Running Dry
We recorded 42 claims from this post. 27 were checked against evidence. 15 can’t be checked. 16 more are held. Read the post at its source.
Claims from this post (42)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: UnresolvedSaid
“As I’ve said for a while, all of this ends when the money runs out — when seemingly-unlimited data center debt becomes distressingly hard to raise, and when AI startups can’t rely on the perpetual availability of venture capital to keep them alive.”
What we tested: The AI boom will end when data center debt becomes hard to raise and AI startups can no longer rely on venture capital.
Assessment: Can’t be checkedSaid
“And it appears the well is running dry, just as it dawns on everyone that the underlying economics of AI, for labs and data center operators alike, don’t really make sense.”
What we tested: AI funding appears to be running dry, and the underlying economics of AI labs and data center operators don't really make sense.
Assessment: UnresolvedSaid
“In other words, that $11 billion round won’t even cover a quarter of OpenAI’s compute costs, and it’s fair to ask both why this round had to close right now and what exactly the company does going forward.”
What we tested: OpenAI's $11 billion round will not cover even one quarter of its compute costs.
Assessment: Can’t be checkedSaid
“If OpenAI needed a paltry $11 billion, it’s likely that it raised it either out of some degree of desperation and/or an attempt to catch up with an Anthropic IPO”
What we tested: If OpenAI needed only a paltry $11 billion, it likely raised it out of some degree of desperation and/or as an attempt to catch up with an upcoming Anthropic IPO.
Assessment: UnresolvedSaid
“The problem is that at $1.25 trillion, OpenAI would have to eventually go public at a $1.5 trillion or $2 trillion valuation to give its latest investors anything close to a positive return, at a time when Anthropic has already seen its expected market capitalization tank by hundreds of billions of dollars.”
What we tested: Having raised at $1.25 trillion, OpenAI would eventually have to go public at a $1.5 trillion or $2 trillion valuation to give its latest investors anything close to a positive return, at a time when Anthropic's expected market capitalization has already fallen by hundreds of billions of dollars.
Assessment: Can’t be checkedSaid
“I think the answer is probably somewhere in the middle — OpenAI needed this money, needed to feel competitive with Anthropic, and realized that it should raise as much money as possible before the capital markets slam shut.”
What we tested: He thinks the answer is probably somewhere in the middle: OpenAI needed this money, needed to feel competitive with Anthropic, and realized it should raise as much money as possible before the capital markets slam shut.
Assessment: UnresolvedSaid
“At $1.25 trillion (as mentioned above), OpenAI would need a much better listing price than Anthropic is expected to have, for a company with objectively worse metrics, including slower growth, lower revenue, higher costs, and hundreds of millions of free users that generate, at best, about a billion dollars in “annualized” ad revenue against the billions of dollars it costs to serve their compute.”
What we tested: OpenAI has objectively worse metrics than Anthropic (slower growth, lower revenue, higher costs), and its free users bring in far less ad revenue than they cost to serve.
Assessment: Can’t be checkedSaid
“And I believe that things will get worse from here.”
What we tested: OpenAI's situation will get worse from here.
Assessment: UnresolvedSaid
“OpenAI will absolutely, without a doubt, need more money in the next few months, and will need to raise at some sort of valuation, one calculated based on both that $1.25 trillion valuation and the performance of Anthropic’s IPO.”
What we tested: OpenAI will need to raise more money within the next few months, at a valuation shaped by its $1.25T round and Anthropic's IPO performance.
Assessment: Can’t be checkedSaid
“Let’s be clear about something: it only took a week for Anthropic to file an amended S/1 to include its Q3 numbers, which means (in my opinion) it was absolutely ready to go for the initial filing but chose to intentionally leave those numbers out.”
What we tested: In his opinion, because it took Anthropic only a week to file an amended S-1 with its Q3 numbers, those numbers were ready for the initial filing and Anthropic chose to leave them out intentionally.
Assessment: UnresolvedSaid
“As I said at the time, this quarter was an anomaly caused by organizations allowing their workers to spend as much as they want on tokens, leading to companies like Uber burning their entire annual token budget in the space of a quarter and much hand-wringing around AI’s ROI”
What we tested: Anthropic's Q2 2026 results were an anomaly caused by companies letting workers spend unlimited amounts on tokens.
Assessment: Can’t be checkedSaid
“The era of token-maxxing — when companies like Meta incentivized their employees to burn as many tokens as possible, only to pull back after costs skyrocketed — ended somewhere in Q2 or Q3 2026, which is why Anthropic is so keen to push this company out the door, ideally before you learn too much about its latest quarter or how much its ongoing expenses might be.”
What we tested: Token-maxxing ended in Q2 or Q3 2026, which is why Anthropic is rushing to IPO before investors learn about its latest quarter and expenses.
Assessment: Can’t be checkedSaid
“In other words, Anthropic deliberately filed less information as a means of misleading the public into believing it was a fast, lean, profitable company with remarkably low compute costs and a large — yet indeterminately-timed — pile of obligations.”
What we tested: Anthropic deliberately filed less information in order to mislead the public about its profitability and costs.
Assessment: SupportedSaid
“While revenue grew about 46.9% ($11.5 billion to a little under $17 billion) quarter-over-quarter, compute costs nearly doubled from $5.6bn to $9.9bn (about 77.4%).”
What we tested: In Q3 2026 Anthropic's revenue grew about 46.9% from the prior quarter while its compute costs rose about 77.4%.
What supported it: The quarterly table reports revenue of $11.555 billion and $16.956 billion, and compute expense of $5.558 billion and $9.863 billion. These produce growth of 46.74% and 77.46%, respectively, supporting the approximate comparison. (Anthropic, PBC; reproduced by Soundcast, )
Assessment: Can’t be checkedSaid
“If I had to guess, Anthropic likely — as it’s mentioned it does — pre-bought capacity at discounted rates from Google and Amazon on sweetheart deals that are either running out or have run out this year.”
What we tested: Anthropic likely pre-bought discounted capacity from Google and Amazon on sweetheart deals that are running out or have run out in 2026.
Assessment: SupportedSaid
“Anthropic wants to sell itself as this fast-growing, lean, “profitable” AI lab that has worked out the messy economics of generative AI, when in fact it’s an overweight hog with nearly a trillion dollars’ worth of commitments, $437.4 billion of which are non-cancellable.”
What we tested: Anthropic has not worked out generative AI economics; it has nearly $1 trillion in commitments, $437.4 billion of them non-cancellable.
What supported it: The filing discloses $895.4 billion in aggregate commitments. Microsoft, Google, Amazon, and Broadcom amounts total $437.4 billion, with specified breach, delivery, or default exceptions to their cancellation restrictions. (Anthropic, PBC; reproduced by Soundcast, )
Assessment: UnresolvedSaid
“So, yeah, just another $300 billion or so of commitments on top of the others, disclosed in an opaque — and I’d argue deceptive — way.”
What we tested: Anthropic has about $300 billion of extra compute commitments that it disclosed in an opaque, arguably deceptive way.
Assessment: Can’t be checkedSaid
“Anthropic wants to con as many retail investors — via the media, of course — into believing that this is a company that will only spend $30 billion on compute in 2027, and will somehow grow to meet its ridiculous, ever-expanding costs.”
What we tested: Anthropic is trying to mislead retail investors into believing it will spend only $30 billion on compute in 2027 and can grow into its costs.
Assessment: UnresolvedSaid
“a savvy move to generate capital for a company destined to burn billions of dollars in increasing intervals.”
What we tested: Anthropic will burn billions of dollars at an increasing rate.
Assessment: UnresolvedSaid
“Nevertheless, Anthropic has a remarkable amount of compute commitments, and itself admits that it will have to cough up as much as $447 billion in costs in the next four years, and unlike previous years, it’ll have to pay for it up front in big, ugly chunks.”
What we tested: Anthropic will have to pay up to $447 billion in compute costs over the next four years, much of it up front in large chunks.
Assessment: UnresolvedSaid
“with the sheer scale of these costs, any mismatch between compute spend and revenues will cost the company billions of dollars.”
What we tested: Any mismatch between Anthropic's compute spend and revenue will cost it billions of dollars.
Assessment: UnresolvedSaid
“This leaves Anthropic in a worrying position where the only guarantees it has are that its costs will expand dramatically, every single year without fail.”
What we tested: Anthropic's costs will grow dramatically every year.
Assessment: UnresolvedSaid
“While we still don’t know what level of junk rating the agencies might award it, Anthropic is all but certain to have to raise debt in the 8% to 11% range — the kind of interest rate that can veer between harsh and unmanageable when you face rough headwinds, especially when you’re raising debt in the tens of billions.”
What we tested: Anthropic will almost certainly have to borrow at 8% to 11% interest.
Assessment: UnresolvedSaid
“As discussed last week, around 80% of OpenAI and Anthropic’s revenues come from venture backed startups that (as I’ll get to at the end) heavily subsidize their userbases, which means that growth is mostly tied to whether VC can continue funding its customers.”
What we tested: Around 80% of OpenAI's and Anthropic's revenues come from venture-backed startups that heavily subsidize their userbases, so their growth is mostly tied to whether VC can keep funding those customers.
Assessment: Can’t be checkedSaid
“And as I discussed a few weeks ago, Anthropic will, outside of the very highest rung of junk, be limited in how much debt it can actually take on, and if it needs (as I estimate) around $50 billion a year, it’s hard to see how it pulls that off.”
What we tested: Unless it reaches the very highest rung of junk, Anthropic will be limited in how much debt it can take on, and if it needs around $50 billion a year (his estimate), it is hard to see how it pulls that off.
Assessment: Can’t be checkedSaid
“While its rush toward IPO was arguably driven by its anomalous second quarter, Anthropic has also realized that the music is stopping, and everybody is beginning to ask worrying questions about the underlying economics that I was asking years ago.”
What we tested: While Anthropic's rush to IPO was arguably driven by its anomalous second quarter, Anthropic has also realized that the music is stopping, and everybody is beginning to ask worrying questions about AI's underlying economics that he was asking years ago.
Assessment: Can’t be checkedSaid
“It’s unclear how CoreWeave, which needs over $100 billion in debt through 2030 per Morgan Stanley, actually survives, let alone the gaggle of yet-to-IPO neoclouds like Firmus or Lambda make it through 2027.”
What we tested: It is unclear how CoreWeave, which per Morgan Stanley needs over $100 billion in debt through 2030, actually survives, let alone how not-yet-public neoclouds like Firmus or Lambda make it through 2027.
Assessment: MixedSaid
“Their overstuffed backlogs and highfalutin revenue growth promises are no longer impressing anyone.”
What we tested: Neocloud backlogs and growth promises no longer impress investors.
What supported part of it: The article describes curtailed bond access and financing difficulties across several neoclouds. The event date denotes the reported status of the offering, not a verified cancellation date. (Oninvest, )
What contradicted part of it: Nebius announced a secured Citi facility supporting expansion. This is material continuing lender backing, although it does not establish that unsecured bond investors remained enthusiastic. (Nebius, )
Assessment: SupportedSaid
“In short, fears around neoclouds are now mainstream.”
What we tested: Fears about neoclouds are now mainstream.
What supported it: A major general-interest broadcaster devoted an article to neocloud financial concerns, including views from established analysts. This directly supports the claim about mainstream discussion. (BBC News, )
Assessment: UnresolvedSaid
“The same week, Amazon and Google both had double-misses on earnings, with only their cloud businesses — which are inflated by billions of dollars of compute spend from both Anthropic and OpenAI — showing any exciting growth.”
What we tested: In the same week, Amazon and Google both had double-misses on earnings, and only their cloud businesses, which are inflated by billions of dollars of Anthropic and OpenAI compute spend, showed any exciting growth.
Assessment: UnresolvedSaid
“That $3.8 billion in revenue growth is most-assuredly from OpenAI’s increased compute spend, especially given the dramatic jump from the end of the last fiscal year.”
What we tested: Azure's $3.8 billion quarter-over-quarter revenue growth came from OpenAI's increased compute spend.
Assessment: UnresolvedSaid
“the vast majority of AI subscriptions are heavily subsidized, letting users burn several dollars’ worth of tokens for every dollar worth of subscription and eating the costs to get them through the door, only to rugpull them the months or years to follow as the company in question tries to pull costs in.”
What we tested: The vast majority of AI subscriptions are heavily subsidized, letting users burn several dollars' worth of tokens per subscription dollar while companies absorb the cost to acquire them, only to rugpull them in the following months or years as they try to rein in costs.
Assessment: SupportedSaid
“Masad claimed in another post that Replit was “never subsidized,” something that you’d only say if you had a deep distaste for the truth.”
What we tested: Replit did subsidize its users, contrary to its CEO's claim.
What supported it: The attributed estimate places Replit's April gross margin below zero. On the stated ordinary reading, that means Replit absorbed delivery costs not covered by revenue. (Steven Forth, Steven's Substack, )
Assessment: Can’t be checkedSaid
“it feels like AI is just a series of different scams”
What we tested: The AI industry feels like a series of different scams.
Assessment: UnresolvedSaid
“In any case, I 100% believe Masad when he says that rugpulls are coming.”
What we tested: Rugpulls of subsidized AI subscriptions are coming.
Assessment: Can’t be checkedSaid
“suggesting that OpenAI’s subsidized subscriptions are coming to an end.”
What we tested: OpenAI's token-based usage limits suggest its subsidized subscriptions are ending.
Assessment: Can’t be checkedSaid
“And I imagine this problem will flow throughout the industry as venture capital dries up.”
What we tested: As venture capital dries up, subscription rugpulls will probably spread across the AI industry.
Assessment: UnresolvedSaid
“What I can tell you is that the bubble bursts when the money runs out.”
What we tested: The AI bubble will burst when the money runs out.
Assessment: UnresolvedSaid
“OpenAI and Anthropic need endless venture capital to fund their operations, and when that runs out — unless they can make it to the public markets — they will be crushed under their trillion-plus in commitments.”
What we tested: When venture capital runs out, OpenAI and Anthropic will be crushed by their trillion-plus commitments unless they reach the public markets.
Assessment: UnresolvedSaid
“AI data center capex is almost-entirely funded by debt, which means that when that debt runs out, NVIDIA’s growth — and overall data center construction — will slow.”
What we tested: Because AI data center capex is almost entirely debt-funded, NVIDIA's growth and data center construction will slow when that debt runs out.
Assessment: UnresolvedSaid
“Assuming that Anthropic makes it to the point that it will need debt, any hesitancy around data center debt will make it even harder for it to handle.”
What we tested: If Anthropic needs debt, any lender hesitancy about data center debt will make its position even harder.
Assessment: UnresolvedSaid
“The signs are already there — OpenAI struggling to raise, Musk tightening rate limits on Grok Bot, debt markets shutting out neoclouds, Anthropic’s lower valuation and increasing scrutiny — and will only get increasingly-louder and more-deranged as things get desperate.”
What we tested: The signs of AI's financial trouble (OpenAI struggling to raise, Grok limits, neoclouds shut out of debt, Anthropic's lower valuation) will keep getting louder as things get desperate.
Held from this post (16)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The AIs reached different verdicts. Our reading, not assessed: Anthropic's filing suggests that fairly soon it will have to send a significant amount of cash out the door as pre-payments, on top of spending $9.3 billion a quarter without fail on SpaceX compute, assuming that agreement does not expand further.
- The AIs read the claim differently
- The AIs reached different verdicts. Our reading, not assessed: OpenAI raised at a lower valuation than planned, which is extremely bad, and could not raise as much as it wanted (or possibly needed) at that valuation, which is even worse.
- The AIs read the claim differently
- The AIs reached different verdicts. Our reading, not assessed: OpenAI has higher expenses and lower revenues than Anthropic and the stench of an also-ran beaten to IPO, which leaves it with few options outside increasingly illogical yet existential funding rounds that either scrape the bottom of the VC barrel or force it into a down round.
- It didn’t meet our rules. Our reading, not assessed: A below-investment-grade rating for Anthropic is inevitable, and it would essentially put Anthropic's debt outside the reach of the most prolific bond buyers, such as pension funds and insurers.
- The AIs read the claim differently
- The AIs read the claim differently
- The AIs read the claim differently
- It didn’t meet our rules. Our reading, not assessed: Google reported that its use of AI Overviews had led to lower monetization of search, meaning that despite hundreds of billions of dollars of AI spending, the only tangible effect appeared to be weakening its core product and creating a dependency on two unsustainable, unprofitable AI labs.
- The AIs read the claim differently
- It didn’t meet our rules. Our reading, not assessed: GPT-6 Astra failed to hide OpenAI's bad fundamentals or energize investors.
- The AIs read the claim differently
- It didn’t meet our rules. Our reading, not assessed: To meet its obligations, Anthropic must roughly double in size by November 2027 while keeping growth and improving margins.
- It didn’t meet our rules. Our reading, not assessed: Microsoft's reliance on OpenAI revenue is growing alarmingly fast while OpenAI struggles to raise money to pay its bills.
- The AIs read the claim differently