not to have an AI story
Brown agreed that piece of it is real, but held the line that the operational gains showing up in upside surprises will not reverse
What AI Actually Does Inside Josh Brown's Firm
Brown's position is that adoption slows rather than stops: "every week we're finding new things that we can plug into AI and do better than we did the week before"
The firm has more than 90 employees and 4,000 client households, with constant orchestration between Salesforce, portfolio accounting software and trading
His framing is operational, not customer-facing — clients do not care about the firm's AI use; the question is whether the firm runs better and faster
He hears the same story company by company on analyst Q&As: an upside surprise attributed to an AI investment made in 2023 or 2024
Zitron's response is that the durable question is not whether some of this works but whether the resulting revenue is enough to keep two unprofitable companies alive, or whether the work migrates to hosted open-source models
"Anthropic and OpenAI must keep growing, and they must keep growing so much faster. This is not even close to where they need to be. They need to be able to afford $440 billion in the next three and a half years."
Model Switching, Token Burn and Whether Gemini Changes the Math
Brown raised Alphabet as the exception, since Gemini does not need OpenAI or Anthropic to spend anything
Zitron's answer is that Gemini is just another large language model company and the differentiation between models is thin
"The actual functionality of this stuff is so malleable and volatile that I think it's exhausting to engage with on the regular" — every switch means new harnesses and new prompts
He says even pro-AI people describe an exhaustion at keeping up, and rationalize it "like any bad relationship"
The cost trap he describes: token prices may hold, but token consumption between models changes arbitrarily. "It's like having a car, the miles per gallon is the same, but you're driving 200 extra miles."
He can see a future with large language models in it, but not an obviously profitable one, even with custom silicon — possibly a very expensive service for certain businesses, which he thinks Gemini could become
Too Big to Fail, and Why He Doesn't Think a Bailout Fixes It
Brown's challenge: too much is riding on this, and bearish investors have learned repeatedly that the system is rigged to work out
Zitron's answer is that a bailout does not solve the arithmetic — rescuing OpenAI and Anthropic still leaves the $440 billion of committed cloud spend unaffordable
He argues the hyperscalers' underlying growth is slowing, "otherwise, they would have never done this"
Brown challenged that directly — cloud revenues are accelerating this quarter, all three came in better than expected — and Zitron conceded the wording, restating it as growth you cannot count on in the future
His evidence is Meta as a control: "You'll notice that AI is not making Meta grow massively." Oracle and Microsoft grew; Meta, which is not renting compute, did not
Brown's Meta bull case: they could reverse at any time, announce they are renting compute, and the stock would rise that day. Zitron agreed it would
On the downside case, he says even a quarter of the planned data centers leaves $600 to $700 billion of demand that has to exist and does not, because essentially every AI startup is unprofitable
Brown noted Oracle stock is down 70% and has caught none of the software bounce, as evidence the market already knows OpenAI is a potential problem
The IPO Window, and Whether Corporate Buyers Have Hit a Ceiling
Zitron thinks Anthropic can come public and that OpenAI has the harder path, then went further: "I've seen OpenAI's. I've seen their actual numbers. They're bad, bad." He cited "$13.07 billion of revenue" against a loss
"Anthropic is rushing this IPO" because of a two-quarter window when enterprise customers were moved off flat monthly pricing onto token billing — "Their revenue exploded" — and he says most companies are now cutting back
A record IPO would push the problem out rather than solve it, because cloud compute agreements require prepayments, so the money goes straight out the door
He thinks corporate buyers have already hit a price ceiling, citing Sam Altman conceding on stage that pricing is a huge issue for customers
Anthropic's most expensive model, Fable — the one marketed as having been banned by the US government for being too powerful, which Zitron says is not true — has petered out in market share, which matters because raising prices is the main way revenue grows from here
On the risk in Anthropic's biggest customers: a story that day had Meta talking about spending $10 billion, and Zitron's warning is that Mark Zuckerberg renamed the entire company for the metaverse and abandoned it within a few years — "this man does not care. He has complete control. He will stop doing something on a whim."
Clinkle, Clubhouse and the Media's Part in Every Bubble
Zitron's origin story, as Brown put it, is the publicist who saw too much: he represented technology companies and became disillusioned with how they treated users
He says he learned early that journalists do not like profitable companies, and he watched money go to companies with no business model — Clinkle, which was going to send payments using sonic waves and died within a year, and Color, a photo-sharing app that also died
Clubhouse is the one that radicalized him. He remembers people saying it would replace radio and that every company needed a Clubhouse strategy
Brown's own Clubhouse story is the best digression in the episode: the era coincided with the SPAC boom, he went on a show called What the SPAC, and told the audience the deals were going to zero. "in a prior career as a retail stockbroker, I sold 100 SPACs. So I knew they were all going to zero." He was never invited back, and says every SPAC involved went to zero
Then the metaverse, then crypto, then NFTs, each written about in newspapers as real, with virtual land being sold
His conclusion is that the media manufactures the conditions: "There's a reason now why bubbles keep happening. and it is the media." Brown pushed back that bubbles predate the media; Batnick offered the Fed
Zitron's refinement is that everything is a social contagion and social media accelerates it, and that AI was special because the 2022–23 tech downturn gave everyone something to sell — consulting, software features, infrastructure
Asked if he is impressed by the products, he answered: "Not for the amount of cost." He separates the technology from the infrastructure, environmental and social cost, and says what it does today is unremarkable against what it took to build
The Rot Economy
Zitron says his answer to whether he hates these people is yes, and he is specific about why rather than general
He says he has published documents showing Mark Zuckerberg requested 12% perpetual growth of all the numbers, and that the company spent years optimizing time spent in app without noticing engagement was falling because it was measuring the wrong thing
On Google, he lays the decline at the 2020 ads coup: Prabhakar Raghavan taking over from Ben Gomes, whose emails published in the antitrust trial warned that pushing more queries would make the user experience worse. "This is why Google sucks now."
Brown connected it to Cory Doctorow's "enshittification" and summarized the pattern as products that start out good, reach critical mass, and then get monetized to the bone
The distinction Zitron insists on: "it's not just they want to grow, it's that they have to grow perpetually, and they will do anything" — and having run out of other levers, the remaining one is price increases
He cites a Guardian report from around eight years ago on an experiment run on 700,000 users, manipulating the notifications and stories they saw, and says that should be illegal
His cultural explanation is the resumes: "Andy Jassy, MBA, Satya Nadella, MBA, Sundar Pichai, McKinsey, and MBA. Mark Zuckerberg, not MBA. Sheryl Sandberg, MBA."
Brown's pushback was that these companies have shareholders who expect earnings growth, and that management are not caretakers of a user population. Zitron's answer: "You mean provide a good service at a fair price?"
Brown made the most personal disclosure of the episode: "I hate Instagram. I love it so much. I'm addicted. I have to use a physical brick device to stop myself from using it." He drew his line at children rather than adults — a 12-year-old is different from a 42-year-old, and what a grown adult does with nine hours of scrolling is not his problem
On that week's settlement, Zitron said the payment is spread over ten years and put it at 12 billion and then maybe 17 billion, with Batnick asking whether that is about 1% of Meta's market cap. "Meta will enforce, the two-hour daily limit on Instagram and Facebook for children as part of a record claim." Both hosts called that part good news; Zitron's complaint is that it was not paid up front
Where He Does and Doesn't Find AI Useful
Brown declared himself on the other side of this one: "I find it to be miraculous, truthfully. I use it all day." He has built skills that replace a five-website copy-and-paste routine with a single Friday-morning instruction
What Brown will not do is have AI write his words, because the reason people read him is a specific voice and point of view; he uses it to check numbers and cite sources before publishing
Zitron's response was that this is search with extra steps, and Brown's own example proves it
The one thing Zitron uses: the Bloomberg terminal's Ask BQL feature, which generates and runs the query code for you so you can test it yourself. "It doesn't impress me for the cost." For a trillion dollars, he says, the product is a better search
His deeper objection is about what gets outsourced: "the distinction between when you are outsourcing work, like searching for something and thinking" — and he thinks far too many people are outsourcing the second one
Asked by Batnick for automated workflows outside the terminal, his answer was no. He tried to fix his kids' Minecraft game with it, watched it chase its tail for half an hour, and estimated the tokens would have cost him about $15
The GitHub Copilot change is his example of what this does to customers: two million customers, and on June 1 Microsoft pulled the pricing out from under them because they were burning $5,000 of tokens for $40
Zitron's answer to the defense that the cost would come down: "The cost of intelligence has come down, but the models spend more tokens."
He is willing to say the technology in a vacuum is interesting, and jokes that if they had been called "library models" nobody would have gotten excited — his objection is the trillion dollars, not the software
The Pale Horses He's Watching For
Brown described the parlor game version — Meta cuts its capex guidance, semiconductor stocks lose 30%, utilities fall like penny stocks, multiples compress — and asked what Zitron actually watches
Zitron agreed a capex pullback is the obvious one, then said: "I actually think there are a few different pale horses."
First, CoreWeave failing to raise debt. Its customers are OpenAI, Microsoft for OpenAI, Google for OpenAI, Anthropic and Meta; it buys a lot of GPUs, is deeply unprofitable, and he says it has raised billions three times this year at rates he recalls as over 9.2%. "So CoreWeave needs to perpetually raise debt. In fact, everyone does." He names IREN, Nebius and Nscale in the same position
Second, a hyperscaler bond sale that barely clears. "We just had an Amazon one that only got 1.6x oversubscribed" — if one comes in at 1.1 or 1.2, he expects stories about data center bond demand drying up. A Google sale afterward was four or five times oversubscribed, which he acknowledged cut against him
Third, a major AI startup going insolvent. He points at Perplexity, which he says Nvidia is considering investing billions into at a $30 billion valuation, because it is one of the few companies actually spending real money on GPU compute — "Nvidia is keeping them alive, because if they fail, well, everyone will notice"
Fourth, an IPO that bombs, particularly a SoftBank one. He says SoftBank's liquidity position is bad and it is having to raise 20 billion in bonds because it raised 40 billion. Brown called the SoftBank founder crazy and said he loves him; Zitron's verdict — "I love his goose math. His goose math rocks."
What he no longer expects is that ethics or social pressure changes anything: "it comes down to money will the money be available"
Oracle, the Ratings Agencies and Nvidia as Its Own Customer
Batnick asked where the mania actually is, given Nvidia trades at 18 times forward earnings and Meta at 16, which does not look like Teflon pricing. Zitron located it in the data center construction and the deal-making rather than in equity multiples — "it's an activity bubble as much as it is a financial bubble"
"Oracle's revenue has been flat for 15 years when you adjust for inflation." He says the stock should trade lower, and that Oracle is the one company here he is certain would be bailed out
He watches Oracle's own corporate account for tells, saying it posts reassurance whenever something goes wrong with OpenAI, and that you can take the opposite of what it says
His aside on Safra Catz: Oracle signed the massive OpenAI deal and about a week later she stepped back and left it to two CEOs
On the ratings agencies: "It's a notch above junk" and he does not believe they will downgrade, adding that they "rated CoreWeave debt investment grade because it was connected to a hyperscaler"
Brown noted the market is already treating Oracle differently from Alphabet and that the rating agencies are looking at it
On Nvidia investing in and lending to its own customers, Zitron's read: "It seems like it's the classic sign of illusory demand or artificial demand"
He flags disclosure in the 10-Q about investment-grade partners paying between three months and a year as unusual, and suggests the 70% guidance number itself may be a confidence play — "I think Nvidia may have said the 70% thing because they're desperate"
Railroads, GE Capital and Why This Isn't Dot-Com
Brown named the analogy that makes him most bearish, and it is his own: he reread a book called 1873 about the railroad build-out and its bust. "I don't think that this looks like dot com. And I don't think that this looks like the great financial crisis." The physicality of the demand is what makes railroads the closest fit, and he said rereading it made him want to sell everything
His second comparison is GE Capital: "Nvidia is acting a lot like Welch." Zitron supplied the phrase — vendor financing — and Brown extended it to GECAS, the aircraft leasing arm that kept airlines alive as America's largest equipment lessor
Zitron's addition is that the collateral is worse: "GPU compute is not really that useful outside of this", and the only buyers paying up for it are unprofitable startups
Brown pushed back with self-driving cars, automation and humanoid robots as other uses. Zitron's answer is scale — at $30 billion of build-out that argument works, but not at what he says requires $700 billion to a trillion a year of income
He notes Tesla, Volvo and Waymo built self-driving systems without this much compute, and that more GPUs have not produced an obvious breakthrough there
He is a fan of the technology he is dismissing as a demand source: he loves autonomous cars and getting into a Waymo, and worries about what happens to taxi drivers — "I wish we actually do some socialism and actually help people"
Brown offered the get-out: this could play out over 30 years rather than three. Zitron thinks decades, because the hard part of autonomy is not the 99% but "it's the 1% issue"
What Would Have to Happen for Him to Be Wrong
Brown asked the closing question his audience would shout — is there any chance he is wrong, and what concretely would change his mind. Zitron's condition was that the same question be put to the next AI bull on the show
"I don't have to be 100% right for things to be really bad." He can imagine OpenAI and Anthropic surviving as much smaller businesses inside Microsoft, Amazon, Google or a consortium
What he rules out is rescue by capability: "There is no breakthrough coming that's going to magically change this. The AGI story is dead." He notes the companies themselves are moving away from the term, and Brown added that nobody can define it
Even Nvidia hitting its number does not resolve it for him, because the same customers would have to take on far more debt to fund it — he names Amazon, Google, Microsoft, Meta and SpaceX, and notes Amazon says it is buying 2 million GPUs
He points out Meta and Amazon both guided lower than expected for the third quarter as early evidence of a slowdown, and that costs are rising even for the same volume of chips
Asked whether even profitability at the two labs would change his mind, his answer was whether they would be profitable enough to afford the compute they have already committed to. Batnick's verdict on the whole exchange: "I don't know how you could disprove that or prove that he's wrong."
No Skin in the Game, and a Market Run on Smoke Signals
Zitron does not have a position: "I don't have cash in the market. People go, oh, he doesn't have skin in the game. I have emotional skin in the game." Brown's reframe was that he carries reputational risk where others carry dollar risk
His reason for staying out is that he does not think fundamentals are what is being traded: "You invest based on reading the smoke signals."
His example is last September, when he counted three or four OpenAI announcements that moved SK Hynix, Samsung, AMD and Nvidia without anything actually happening. "I can't play in a market where companies do fake announcements."
He does talk to hedge fund managers occasionally, but says he never gets near the actual trades, which he says is how it should be
Where he draws the line between equities and credit: if you want to ride the crazy train in equities, he says that is not his expertise and go ahead — but on anything long term, especially the debt underlying a data center, he will explain why he thinks the demand is illusory
"There's not no demand, but it's a teeny tiny amount compared to how much we're building."
Why He Thinks There Is No Salvage Value
"There's nothing after this. Nothing. There's no dot-com bubble-style fixer here. AI GPUs are not useful for other stuff."
An unfinished data center costs just as much to finish in three years, the electricity to run the GPUs costs the same or more, and gas turbine power makes it worse depending on the price of gas
The dot-com comparison he says people get backwards: that bust left server hardware and dark fiber that were cheap to light up later. "Amazon's total capex was $29.7 billion adjusted for inflation" between 2003, when AWS was created, and 2015, when it became profitable — and that was all of Amazon's capex, not just AWS
His ask of listeners is to stop reaching for a historical template at all, and if they disagree with him, at least not to rely on a happy ending
What he thinks will actually shock people is not a startup failing: "The data center's not being used and the data center debt not being paid. That is going to be what shocks people."
Private Credit, and the Damage He Says He Cannot Size
Brown drew out the implication: unlike a Nasdaq drawdown that left the wider economy unscathed, the lenders here are banks and insurance companies, and "what you're describing does have the ability to literally take down the economy"
Zitron agreed and said the private part of private credit is exactly why the damage is hard to quantify — there is activity nobody can see
"I think a sixth of insurance annuities are private credit funded", alongside public and private pension funds, with Blackstone named as a lender funding it with insurance and private credit money
He credits Pablo Torre and Sam Copleman at Hunterbrook with illuminating the private credit story — the Mark Walter situation was on everyone's mind that week — and says the reporting still is not connected to data centers the way it should be
The point he wants connected: OpenAI and Anthropic do not have to die for this to break. "They could grow like five times as big. And they still won't have the demand for these things."
On the construction risk, Brown's line was the one that landed: "When have you heard of a building project that didn't run over budget?" Zitron's addition is that these are the most ambitious infrastructure projects of all time, underwritten by private credit, with tax breaks being cut in Arizona and Illinois and opposition in Texas and Pennsylvania
Nvidia's Receivables and the Limits of Accounting
Batnick's own question, saved for the end: how and when does Nvidia recognize revenue?
Zitron pointed at the accounts receivable line and asked why it and days sales outstanding are growing at the same time the company is making more money: "Why are your accounts receivable? Why your days sales outstanding growing?"
His interpretation is that it may show the edge of how quickly, or whether, Nvidia can get paid — and asks how much more accommodation customers will need to get the company to its next revenue level
Batnick made the accounting point that you can ship product and book revenue with no money passing, and that eventually you have to disclose it
Zitron's expectation is that the limits get tested: "They're going to push the absolute limits of accounting shenanigans here." He notes inventories growing as well
The circularity that he says proves the thesis: "Nvidia has $30 billion of cloud compute agreements, as in to rent back their GPUs. They have $25 billion of data center lease agreements." His question to Jensen Huang is why the company selling to data centers is also renting them
Why He Says He Doesn't Want to Be Right
Brown put it to him that he seems enthusiastic, and would be cackling if it all came down
Zitron reached for the line from The Big Short about unemployment and deaths, and drew the distinction himself: "I'm not going to, I might be satisfied. I'm not going to be happy."
His worry is who is exposed at the retail end — he says "we're at the highest point of retail leverage in history", and that this is what chills him
"Because we could have stopped this. We can't now." What is left, in his framing, is how large the system lets itself grow before it runs out of cash — not because the money does not exist, but because there are limits to how fast it can be mobilized
Brown closed by defending the booking: some listeners will say never have him on again, but if there is even a minor catastrophe, people who heard this will understand why it is happening. He noted the show has had Sam Copleman on, and Nick on the YouTube live stream
Zitron's bottom line is that none of this needs a technology failure to go wrong — OpenAI and Anthropic could keep growing and still never generate enough demand to pay for the data centers, the debt behind them, or the hyperscaler growth story that has been built on their spending.
Products, Companies & Tools Mentioned
Nvidia (The center of the episode: first full-year guidance, 106% revenue growth at 75% gross margins on Batnick's chart, and the customer concentration, receivables and vendor financing Zitron says undercut it)
OpenAI (The single spender Zitron says drives roughly 60% of AI data center demand, with $122 billion raised in its last round and commitments to Amazon, Cerebras, CoreWeave, Oracle and Google)
Anthropic (Rushing an IPO on the back of a two-quarter token-billing surge; Zitron disputes the $60 billion run-rate framing and says its most expensive model, Fable, has lost market share)
Microsoft and GitHub Copilot ($34.33 billion of AI revenue in fiscal 26 with $24.1 billion of it OpenAI; Copilot's switch to token billing on June 1 is Zitron's example of a rug pull on two million customers)
Oracle (Down 70% and catching no bounce; flat real revenue for 15 years, a notch above junk, and the one company Zitron is certain gets bailed out)
CoreWeave (The first pale horse — deeply unprofitable, dependent on perpetual debt raises at rates Zitron recalls as over 9.2%, and rated investment grade because of its hyperscaler ties)
Meta (Zitron's control case for the AI growth story, plus the week's settlement, a two-hour daily limit for children, and reported talks about spending $10 billion with Anthropic)
Alphabet, Google Cloud and Gemini (Brown's counterexample of an AI business that needs no OpenAI spend; Zitron says Gemini is just another LLM and Google's ads coup is why search got worse)
Amazon and AWS (Cloud growth Zitron attributes to two customers, a bond sale only 1.6x oversubscribed, 2 million GPUs on order, and the $29.7 billion of total inflation-adjusted capex behind AWS's first profitable decade)
Perplexity (The insolvency candidate; Zitron says Nvidia is weighing an investment at a $30 billion valuation because Perplexity is one of the few real GPU spenders)
SoftBank (A failed IPO here is Zitron's fourth pale horse; he says its liquidity position is bad and Brown and he both enjoy the founder's "goose math")
Blue Owl and Blackstone (The private credit side — Blue Owl's reported ten-minute decision on Stargate Abilene, Blackstone funding infrastructure with insurance money)
Airbnb (Brown's strongest counterexample: Brian Chesky's specific numbers on AI closing 45% of support calls, which Zitron countered by noting Chesky uses open-source models)
AT&T (Zitron's example of an AI savings headline that shrinks on reading — 80–90% in some functions, across thousands of integrations)
Claude and the Bloomberg terminal's Ask BQL (The two products the room actually uses — Brown calls Claude miraculous; the terminal's query generation is the only tool Zitron concedes)
SpaceX (Zitron's guess at Nvidia's unnamed 16% customer, and a company he notes is carrying a lot of debt)
IREN, Nebius and Nscale (The other neoclouds Zitron says have no underlying business to fund their debt)
Cerebras (One of OpenAI's post-exclusivity compute commitments, at 20 billion)
Waymo, Tesla and Volvo (Autonomous driving as the counterargument to GPU obsolescence, and the technology Zitron says he genuinely loves)
Salesforce (Part of the operational stack Brown's firm orchestrates with AI across 90-plus employees and 4,000 client households)
Clinkle, Color and Clubhouse (The pre-AI hype cycles that made Zitron a skeptic — sonic-wave payments, a photo app, and the audio product people said would replace radio)
Andreessen Horowitz (Zitron's example of manufacturing a hype cycle by assembling the right people on Clubhouse)
Books & Resources Mentioned
Where's Your Ed at (Zitron's newsletter, launched 2020 and published on Ghost, home of the 2023 rot economy essay)
Better Offline (Zitron's Webby Award-winning podcast with iHeartRadio and Cool Zone Media)
AI is eviscerating global decision making – Nick Suresh (The blog Zitron cites for the argument that enterprise AI adoption is often defensive rather than useful)
1873 (The book on the railroad build-out and bust that Brown reread, and calls the best analog to the data center build-out — "I should not have re-read that book because it made me want to sell everything")
The Death of Stalin (Zitron's analogy for executives agreeing that AI integrations look healthy)
The Big Short (The line Zitron reaches for to explain why being right would not make him happy)
Hunterbrook, and reporting by Sam Copleman and Pablo Torre (The private credit reporting Zitron credits, and says is still not connected to data centers)
Bloomberg's reporting on Microsoft's AI revenue, attributed to Brodie Ford (Zitron's source for the claim that Microsoft lacks native AI demand, based on earnings documents rather than opinion)
Steven Drew's UBS analyst note (The source for OpenAI spending about $12.5 billion a year at Google)
Sightline Climate's February data center report (190 gigawatts in planning, which Zitron converts to $1.5 to $3 trillion a year of compute demand that would need to exist)
The Information's report on Stargate Abilene (The source for Blue Owl's reported ten-minute investment decision)
The Guardian's report on Facebook's notification experiment (Around 700,000 users, which Zitron says should have been illegal)
Cory Doctorow on enshittification (The parallel Brown draws to Zitron's rot economy)
Get the latest market chatter and takes as they happen:
X | Threads | Instagram | YouTube | TikTok | Facebook