Intro
Ed Zitron goes through Nvidia's latest quarter with Isaac Pound — the handful of customers behind the revenue, the debt those customers need to keep buying, and the companies Nvidia funds so they can become its customers — then turns to Anthropic's reported market-size claim, the timing of its IPO, what he actually uses language models for, and why he thinks model progress is running out of training data rather than money.
Guest: Ed Zitron, writer of the newsletter Where's Your Ed At and host of the Better Offline podcast
Host: Isaac Pound
Published: 28 August 2026 on The Tech Report
Episode page | 47 min
Key Takeaways
Nvidia's revenue is real money from a very short list of people
"So 70% of their accounts receivable this quarter was from five customers."
"16% of their entire revenue was from one customer."
Each quarter now turns on whether Nvidia's customers can raise debt, not on Nvidia's chips
Amazon and Google are cash flow negative and have to keep borrowing; SpaceX sits a rung or two above junk
Nvidia is funding the companies that buy from it
Poolside: "Nvidia gave them $6 billion, hired most of their staff, gave them another billion dollars in investment"
"Nvidia is backstopping $105 billion of that if it gets built, but only if it gets built."
The financing is vendor financing with the label taken off
Morgan Stanley calls it balance sheet as a service; Zitron: "I think it's vendor financing by proxy."
Extended payment terms are buried in the filing and the beneficiaries are unnamed
"for some investment grade customers they are extending payment terms from 90 days to a year"
One disclosure rule would settle most of it
"If you have a customer over 10%, you have to name them unless it is national security."
The $30 trillion AI market figure is a number chosen, not calculated
"So the 30 trillion number is just make up a number."
"There is no value in a company saying $30 trillion total addressable market."
Anthropic's IPO timing is about which quarter's books it can avoid showing
"Q2 was really the peak of token maxing" — and annualizing token spend treats a spike as a subscription
The whole build-out fails the same arithmetic question
"So Anthropic just has to find another 29 trillion."
Nobody can articulate what separates one frontier model from another
"There is still no killer app."
The use cases he rates are small, and he says the industry priced them as if they were not
"That is not a 30, 40, 50 trillion TAM industry. It's not even a 30, 40, 50 billion dollar industry."
Progress ends where the training data ends, not where the money ends
"the only progress they make with video is good for 10 seconds because you need billions more hours of video than exist"
Nvidia Made Real Money, From Very Few People
Pound opened on Nvidia's results, noting the CFO put the backlog across major cloud providers at $2 trillion, and that Jensen Huang has said the industry is now in a golden age. Zitron's response was two words: "Good for him."
Zitron drew a hard line between the earnings and the announcements around them. The much-discussed $500 billion deal, he said, is not real — nothing happened, no money existed — but the quarter itself was real money, and he said so twice.
On the accounting, his point was concentration rather than impropriety. He walked the audience through accounts receivable — goods shipped and not yet paid for, an IOU that is official and entirely legal — and said Nvidia is doing the accounting correctly.
The concentration figures are the ones he wanted people to hold onto: "So 70% of their accounts receivable this quarter was from five customers."
"16% of their entire revenue was from one customer."
And "for the last 6 months, 44% of their revenue is from three customers" — against which he set the ordinary business preference for diversified income, so no single departure takes a chunk of revenue with it.
On pricing, he added a number he says came from his own reporting. "Nvidia has raised prices by 17%", and "I have actually heard from sources who deal with the resellers that it's more like 22 to 40% in some cases".
He noted that borrowing costs for Nvidia's customers are rising, pointing to CoreWeave's most recent debt as the example, and said the pattern is the familiar one: "Until something goes wrong, nothing is wrong." Everyone, he said, is thinking a quarter or two ahead while the number is going up.
Why 70% More Revenue Next Year Is the Hard Part
The guidance is the pressure point. Nvidia has said it will make 70% more revenue next year. Zitron laid out the arithmetic with an apology for the fiscal calendar: Nvidia's years run 1 February to 31 January, it is in fiscal 2027 now, consensus for that year is 396 billion, and 70% more is roughly 674 billion in fiscal 2028.
He named the likely five: Microsoft, Google, Amazon, Oracle and SpaceX — with the caveat that Nvidia does not specify, and one of them might be CoreWeave.
Two of those five are burning cash and have to keep borrowing. Amazon and Google are cash flow negative on his read, and have to raise debt now that chips, construction and debt itself are all more expensive than they were. SpaceX he placed a rung or two above junk, still investment grade.
The build side is slowing at the same time: data centers take a while, power is hard to get, and local authorities are pushing back. He listed Illinois and Arizona among states cutting tax breaks, with one more he could not recall, and said Pennsylvania and Texas are moving against data centers and stripping out the fast-tracking.
His summary of what that forces Nvidia into: "we're going to enter the Jensen verse" — circular financing tightening into a perfect circle because selling 70% more chips at higher prices into more expensive debt is, in his words, difficult rather than impossible.
The failure mode he expects is not a bad quarter. It is the moment someone cannot raise debt, with CoreWeave and Nebius named as the likeliest candidates.
"a normal healthy diverse industry does not act in this way" — and he pre-empted the rebuttal that this is simply what a new industrial revolution looks like: "people are overly reliant on history as a coping mechanism to the point that they will ignore reality".
The Neoclouds That Exist Because Nvidia Invested
Pound asked the question directly: "how much of this demand involves Nvidia promising money to its own customers" — while acknowledging an exact number is impossible.
Zitron said it depends on the customer, and started with Sharon AI, the first participant in what he called a new and nebulous Nvidia revenue-share program: Nvidia invests, helps the company raise debt, the data centers get built at some unspecified point, the GPUs rent out at a rate, and after interest, depreciation and amortization, Nvidia takes what is above.
The status of that program is contested on the record. CFO Colette Kress said on the earnings call it was very real; the Wall Street Journal reported the revenue-share deal was dead.
The scale mismatch is the whole point: on Sharon, "Their quarterly revenue is $1.8 million, but Nvidia just signed a $4.9 billion deal with them."
"there are just these bordering on fake companies that they're real" — real in the sense of having a bank account and a person running them, and a data center they intend to build.
On the second Australian neocloud: "Firmus, I was told by a source, was near death until Nvidia invested in them."
Poolside got a longer aside. The CEO said it was not an acqui-hire; Zitron's account is that "Nvidia gave them $6 billion, hired most of their staff, gave them another billion dollars in investment" and that all of those people will now work on Nvidia things, possibly in Nvidia's building. He was pointed about the label: stop calling it anything else.
Nvidia invested $30 billion in OpenAI, which is building a large data center in Ohio. "Nvidia is backstopping $105 billion of that if it gets built, but only if it gets built."
Vendor Financing by Proxy, and the Terms at the Bottom of the Filing
On SpaceX he called the arrangement the biggest of the circular deals — a transaction with Valor Equity Partners earlier in the year in which Nvidia invested $2 billion.
He explained why the money never travels in a straight line: telling a customer to spend the money you gave them back with you creates a legal mess. His precedent is Winstar, which sued what was left of Lucent Technologies and won, with the bankruptcy courts ripping the arrangement away.
So the structure is one of two moves. Either Nvidia makes an equity investment, which gives the banks collateral and cash to lend against, or Nvidia becomes the customer itself.
Morgan Stanley's term for this is balance sheet as a service. Zitron's is blunter: "I think it's vendor financing by proxy."
His analogy was a mortgage application. If your parents fund your deposit, the bank asks them to confirm in writing that the money is non-recourse. If all of your income came from people with your surname, the bank would say that is not distinct income. Route the same thing through Nvidia — a real business with good credit — and the banks say yes.
The disclosure he found most alarming was a note at the bottom. "for some investment grade customers they are extending payment terms from 90 days to a year" — and he stressed the wording says investment grade, not investment-grade credit, which he reads as wide enough to include SpaceX, Nebius or CoreWeave, both of the latter having raised investment-grade debt tied to Meta contracts.
He is clear the arrangement holds while Nvidia keeps generating cash. It breaks if a customer cannot raise debt — and on his estimate this set of companies wants roughly half a trillion dollars of debt next year, on top of everything else the world wants to borrow for AI chips.
What He Would Force Nvidia to Disclose
Pound asked what should be on Nvidia's earnings reports that is not. Zitron's answer was a single regulation: "If you have a customer over 10%, you have to name them unless it is national security."
The second half of the rule covers the payment terms: if you mention investment-grade clients getting three months to a year to pay, you have to name them too. Not what they bought, not when it ships — just who.
The reason names matter is that the credit quality behind them is wildly different. Ninety-day terms extended to Microsoft are nothing; Microsoft has the cash and has barely had to borrow through the AI build-out. Oracle worries him more. "If it's CoreWeave, I'm calling a cardiac surgeon."
He expects the AI boosters to read the same filing as vindication — investment grade, the keys jingled in the right way — and warned they had better be right, because "SpaceX does not have great credit." SpaceX loses billions, will need more debt, and its debt trades poorly.
On CoreWeave he flagged the uncertainty in his own numbers, saying he should look it up, but put the year's raises above 10 billion dollars, with billions more needed next year and every year after.
The framing he keeps returning to is that headline revenue tells you nothing on its own. A company earning roughly 96 billion dollars from hundreds of customers is a different company from one earning it from ten, or five, or three. Each quarter's success, on his account, is contingent not on Nvidia's ability to make great chips, be great salespeople or provide great service, but on the financing of the people buying.
Anthropic's $30 Trillion Market and the Case Against Long-Dated TAM
Pound raised the Wall Street Journal report that Anthropic sees a total addressable market of possibly $30 trillion, alongside a $2 trillion IPO target, and asked what assumptions get you there.
"So the 30 trillion number is just make up a number." Zitron's counter-suggestion was that they should have said a hundred, and Pound's follow-up — why go so low — set the tone for the rest of the segment.
He relayed Steve Eisman's reading of it. Eisman, the investor behind Steve Carell's Mark Baum in The Big Short, framed the figure as the midpoint between the entire US economy and SpaceX's own stated TAM. The stated basis is that all labor gets automated.
His regulatory answer is a time limit. "I don't think they should be able to say TAM further than two years." He dismissed the objection that this makes valuation impossible: "There is no value in a company saying $30 trillion total addressable market." It exists to mislead investors and add obfuscation.
On the raise itself: "they will be raising a hundred billion dollars which won't be enough to sustain them longer than like a year" — and he thinks it will be the last money out.
Why the IPO Has to Get Out Before the Q3 Books Close
Zitron's theory of the timing is that Anthropic wants September or October so it never has to close the books on Q3. "Q2 was really the peak of token maxing", and the slowdown that followed has not fully landed in the reported numbers.
The accounting objection is to annualizing a spike. Multiplying a period by 12 or 13 assumes recurrence: "if you're spending $1,000 on tokens in a week, that's not a recurring expense" — it is not a $200-a-month subscription.
He was careful about what he is not claiming: "I'm not saying people have stopped paying Anthropic. I'm saying that people are not paying Anthropic as much." He cited a Reuters story about customers lowering token spend by moving to open-source models.
On the first day, he allowed the trade could work. A 15, 20 or 25% pop is possible, and investors may look past the numbers to enjoy the fluff of AI. The long-term problem is that the company needs hundreds of billions more under public-market scrutiny.
He expects any debt to price as junk or high yield, on the grounds that this is an unprofitable company.
His specific prediction is an accounting one. He wonders whether Anthropic will try to capitalize training expenses, or run a non-GAAP presentation where you remove training and the margins turn positive: "Checkmate, atheists."
The last valuation, he said, was 900-and-something billion, and he questioned what has doubled to justify twice that — asking why a real company is quoting annualized run rate rather than revenue in dollars and cents.
On how the stock would trade: the same people who scream, cry and laugh at every piece of Anthropic news will be setting the price, so "trading it is going to be like well throwing knives and catching them at the same time".
Where Is the Money Going to Come From
Pound pushed the logic of the TAM claim further — if the basis is replacing all labor costs, why stop there, given there is new work humans could not do? He offered 60 trillion, 90 trillion, 100. Zitron's answer: why not.
The size of the number is not what he objects to. The question underneath it is where the money comes from, and there is not that much sitting unspent in the labor market waiting to move.
On the compute side he ran the sum out loud. Building the 190 gigawatts of capacity the industry has been discussing since February implies $1.5 to 3 trillion a year in compute costs.
Against that he set the actual cloud market: Microsoft Azure at a hundred-something billion, "Amazon Web Services buck 50, buck 70." and Google Cloud at 30, 40, 50 billion — most of it, he claimed, Anthropic. That totals 3, 4, 500 billion at most for the entire cloud compute industry.
The global software market he put at around 800 billion, generously a trillion. "So Anthropic just has to find another 29 trillion."
He worked through where that money would have to come from and found no candidates. Non-classical vendors, ordinary people, airlines and logistics companies — all businesses he notes are not high margin. On lawyers, his view is you are not replacing associates, though he joked some partners might be easier.
For scale he reached for the two biggest numbers in public finance. "The US national debt is now over $40 trillion." Bailing everyone out in the great financial crisis, counting the facilities on the side, ran to about 7 trillion, maybe 10.
Even a fraction of the promise breaks the comparison. Anthropic growing to three trillion a year of revenue would be roughly a hundred times Microsoft's, and the spending on compute has to grow alongside it.
His answer to the charge of cherry-picking was to point at the promises themselves. "They just agreed to rent $45 billion worth of chips from Nscale for a data center that doesn't exist." Set against $1.2 trillion of commitments, he said, nothing he says is as bonkers as what is being claimed.
Nobody Can Explain the Difference Between the Models
Pound brought in a conversation he had earlier in the week with Eli the Computer Guy about Anthropic's Fable 5 usage rates, which he said plateaued straight out of the gate, and asked whether there is a business case left for large frontier models.
Zitron credited the framing to Roger McNamee: "The problem with selling a general purpose tool is you have to come up with thousands of use cases."
That forces differentiation onto benchmarks, and a benchmark is a wasting asset — the moment one is beaten, the selling problem returns.
The substitute for an explanation is testimony. People on Twitter saying a model changed their life is not, on his account, a tangible difference between products, and the labs need one because they have to keep growing into $2 trillion valuations each.
The stack of harnesses and workarounds struck him as the opposite of a finished product: "a digital busy box".
He borrowed Robin Sloan's coinage, Jarvising, for the trap he sees users falling into — building systems about your own work, a wiki of everything you have written that tells you what you think, which feels productive and does not translate into a mass-market product.
The office version has a price ceiling he thinks people have not done the arithmetic on. A little report every morning that you are not sure you trust: "Would you pay five bucks a day for it?" Three? At ninety dollars a month, and with each additional tool stacking on top, the case gets harder the smaller the business.
At scale the return is unmeasurable — a thousand people with four minor use cases that do not sound essential when described out loud — because the thing people imagine buying is an autonomous intelligence that anticipates needs, and what they get needs re-prompting every time a model or a harness changes.
On the money already spent, his objection is proportional. Small automations would be fine if the industry had spent $30 billion; it has spent over a trillion and wants to spend trillions more. He also noted how rarely anyone will show him the workflow when he asks — the steamed hams scene from The Simpsons.
The selling levers left are hype, FOMO and executive incompetence, and he says those have a ceiling: you cannot tell someone they are an idiot for not spending a billion dollars.
His aside on the biggest single commitment: "Fun fact, Mark Zuckerberg's planning to spend 10 billion a year on Claude."
On pricing power he pointed at a recent cut — "They just discounted Soul by 20%." — and asked why they had to. "There is still no killer app." He noted "Clammy Sammy said it himself recently." that adoption is slower for exactly that reason.
The Use Cases He Actually Rates
Pound noted Zitron's free newsletter that week described his own use of LLMs to stay current on their capabilities, and that his Bloomberg terminal uses them, and asked what he considers valuable.
The distinction he draws is between outsourcing a task and outsourcing thinking. In a vacuum, LLMs are useful for minute, distinct tasks.
His example was fixing a bug in his kid's Minecraft game — the Wither Storm mod, or Cobblemon — and the loop it produced: install this, something breaks, it apologizes, fixes that, something else breaks, and thirty or forty minutes later it thinks it has fixed it and another thing is broken.
What it is doing is PowerShell and scripting work he says you could do yourself, only automated. "OpenAI spent tens of billions of dollars on training at this point, this is mediocre to me." It has improved over the year; he does not care.
He relayed Nik Suresh's reaction — a writer and software engineer — that this would have hit hard at 16. You can tell the computer to make things more efficient, understand nothing of what it does, and feel powerful, like a kid at an arcade playing a racing game without putting money in.
His own operational rule is a quarantine: "the only times I use these things are on a laptop disconnected from any personal information", because it adds and deletes at volume and he does not know what it is doing.
The one use case he endorsed without hedging: "dump a troubleshooting log into one of these things, any of them, and you'll get more of an idea than you would on your own".
The size of that, in his framing, is the whole argument: "That is not a 30, 40, 50 trillion TAM industry. It's not even a 30, 40, 50 billion dollar industry." It is computer automation.
On running models locally, he said it might be interesting, but he is not standing up a $5,000 Mac Mini to read a log from a Minecraft game. He pays the 20-dollar subscriptions to check in occasionally.
The Bloomberg example he gave is a genuine workflow improvement. Requesting data used to mean writing BQL, a Python-like query language; now generative AI writes the BQL from a plain request and runs it, and he takes the generated code and runs it himself to confirm the numbers match. Cited claims are clickable back to the source.
He does not accept that this scales into the case being made: "a lot of AI boosters conflate any progress with the progress that actually matters". Fixing a Minecraft problem is not the progress the spending requires; that would be automation you can trust.
His example of why he cannot trust it is a number it invented. Checking historic stock prices on Bloomberg, everything verified until one figure stopped him — "Microsoft stock's never been at 575 bucks." — and it had made the number up for a specific day he had a record of.
A Race to the Bottom, and the Training Data Running Out
Pound closed by putting Gary Marcus's comparison to him — that the broad AI business ends up like airlines, racing to the bottom on minimal profits — and asked whether that assumes ubiquity in the specialized use cases.
Zitron answered with Nik Suresh's blog post, AI is eviscerating global decision-making, which describes modern business culture as pressure to praise AI, and compares it to the opening of The Death of Stalin: everyone around the corpse saying how healthy he looks, watching each other, afraid of saying the wrong thing.
"You have executives who are saying they're 100 times more productive with AI" because customers and partners take offense otherwise, and nobody is actually feeling it.
He cited Suresh, a software consultant he described as gifted and successful, saying "he's seen 0% success of these AI initiatives" at scale — with the caveat that most software initiatives fail anyway and AI makes that worse.
On the destination he agreed with Marcus: "any future of this is going to be a race to the bottom commoditized and small industry".
The open variable is the one nobody outside the labs can see. "How much does it actually cost to run inference? We truly don't know." There are per-hour figures and cost-of-revenue presentations, but not the operating cost of standing up inference for a given customer base, or how much it varies with a guaranteed one. If it turns out to be messily profitable at low-10% gross margins, the surviving shape is an open-source GPU rental business making a couple hundred million a year at most.
He expects the improvement curve to flatten with the money: "I think we'll see the end of LLM progress", because progress so far has depended on Anthropic and OpenAI burning tens of billions, Chinese and open American models distilling the results, and billions spent on training data. What remains is small iterative tweaks — and the pressure campaign ending, so people use AI only when they need to.
On replicable use cases, he says the industry still has few. Research quality varies with the subject and with how much you ask for. Quick transcription he expects to stick around, with the same unknown attached: nobody knows what it costs to run.
Cal Newport's explanation is the one he finds most useful for predicting where the labs go next. Newport, a Georgetown computer science professor, told him the move into cybersecurity is the same move as code: enormous volumes of vulnerability data already exist online, the way code sits on Stack Overflow and GitHub. Images worked for the same reason.
Video is where the pattern shows its limit: "the only progress they make with video is good for 10 seconds because you need billions more hours of video than exist". He noted Adobe is paying for video, and expects the labs to keep hopping to whatever domain has data lying around.
On cybersecurity specifically, he questioned the economics rather than the capability: a tool that surfaces vulnerabilities with false positives is useful to someone who can spot a false positive, but he does not know what anyone will pay, and he had just read that someone built a harness that cut the cost by 80% — good for them, bad for the vendor.
He ran out of frontiers on air. Financial analysis has been tried and no hedge funds appear to have been replaced; prediction markets came up as a guess; he invited listeners to email him with a data-heavy field he has missed.
The same logic explains the rest of the map: 3D world generation off the volume of Minecraft and gameplay video on Twitch and YouTube, and generated writing that all sounds alike because all of the writing is from the internet. Music, he said, has hit a wall — there is only so much of it, and "music is an incredibly emotional thing that you cannot train the vagaries of human emotion".
His closing formulation: the limits on the models are what they always were, and the only thing without a limit has been the money spent forcing past them by buying data, creating data and paying data companies. "that's where the limitations will be when the bubble actually pops".
Zitron's bottom line is that Nvidia's numbers are real and largely beside the point: the quarter is underwritten by a handful of buyers who need ever more expensive debt, several of whom Nvidia funded into existence, and the trade holds only until one of them cannot borrow.
Products, Companies & Tools Mentioned
Nvidia (Real money, concentrated: "So 70% of their accounts receivable this quarter was from five customers." — and, Zitron says, the source of the financing that lets its customers buy)
CoreWeave and Nebius (The two he expects to break first; both have raised investment-grade debt tied to Meta contracts, and on CoreWeave, "If it's CoreWeave, I'm calling a cardiac surgeon."**)
Sharon AI and Firmus (Two Australian neoclouds Nvidia invested in; Sharon's "quarterly revenue is $1.8 million, but Nvidia just signed a $4.9 billion deal with them."**)
Poolside (Not an acqui-hire, per its CEO: "Nvidia gave them $6 billion, hired most of their staff, gave them another billion dollars in investment"**)
OpenAI (Took a $30 billion Nvidia investment; its Ohio data center has "Nvidia is backstopping $105 billion of that if it gets built, but only if it gets built."**)
Anthropic (The $30 trillion TAM and $2 trillion IPO target at the center of the second half; Zitron thinks the timing exists to avoid closing the Q3 books)
SpaceX (A likely top-five Nvidia customer and, via a $2 billion Valor Equity Partners deal, the biggest circular financing case; "SpaceX does not have great credit."**)
Microsoft, Amazon and Google (The three hyperscalers; Microsoft has the cash and has barely borrowed, while Amazon and Google are cash flow negative and must keep raising debt)
Microsoft Azure, AWS and Google Cloud (The revenue base he weighs the build-out against: Azure at a hundred-something billion, "Amazon Web Services buck 50, buck 70." and Google Cloud at 30, 40, 50 billion)
Oracle (The name that would worry him if it were the one getting extended payment terms)
Nscale ("They just agreed to rent $45 billion worth of chips from Nscale for a data center that doesn't exist.")
Morgan Stanley (Coined balance sheet as a service for the structure; Zitron prefers vendor financing by proxy)
Winstar and Lucent Technologies (His precedent for why routing your own money back to yourself ends badly — Winstar sued and the bankruptcy courts took it away)
Meta (Zuckerberg's spending is his single sharpest aside: "Fun fact, Mark Zuckerberg's planning to spend 10 billion a year on Claude."**)
Bloomberg terminal (The workflow he rates: generative AI now writes the BQL query he used to write himself, and he re-runs the generated code to check it — and it is also where he caught the model inventing a Microsoft stock price)
Minecraft mods, including the Wither Storm and Cobblemon (The bug-fixing loop that took thirty or forty minutes and kept breaking something else)
Suno (Generates music but has hit a wall, on his account, because there is only so much of it and emotion cannot be trained)
Adobe (Paying for video data, in a market he says needs billions more hours than exist)
Stack Overflow and GitHub (Why code was an early win — the training data was already sitting there, the same reason he gives for the move into cybersecurity)
Books & Resources Mentioned
Where's Your Ed At (Zitron's newsletter; the free edition that week covered his own use of LLMs, which prompted Pound's question)
Better Offline (Zitron's podcast)
AI is eviscerating global decision-making – Nik Suresh (The post behind the Death of Stalin comparison, and the source of his 0% figure on AI initiatives)
Jarvising – Robin Sloan (The coinage for building elaborate personal systems with an LLM that feel productive and do not scale into a product)
Get the latest market chatter and takes as they happen:
X | Threads | Instagram | YouTube | TikTok | Facebook