Ed Zitron says OpenAI and Anthropic have signed $1.3 trillion of compute commitments while losing tens of billions of dollars a year, and that bankers are now pushing rating agencies to call them investment grade.
The usual answer to that is that Uber and Amazon Web Services lost money for years too. Zitron's answer is that AWS spent $29.7 billion of capex between launching in 2003 and turning profitable in 2015, which is not the same order of magnitude as what the AI labs have already promised to pay.
"Anyone who tells you not to worry about Anthropic, a company that loses billions of dollars a year, if not tens of billions of dollars a year, and has made $517 billion in compute commitments, is a con artist."
Zitron writes the technology newsletter Where's Your Ed At and has spent long enough making this case in public that, as he puts it, "I'm considered a firebrand for asking these questions."
I listened to the full episode so you can skip it. 17 minutes of audio, 13 minutes of reading.
Here are the 10 takeaways that matter.
👤 Speaker: Ed Zitron, who writes the technology newsletter Where's Your Ed At and hosts Better Offline, a weekly show about the finances and claims of the technology industry
📰 Published: 9 September 2026 on Apple Podcasts
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 17 min
Key Takeaways
Crises get managed so the system survives, not fixed so the problem goes away
His frame is Adam Curtis's "Hypernormalization" — a simpler invented world, kept because it reassures
After 2008, over $14 trillion went to institutions and about $46 billion to homeowners
He reads a 2009 congressional panel finding no evidence TARP was used to prevent foreclosures
Every defense of the AI industry is the same systemic apologia, just reworded
It's early, it's like the internet, they're the smartest people, trust the process
The anti-AI conspiracy theories make the same mistake as the boosters
Believing the spending hides a secret surveillance plan assumes there is a plan at all
Bankers want OpenAI and Anthropic rated investment grade so they can borrow more
One rung above junk means expensive debt, on companies with no profits to service it
OpenAI is expected to spend $750 billion through 2030, against Meta's roughly $141 billion a year of operating expenses
80% of the two labs' enterprise revenue comes from the top 1% of their customers
Mostly unprofitable AI startups subsidizing their own users' token spend
The compute deals are take-or-pay, so the bill lands whether or not the demand shows up
The whole structure is a few hundred companies paying two companies to justify five companies paying one company
He expects a correction that dwarfs the dot-com bubble, and says nothing breaks until one payment is missed
1. Managing, Not Fixing
Zitron opened by explaining that this was monologue week because a guest pulled out at the last minute, then set out the frame from his newsletter piece Hyperscale Normalization, named after Adam Curtis's documentary Hypernormalization.
He read Curtis's thesis out directly: over the past 40 years, politicians, financiers and technological utopians declined to face the real complexities of the world and retreated instead, building "a simpler version of the world" to hang on to power, and the rest of the world went along with it because the simplicity was reassuring.
His compression of a two-and-a-half-hour film is that societies postpone problems instead of removing their causes. "That we as a society manage problems rather than dealing with the messy and complex elements of the world around us"
Problems deferred that way do not disappear, he said — they compound
The pattern holds because maintaining the system is easier than having a vision: feeding it money, cleaning up its messes and explaining away its successes
Thinking of an alternative becomes near impossible, he argued, because of "the dogma and confidence from those in power" — even as crisis after crisis shows the system is not functioning
2. The 2008 Bailout Choice
His first worked example was the subprime mortgage crisis. Dodd-Frank and what he called the now-weakened Volcker Rule curbed some excesses, but he said almost nobody involved lost money, a job or a reputation, and that he cannot find a single figure in the media of that period who was ousted or seriously criticized for how badly they got it wrong.
The split of the money is his evidence for what was actually being rescued. "While financial institutions and companies saw over $14 trillion in bailouts, only $46 billion, about 10% of the TARP plan, focused on trying to save homeowners from foreclosure"
He quoted a 2009 congressional panel finding "there was no evidence that the Treasury has used TARP funds to support the housing market by avoiding preventable foreclosures"
He described the outcome as a deliberate choice of target. "Government funds were used to manage a systemic crisis by making sure the system survived, rather than building a better society or fixing the system"
The rest was managed by cutting rates to zero and bailing out banks, automakers and anything systemic enough to pose a risk to the country — "rather than the people in the societies"
3. COVID Ran the Same Play
The pandemic, Zitron said, was handled the same way: "COVID was managed by the loosest lockdown protocols possible, giving businesses hundreds of billions of dollars via the Paycheck Protection Program." He noted the program covered salaried employees rather than contractors, and that banks received trillions through the same channels used in 2008.
The problems the pandemic exposed — wealth inequality, racial injustice, discrimination — were pushed aside rather than addressed
On the corporate response: "Companies picked up DEI as long as necessary to appease markets that believed there was actual social change without actually ingesting it"
He listed the structural changes that did not happen: Medicare for All, a social safety net, making price gouging illegal
The net result he identifies is the 2021 flood of cheap money — trillions of dollars of capital that "made the rich richer and the poor poorer"
4. Greed, Not Supply Chains
Zitron's account of the inflation that followed puts the cause with companies rather than with supply chains: "Meanwhile, corporations have been using supply chain crises and the specter of inflation to raise prices across the board, though it became obvious that the real reason was sheer, unabated greed, with companies posting record profits to soak up the cash from a frothy society excited to be back in the real world."
He named the policy response and its cost. "By the end of 2022, the Federal Reserve would raise interest rates by a dramatic 4.25%, leading to tens of thousands of people losing their jobs in the tech sector alone"
His view is that things have never really recovered since
The mismatch he keeps returning to is between what caused the price rises and who paid for stopping them. "Regular people have spent years watching the price of goods increase due to inflation, even after those interest rate bumps, despite the fact that the increase in pricing was mostly driven by, get this, corporations raising prices"
5. The Media Sneered Back
By "the system" Zitron means the conjoined forces of the media, the markets and the American government, and he spent this stretch on the first of the three. Parts of the legacy media, he said, spent years chiding readers for disbelieving that the economy was good, contorting to prove prices were not higher, and going against their own reporting in the name of balance, even as companies like Pepsi boasted about raising prices.
His description of what that did to an ordinary reader: "A regular person experiencing aggressively worsening standards in their own lives would open the news only to be sneered at for feeling bad, mocked for questioning the numbers, told to sit down and shut up because the media and those that informed them knew better"
When people took advantage of remote work, job flexibility and purchasing power, he said, they were told they were lazy, that it was temporary, and that they expected too much
Blame, in his account, is routed anywhere except the system — usually to a foreigner, or to the reader
Criticizing the system itself is treated as either stupid, because it always works, or hopeless, because it is too powerful
6. The AI Apologia Script
The turn to AI is the point of the whole argument: think about how people defend the AI industry, Zitron said, and the pattern becomes recognizable.
The loss-making comparison is the first move. "If you talk about how OpenAI and Anthropic are losing tens of billions of dollars a year, you'll get told that Uber and Amazon Web Services lost a lot of money"
His counter is scale. Both burned roughly $30 billion each: "And in Amazon Web Services' case, it was $29.7 billion of total capex between the year it started in 2003 and the year it became profitable in 2015"
On reliability, the answer is always the calendar. "If you say that generative AI's hallucinations make massive mistakes that make LLMs inherently untrustworthy, everyone will tell you it's the early days"
On usefulness, a person who cannot find a use case is told to use it more — a product described as simultaneously stupid and super smart. "You see, it's smart enough to do stuff if you make it do it, but not smart enough to actually help you in the way that we associate with software"
He groups the responses into one thing. "If you critique basically anything about generative AI, you're given a shot of systemic apologia" — it's early, it's like the early days of the internet, these are the smartest people and the largest companies in the world, so trust the process
7. There Is No Secret Plan
Zitron then turned the same criticism on people who agree with him. He said he regularly hears that the multi-trillion-dollar data center build-out is secretly a government surveillance program, and that this is the same systemic thinking wearing different clothes.
The conspiracy theory smuggles in the assumption that the people running the system are competent. "The smart people running the system have a big secret plan. They don't"
He insists on the point because the alternative is uncomfortable: without a plan behind the spending, "the real world seems kind of insane"
His counter-history is that failures come from an absence of planning, not the presence of one. "Yet every systemic failure has come from a lack of any real plan, strategy, or forethought"
On 2008: everyone assumed the institutions had hedged correctly, that the investments were high quality, and that anyone with an adjustable-rate mortgage could refinance or sell an ever-appreciating house
He calls those assumptions mantra after mantra, built on "a lack of curiosity" and complete faith in systems that had already failed once
8. Investment Grade as Rescue
The trigger for the episode was a report published the previous day: "Yesterday, the Financial Times reported that bankers for OpenAI and Anthropic's IPOs were seeking investment-grade ratings for both companies, pushing agencies to rate two unprofitable, unsustainable companies as if they were, well, real companies."
He reads the stated rationale as a euphemism. The FT's framing was that public listings would unlock liquidity and improve balance-sheet health; Zitron translates that as access to a lot more debt
His verdict on it is blunt. "This is a deeply dangerous way of managing two companies that will eventually run out of money"
The scale of what the rating would support: "Anthropic and OpenAI will need hundreds of billions of dollars a year to meet their $1.3 trillion of compute commitments"
Even a successful rating leaves the debt expensive. "They're going to need so much debt. So much debt with onerously high interest, because there'll be one rung above junk" — which, absent sudden profitability, he does not see them servicing
He broke off to address S&P and Moody's by name, in language that cannot be printed here, telling them they would deserve everything that followed if they granted the rating — then said he expects them to grant it anyway
9. Where the Revenue Comes From
Against those commitments Zitron set two comparisons and one revenue fact. "And I must be clear, their compute commitments, OpenAI is expected to spend $750 billion through 2030. How do they afford that? They lose tens of billions of dollars a year."
The benchmark he chose is what the largest technology companies spend to run themselves. "An unbelievable sum that dwarfs the operating expenses of Meta at around $141 billion a year, and Microsoft at about $175 billion a year"
On the projections: "I'm pretty sure Anthropic and OpenAI is going to spend more than those two numbers combined at some point" — and he clarified that by spend he means intend to spend
The revenue base underneath it is narrow and, he argues, circular. "Both companies are seeing growth decelerate and 80% of their enterprise revenues per Ramp, as I discussed last week, come from the top 1% of their customers who are predominantly unprofitable AI startups that are subsidizing their users' AI token spend"
The mechanism: for every dollar a user pays in subscriptions, the startup burns several dollars in tokens, and that money flows back to Anthropic and OpenAI — who are funded by the same venture capital their customers need
A rating changes the timing, not the arithmetic. "Any attempts to mark OpenAI or Anthropic as investment-grade will simply delay the inevitable point at which they run out of money"
10. A Death Cult for Growth
The commitments are take-or-pay, which is what removes the usual escape route: "Said commitments are also take-or-pay agreements that mean that regardless of whether they have the demand to actually need the data centers, Anthropic and OpenAI will have to pay Amazon, Google, Microsoft, Cerebras, CoreWeave, or any number of other providers once the capacity comes online."
His description of the failure mode is that nothing looks wrong until it suddenly is. "However you may feel about me or the greater AI bubble is immaterial to the fact that everything will seem like it's fine right up until somebody can't raise money and make a payment to either a neocloud, hyperscaler, or AI lab"
The chain he lays out runs from AI startups raising to pay the labs, to the labs raising to pay the hyperscalers, to the hyperscalers borrowing to buy GPUs from Nvidia and TPUs from Broadcom
He reduced it to one sentence. "In other words, the AI bubble is based on the whims of maybe a few hundred companies spending money on two companies to justify five companies spending money with one company"
On the people telling readers not to worry, he was categorical. "And anyone who prints a quote saying that you shouldn't have to think about it too hard, or you shouldn't have to worry about it, or that this is just like the dot-com bubble, really doesn't give a shit about whether you live or die"
His characterization of the industry: "A death cult obsessed with growth, empowered by a media ecosystem obsessed with measuring and celebrating how much it's growing and what might grow it in the future, always framed in the terms set by the rich and powerful"
Bonus Insights
On Broadcom's place in the chain, he allowed listeners to leave it out of the count: "I don't like thinking about Broadcom very much"
He treats being asked how the labs afford their commitments as a marginal position rather than an obvious question — "I'm considered a firebrand for asking these questions"
He closed by saying he would be back Friday with another monologue
Zitron's bottom line is that the AI industry is being managed rather than fixed, and that a ratings upgrade would buy two loss-making companies more debt rather than a way out of $1.3 trillion of commitments they cannot currently pay.
Products, Companies & Tools Mentioned
OpenAI and Anthropic (Loss-making, IPO-bound, and carrying $1.3 trillion of compute commitments between them)
Amazon Web Services and Uber (The standard rebuttal to AI losses; he says each burned roughly $30 billion, AWS between 2003 and 2015)
Meta and Microsoft (His yardstick for the commitments — around $141 billion and about $175 billion a year of operating expenses)
Nvidia and Broadcom (The end of the spending chain — GPUs and TPUs bought with money raised further up it)
CoreWeave, Cerebras, Amazon and Google (Owed under take-or-pay contracts once capacity comes online, demand or no demand)
Ramp (Source of the figure that 80% of enterprise revenue comes from the top 1% of customers)
S&P Global Ratings and Moody's (The agencies being pushed toward an investment-grade rating; he expects them to do it)
Financial Times (Reported the day before that IPO bankers were seeking those ratings)
Pepsi (Boasted about raising prices while the media debated whether price gouging was happening)
Books & Resources Mentioned
Hypernormalization – Adam Curtis (The two-and-a-half-hour documentary the whole argument is built on)
Hyperscale Normalization – Ed Zitron (His own piece from a couple of weeks earlier, applying Curtis's frame to the AI build-out)
Concentration Risk – Ed Zitron (The companion newsletter for this episode, linked from the show's own notes)
A 2009 congressional panel report on TARP (Source of his line that there was no evidence Treasury used the funds to prevent foreclosures)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: