Ed Zitron 解析 OpenAI 1.25 万亿美元估值融资与 Anthropic S-1/A 中的万亿级算力承诺

If you haven’t subscribed to Where’s Your Ed At, now’s the best time. I’ve extended the sale — 10% off of the $70-a-year subscription for life (click here!) — for new subscribers for another week. It’s a great way to support my work, and you get an absolute ton of value at a lower price than basically every competing newsletter.

Over the past few months, I’ve published a two part guide to the world of AI debt, an analysis of the (dubious) impact of generative AI on the global economy, deep analyses of the biggest players in the AI infrastructure space like NVIDIA (part one and two) and Oracle (part one and two), how OpenAI will eventually kill Oracle, and how venture capital is killing itself.

Friday's piece is about fault lines - the areas most-exposed to the collapse of the AI bubble, the datapoints to watch to see the industry's deterioration, and what might cause things to fall apart.

If you want to get in touch — and especially if you have any juicy information about Anthropic, OpenAI, Blackstone or any other companies in the AI bubble — hit me up on Signal at ezitron.76. I’m also on IB on your Bloomberg Terminal.

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.

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.

On September 16, 2026, the New York Times would report that OpenAI was “considering” a new funding round valuing it at $1.5 trillion, a little under three months after it delayed its IPO to 2027. Ten days later on September 29, Bloomberg would report that it was aiming to raise $30 billion at a $1.4 trillion valuation, and on October 5 revealed that it was in talks with BlackRock, a syndicate, and funds from the UAE which had “discussed putting in as much as $10 billion altogether.”

Last week — less than a month later — OpenAI raised $11 billion at a $1.25 trillion valuation “after receiving tepid interest from some large investments,” per the Wall Street Journal, who added that OpenAI had “...told shareholders it expects to try to raise more money…[and hoped] to raise up to $20 billion,” targeting “family offices and high net-worth individuals,” which means that sovereign wealth funds, venture capitalists, and private credit firms have given all they’re willing to.

Per The Information, OpenAI spent a total of $12.1 billion across its cost of revenue and research and development segments in Q1 2026, and per Microsoft’s own earnings, OpenAI’s compute spend and revenue share payments contributed $10 billion in revenue to its latest fiscal quarter.

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.

Per the same Information story, OpenAI ended Q1 2026 (so March 31, 2026) with “more than” $73 billion in cash and marketable securities. Revenue grew tepidly from $5.7 billion in Q1 to $6.7 billion in Q2 2026 (per the WSJ), and since then, Amazon has sent OpenAI $35 billion (around July 31, 2026) and SoftBank $30 billion (wired in April, July and October).

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 that investors are now expecting to, per Bloomberg, price at $1.4 trillion to $1.5 trillion, down from less than a month ago when it was expected to price at $2 trillion or more. Anthropic is now, per the WSJ, being “urged by institutional investors to target a lower valuation of $1.5 trillion to $1.6 trillion.”

It is extremely bad that OpenAI raised at a lower valuation than planned, and even worse that it couldn’t raise as much as it wanted (needed?) at said valuation. In the absolute best case scenario, this was an attempt by OpenAI to bump its valuation up from its $852 billion par from March, I imagine to make its investors feel better about Anthropic’s upcoming IPO…potentially? 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.

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.

The question I ask is simple: what the hell does OpenAI do next?

OpenAI’s $1.25 Trillion Round Is A Bad Sign

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. Despite all the buzz around GPT-6 Astra — which was, according to NVIDIA’s Jensen Huang, “AGI” — it’s clear that its effect on revenues was insufficient to excite even the cash-rich and hype-friendly investors in the Gulf.

GPT-6 Astra was, I believe, expected to be the lipstick on an unmistakably feculent pig of a company — something that’ll disguise the horrendous fundamentals of OpenAI’s business and energize investors — and, based on its disappointing latest funding round, it failed. And I believe that things will get worse from here.

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.

Making things a little more challenging is the Financial Times’ report that Blackstone has told some investors that Anthropic “may not receive an investment-grade rating immediately after going public,” otherwise known as a junk rating, a point I made about a month ago based on the fact that it’s a negative-EBITDA company.

While we don’t know what the rating will be — as the big three (Moody’s, S&P and Fitch) are yet to rule — Anthropic’s $35 billion Broadcom-backstopped deal to lease TPUs received a BB+ rating from KBRA a few months ago, which is one rung below investment grade, but I must be clear that $30bn of the deal is backstopped by a company with a high investment-grade credit rating.

While inevitable, a below-investment grade rating essentially puts Anthropic’s debt outside the reach of the most prolific bond-buyers, like pension funds and insurers. And again, I couldn’t see this going any other way, but I would argue that this rating represents a psychological cold shower for investors — a reminder that this space is, using the most charitable language at my disposal, inherently speculative and rife with risk. Given that OpenAI’s past and future fundraising is contingent upon an investor class that’s drunk on exuberance and that has lost its senses, cold showers are a very bad thing.

OpenAI is a company with higher expenses, lower revenues, and the stench of an also-ran that got beaten by an aggressive competitor gunning at full speed to beat it to IPO, leaving it with few options outside of increasingly-illogical-yet-existential funding rounds that either scrape the bottom of the VC barrel or force it to take a down round to get cash through the door. Future investors will have to believe, somehow, that OpenAI will both actually go public and one day be worth $2 trillion to $3 trillion, as well as somehow being able to raise hundreds of billions of dollars of debt to keep up with its $750 billion in planned cloud spend through 2030.

And given that junk rating and the grim new data from Anthropic’s newly-amended S1 (called an S1/A), it’s unclear how any of this is actually meant to work.

Anthropic Has $895 Billion In Spending Commitments, With A Minimum Of $319.9 Billion Due In The Next Four Years

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.

You’re about to find out why.

Anthropic’s Q2 2026 Numbers Were An Anomaly Of Artificially-Depressed Costs and Revenues Inflated By Token-Maxxing

As I discussed last week, Anthropic’s initial S-1 filing — covering up to Q2 2026 — described a company that effectively no longer exists, with a “profit” driven by discounted SpaceX compute, $2.08 in revenue for every dollar of compute spend, $11.5 billion in revenue, and $430 billion in non-cancellable compute commitments, including a $161.2 billion deal with Broadcom to lease Google’s TPU chips at a remarkable scale.

Though we already knew some of the Q3 numbers thanks to the Financial Times’ reporting around Anthropic’s investor day, boosters and the otherwise cope-poisoned chose to do several consecutive victory laps, claiming that this was proof that A) Anthropic was a “profitable company” and B) that “the economics of generative AI work.”

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:

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.

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.

Having read the amended filing, I kinda get why it did so.

Anthropic’s Q3 Numbers Show Exploding Costs, Slowing Revenues, and $895bn to $1.07tn In Obligations, With At Least $69.8bn In Compute Costs In The Next Year and $437bn In Non-Cancellable Contracts

As part of its original filing, Anthropic included this delightful diagram of its “revenue per dollar of compute” — but left out Q3 numbers that we already know it had (because the Financial Times reported them weeks ago) for the obvious reason that its economics got worse.

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%). This was partially driven by the fact that, per Bloomberg, Anthropic’s compute spend made up 34.5% — around $5.028 billion — of SpaceX’s Q3 revenue, making it Musk’s largest customer.

Sidenote: I’ll also add that Anthropic’s compute costs still, to this day, do not make sense in comparison to those borne by OpenAI and basically every other company renting GPUs and TPUs. While it still lost $567 million in Q3 (and had a negative 3% operating margin) it only spent $4.9 billion on compute (outside of SpaceX) across Google, and Amazon, despite Amazon opening an entire data center in Indiana for it last year and signing a deal worth “tens of billions of dollars” with Google in October 2025 that would “bring over a gigawatt of capacity online in 2026.”
As I reported a year ago, OpenAI — in a year where it made $13.01 billion in revenue —
spent $3.64 billion on inference in Q3 2025. While it still had the burden of the free Sora app at this time, not to mention its free users, I find it hard to understand the significant gap between the companies’ costs, and do not buy that Anthropic has some super secret sauce that makes it significantly more efficient than OpenAI to the extent that we see in the numbers.Per my own reporting on Anthropic’s Amazon Web Services bills from 2025, Anthropic spent $1.225 billion on AWS alone in Q3 2025, in a quarter that (per the S-1) had a total of $2.047 billion in compute costs, leaving about $1.71 billion in costs in that 9 month period unaccounted for, likely going to Google Cloud.
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. For example,
UBS’ Stephen Ju cited in a note from earlier in the year that Anthropic had a deal to pay Google a mere $1.25 billion in annual revenue outside of a $25 billion commitment to TPUs, and another agreement beginning in FY2027 at $6.25 billion a quarter.
I guess we’ll never know, unless one of you wants to share Anthropic’s Google Cloud or Amazon Web Services spend! Ezitron.76 on Signal.

Anyway, Anthropic also filled in that contractual obligations chart, showing a minimum of $319.9 billion in payments through the end of 2030.

Yet before anyone starts suggesting that Anthropic has fewer obligations than we thought, it immediately goes on to add that there are $451.2 billion in leases yet to commence…

Further, the table above does not include leases executed and not yet commenced as of September 30, 2026, with total estimated future undiscounted lease obligations of approximately $451.2 billion. Generally, these lease obligations are expected to commence between 2026 and 2028, with lease terms ranging from less than one year to 20 years, with a weighted-average term of approximately 10 years.

…and, for whatever reason, leaves out over $100.2 billion in future obligations to hyperscalers and data center lessors, as well as “certain agreements that are short-term in nature” with a minimum spend of $6.6 billion.

In addition, the table above does not reflect commitments entered into after September 30, 2026, with aggregate future minimum payment obligations of approximately $100.2 billion, primarily consisting of data center leases and CSP hosting and other third-party infrastructure service agreements. We also entered into certain agreements after September 30, 2026 that are short-term in nature. Minimum payments under such contracts are approximately $6.6 billion. See Note 17 to our consolidated financial statements included elsewhere in this prospectus for more information on commitments entered into subsequent to September 30, 2026.

Anthropic does add that it expects to pay $69.8 billion of its contractual obligations in the 12 months starting October 1, 2026 — meaning Q4 2026 — which makes that “$29.1 billion in payments due by year” number extremely suspicious, to the point that it almost feels like Anthropic intentionally wants readers to get the wrong idea about the company.

I mean, I get it! 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.

Anthropic Appears To Have More Than $1 Trillion In Compute Commitments, And Uses Legalese To Obfuscate Its Truth Obligations

Just to review, that’s…

$319.9bn in purchase commitments and commence leases
$451.2 billion in signed leases that haven’t commenced.
$17.4 billion in short-term commitments.
$100.2 billion in commitments signed after September 30.
$6.6 billion in additional short-term commitments signed after September 30.
$69.8 billion due to be paid within the twelve months starting October 1, 2026

Yet even then, there’s potentially hundreds of billions of dollars more in compute commitments that Anthropic vaguely describes, the first being underneath that chart in a footnote, emphasis mine:

Excludes certain other compute and infrastructure arrangements that do not specify a fixed dollar commitment which are not included in the aggregate above. Payments under these arrangements will depend on the capacity made available and deployed, delivery timing, pricing mechanics, and the satisfaction, or our waiver, of certain other deployment conditions. Because these factors are not yet determinable, the total amount of payments under these arrangements is uncertain, but we currently estimate such payments to be approximately $150 billion to $175 billion. We expect these amounts to be payable over time after the capacity is delivered.

No worries dude! Just $150 billion in compute commitments that might happen at some point or another with no explanation as to who the commitments are to, what they’re for, and when they might begin, other than that they might be part of a deal with a hyperscaler!

Say, talking about hyperscalers, Anthropic also apparently has another massive commitment to lease TPUs — apparently separate to the Broadcom deal! — that may or may not be worth as much as $125 billion to $150 billion, and deliver sometime in the next two to five years:

In addition, we also committed to procure additional TPU compute, without a fixed dollar commitment, via one of various delivery methods, including either a CSP hosting agreement with Google, direct purchase from Google with five-year financing, or direct purchase and/or lease from a third-party TPU provider with five-year financing. Such compute is expected to be delivered between 2028 and 2031 and the associated payment obligations are determined by the capacity made available, the delivery method, and the pricing and financing mechanics. We estimate the aggregate future payments for such additional compute to be approximately $125.0 billion to $150.0 billion, which are not included in the contractual obligations table above.

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.

But that’s just how Anthropic (and its former Blackstone private equity CFO) rolls. 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.

Anthropic’s Discounts Are Eating Into Its Revenues, Costing It $600 Million In Q3 2026, An Increase of 757% Year-over-year

Underneath the quarterly trends section of the S/1A, Anthropic added an entire paragraph underneath “quarterly revenue trends” on page 161 (158 in the original S-1):

For the three months ended September 30, 2025 to the three months ended December 31, 2025, the increase in revenue due to higher usage was partially offset by a decrease of approximately $70 million from price reductions introduced as part of new model versions and contractual discounts. For the three months ended June 30, 2026 to the three months ended September 30, 2026, the increase in revenue due to higher usage was partially offset by a decrease of approximately $600 million from price reductions introduced as part of new model versions and contractual discounts.

In other words, for every dollar that Anthropic makes, it’s cutting that by 3.4% or so via discounts, a fact it’s only including as a means of key-jingling for boosters to say “actually revenue would be so much higher, if…” even though this number doesn’t really tell us very much about Anthropic’s business. Are these discounts for enterprises buying tokens direct? What were the discounts in Q1 and Q2 2026? Don’t worry about it man, these numbers are not here for you to learn more about Anthropic, they’re about making sure you invest in it.

With A Junk Rating, It’s Unclear How Anthropic Affords Its $330bn to $447.5 Billion In Compute Commitments In The Next Four Years and Decaying Margins

Anthropic’s cash position is one of the weirdest things I’ve seen in a company, and I’ll explain why with this quote from economist Paul Kedrosky when he dug into its original filing, where he called it “a bank with a money-losing AI side gig”:

Anthropic held $112 billion in cash and equivalents at June 30, plus another $19 billion in short-term investments. It generated $1.1 billion in interest income on that balance in the six months. Granted, this won't get you a gig at PIMCO or Doubleline, but it is massive.
To put in context, Anthropic's interest income is larger than their operating losses of $700m+. It is also larger than many regional U.S. banks. While the interest income provides a cushion beneath operating losses, it doesn't change underlying operating economics. For now, being well-funded is itself a meaningful source of earnings. That benefit will go to zero as the cash gets spent.

In Q3, Anthropic’s cash position is largely left unchanged, with $16.8bn in cash and cash equivalents turning into another $17 billion in short-term investments (US Treasuries and securities), generating $1.248 billion in interest — a savvy move to generate capital for a company destined to burn billions of dollars in increasing intervals.

And the problem for anyone trying to digest this S-1/A is that there’s very little guidance as to how that might actually happen.

Let’s go back to the top of page 6:

These commitments include approximately $447.5 billion expected to be incurred before December 31, 2030, and substantially all of the remainder is expected to be incurred between 2031 and 2045. Approximately $69.8 billion of these commitments are due in the 12 months following September 30, 2026.

Now, that $69.8 billion could, in theory, include some of SpaceX’s $3.1 billion a month in compute costs — which are, of course, “short-term leases” that can be exited with 90 days of notice — but it’s impossible to tell thanks to this insipid company’s weasel wording. There’s also a compelling argument that it only includes a quarter’s worth, which would add another $27.9 billion on top of the $69.8 billion in the event that Anthropic used it for the entire year.

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.

In Anthropic’s own words, it “expects an increasing portion of [its] compute and infrastructure investment to take the form of capital expenditures and other commitments,,” and that said commitments require “significant upfront payments well in advance of the related compute becoming operational.”

While one might be tempted to say that its current liquidity — about $95.5 billion in cash and $36.2 billion in short-term investments — in addition to potentially raising $100 billion at IPO might be enough to get it by for years, even Anthropic suggests that might not be the case:

We believe our existing available cash and cash equivalents, short-term investments, together with cash generated from sales and amounts available under our A&R Credit Facility, will be sufficient to support working capital, operating and finance lease payments, and capital expenditure requirements for at least the next 12 months from the date of this prospectus.
However, our future capital requirements will depend on many factors, including our revenue growth, the timing and extent of our investments in compute and infrastructure capacity, including additional or expanded CSP, equipment lease, and data center arrangements, the success and expansion of our sales, marketing, and business development strategies.

This suggests that fairly soon Anthropic will have to send a significant amount of cash out the door in the form of pre-payments, on top of the fact that it’ll be spending $9.3 billion a quarter, without fail, on SpaceX’s compute, assuming that that agreement doesn’t expand any further.

Anthropic’s Future Is Brittle, Its Obligations Ridiculous, And Every Bit Of Debt More Dangerous Than The Last

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. While revenue growth might continue (though Ramp, which was accurate with its predictions of Anthropic’s Q3 growth, now expects its Q4 numbers to be far more modest, ranging between 15% and 25%), said growth now has to compete with hard, inflexible costs regardless of any seasonal shifts, price increases (or decreases) or product changes that might hamper growth, and with the sheer scale of these costs, any mismatch between compute spend and revenues will cost the company billions of dollars.

And then there’s the problem of that junk rating. 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.

Anthropic’s current obligations are priced for perfection, requiring the company to, by this time next year, be roughly twice its size and continuing at a similar (or better) growth rate, because the sheer scale of its compute commitments demands Anthropic maintains both perpetual growth and improved margins.

It is not enough for this company to be the fastest growing startup of all time — it must also be attractive enough as a potential debtor to appeal to an already-strained junk debt market, which will require it to become a better, stronger company in stark economic terms unrelated to any hype around AI, especially as Wall Street becomes, to quote the WSJ, “choosier about where to place their bets and are demanding tougher terms.”

To be clear, that’s referring to AI data centers that have physical infrastructure and actual assets rather than a negative-EBITDA software company with a very annoying CEO.

While Anthropic’s cash pile might make lenders a little bit calmer, we still have very little clarity as to the size and velocity of its prepayments, which makes it difficult to tell how big that cash pile might be in three, six or twelve months.

My question is simple: what happens when Anthropic’s revenue growth stalls as its costs naturally increase as capacity comes online? It’s deeply unattractive to lenders to say “well, we made all these commitments to pay hundreds of billions of dollars and it turns out our revenue growth slowed down, can you lend us money to cover the gap?” While in theory that might work, it’s the kind of thing that makes a lender ask “are you concerned about defaults?” or “what happened to all that cash you had?”

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.

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.

It’s getting dark outside, whether boosters will admit it or not.

The AI Trade Is Collapsing As Reality Dawns On The Industry, With Debt Markets Closing, AI Making Google Search Less Revenue, and The ROI Question Unanswered

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.

Everything started to accelerate last week off the back of Anthropic’s S-1. The day before it was filed, Elon Musk closed a $40 billion debt deal to buy NVIDIA chips with interest rates for some parts of it as high as 8.75% per Semafor. A day later, Musk claimed that the markets had begun to shut out neoclouds entirely, per CNBC:

“As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers,” Musk, CEO of SpaceX and Tesla, said on a post on X on Sunday. “That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many,” he added.

A few days later, the Wall Street Journal would report that the markets have begun rejecting neoclouds:

The new terms of trade are visible in every corner of markets. Bonds issued even by big tech companies such as SpaceX and Meta are trading at steep discounts. AI cloud companies such as CoreWeave, Nebius and Nscale that borrow big to build and sell capacity have effectively been shut out of the bond market in recent weeks, people familiar with the matter said.

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. Their overstuffed backlogs and highfalutin revenue growth promises are no longer impressing anyone.

More tellingly, the mood around neoclouds have shifted — demonstrated most vividly by a piece published by the BBC’s North America technology correspondent Lily Jamali over the weekend, which essentially talked about the growing skepticism about the long-term health of neoclouds on Wall Street, and noted that many (in particular, Coreweave) have racked up debts that they really shouldn’t have been able to.

Neocloud firms often finance themselves by signing long-term lease contracts with major tech customers who need quick access to computing power.
Firms can then borrow money using those contracts – plus the AI chips they purchase – as collateral, explained analyst Cory Johnson of "The Quiet Period" IPO research outfit.
He said debt is common across neocloud firms because the hardware has to be paid for up front.
Take the industry leader: CoreWeave. It's a fraction of the size of some of the Big Tech companies it serves.
But Johnson believes lenders sometimes consider the creditworthiness of CoreWeave's customers, like Microsoft or Meta, more than the company's when they lend.
"This helps explain how CoreWeave accumulated $35.1bn (£26.3bn) in debt," he said.

I’ll simply note that I raised the alarm on Coreweave in March of last year (nearly 18 months ago), and many of the points that the article makes are ones that I’ve been raising in this newsletter ever since. In short, fears around neoclouds are now mainstream.

And while hyperscalers might still have been able to raise debt this week, last week’s earnings have thoroughly panicked Wall Street.

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.

Per page F-16 of the amended S-1, 44% of Anthropic’s revenue (around $14.6 billion) for the first three quarters of the year came from Google and Amazon (and Microsoft to a lesser extent) selling its models through Vertex and Bedrock, which means both have seen billions of dollars of extra revenue both from selling Anthropic’s models and providing it with compute to deliver them.

In other words, in 2026, Anthropic got $10 billion in funding from Google, $5 billion in funding from Amazon, then spent billions of dollars on both Google Cloud and AWS to serve compute to sell access to models, which the hyperscalers then booked as revenue, on top of the revenue the hyperscalers booked from selling those models, which (due to the unique way that Anthropic itemizes its revenues) Amazon, Google and Anthropic all book as their own revenue.

While Amazon could explain away a bad quarter as a function of a flailing economy with spiralling inflation, Google had far grimmer news — that its use of AI overviews had led to “lower monetization” of search, meaning that despite spending hundreds of billions of dollars on AI, the only tangible effect appeared to be weakening the core product and creating a dependency on a pair of unsustainable, unprofitable AI labs.

As a result, Google had its single-worst day of trading in history.

OpenAI Was 11% of Microsoft’s Q1 FY2027 Revenues, And Has Become A Load-Bearing Part Of Microsoft’s Growth

While Microsoft beat expectations, it did so at a gruesome cost — as mentioned earlier, OpenAI’s compute spend and revenue share payments contributed $10 billion in revenue to its latest fiscal quarter, or almost 11% of its quarterly revenues.

This is extremely bad in a way that I’m surprised more people aren’t discussing. Per Bloomberg earlier in the year, OpenAI accounted for around 7% of Microsoft’s Fiscal Year 2026 revenues, or around 70% ($24.1 billion) of AI revenues. Within a single quarter, OpenAI has already accounted for nearly half of its last fiscal year contribution to Microsoft, with its concentration of overall revenues growing at an alarming rate at a time when, as discussed, the company is having trouble raising money to pay its bills.

There’s also the messy issue of Microsoft’s new revenue disclosures, announced back in September, that mean that we now know the quarterly revenue of its Azure cloud platform, which was an impressive-seeming $33.2 billion in its latest quarter until you, as nobody appears to have done, compare it to the $29.4 billion it made in Q4 FY2026.

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. If that’s the case, Microsoft’s “impressive earnings beat” is far more a function of OpenAI’s cash furnace than any success of its AI plays — especially as Microsoft neglected to update us on paid Copilot licenses outside of claiming that “Copilot daily users grew nearly 50% in the three months to September,” the kind of thing you tell an audience you are convinced is too stupid to see through your bullshit.

The Subprime AI Crisis Continues, And The Rugpulls Await

And just when we thought we’d escaped the fearmongering of a shitty earnings season, Uncle Elon decided to say the worst thing possible as users complained about their weekly usage of “Grok Bot” suddenly evaporating, per TechCrunch:

SpaceX CEO Elon Musk confirmed in an X post on Sunday that this was not a glitch.
“The cost of serving Grok Bot was unsustainably high relative to revenue,” he wrote in a reply to a question from a user. He has not addressed the issue on his main feed but has continued boasting that Grok Bot can help you buy gifts for your family.

Musk, who had said three days beforehand that saving money for the distant future was pointless with the advent of “super intelligent abundance,” added that these changes (which have yet to be fully defined) are a result of him “just being sensible.”

Sidenote: The cost of running any kind of “agentic” service like Grok Bot is also much, much higher, as on top of the cost of inference, each one of them is given a virtual machine with an allocation of RAM and storage, and can spin up multiple different LLMs at once to work on a problem, with the underlying orchestration of all their actions (EG: connecting to APIs, databases, web searches, using applications) putting a heavy burden on the CPU.
Considering most AI services are billed on a monthly basis — and services like Muse and Instinct are free — these “agentic” features are a brand new infrastructural cost layer above and beyond the regular costs of LLMs. It’s a testament to the stupid amount of compute power necessary to get these things to do anything close to what we were promised with ‘agents’ over the last few years, and even then the results are mediocre.

This freaked out the entire AI industry, because they realized, as I warned two years ago (and revisited again in March), that 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. This all began in June last year when Anthropic and OpenAI forced AI startups dependent on their models to start paying priority processing fees.

A few hours later, Replit CEO Amjad Masad gave a dire warning: that if you’re enjoying a subsidized AI subscription [you should] get ready to get rug pulled.

After all, Masad would know — shortly following the June 2025 changes to priority processing, Replit introduced “Effort-Based Pricing” that jacked up pricing for users, a problem made worse by the launch of its “Agent 3” product a few months later, per The Register:

Part of the problem might be with Replit's June introduction of "effort-based pricing.Previously, it would charge $0.25 for each checkpoint, and a task with multiple checkpoints would simply be tallied up one by one to arrive at a total cost. But with effort-based pricing, it now bundles complicated tasks into a single more expensive checkpoint. The next month, Replit admitted "can end up being more expensive over the lifetime of a project." But the costs didn't really seem to hit users until Agent 3.
"The effort-based pricing never ran me as much before but Agent 3 has been exceptionally high," explained the user we spoke to.

A short trip to the Replit Subreddit is all it takes to find unhappy customers, but my own thread soliciting experiences about effort-based pricing led to arguably the most damning response that, well, uses exactly the word that Masad chose:

Replit's new pricing is just the latest rug pull in a long line of half-baked rollouts. I’ve spent months building serious infrastructure on their platform—AI-driven automation, multi-role dashboards, recursive agent logic. What I got in return? An unstable, overpriced Agent that breaks more than it builds.
…
And now they’ve changed pricing again—with no warning—making it even less predictable. Their “effort-based” system means you can get charged more for the Agent’s failure loops than if you wrote it yourself. I’ve seen single-line fixes priced higher than full feature requests—and that’s assuming it even finishes.

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.

Per The Information from August 2025, emphasis mine:

For Replit and other coding startups, one of the biggest factors weighing on their gross margins is the fees they pay every time their assistants access large language models developed by Anthropic and others. Replit, which targets professionals without coding backgrounds, includes payments for Anthropic’s Claude models and Google Gemini models, as well as cloud computing costs to deliver its products, in the cost of its revenue.
In February, as revenue surged, Replit’s gross margins hit 36%. Then they fell to negative 14% in April when the company launched a new version of its agent product, which was more autonomous and spent more accessing large language models than the prior version.
As a result, the company repriced the product—charging for the amount of time and compute used for the agent to complete a task rather than for each completed request—to account for the additional costs. In July, the margins bounced back to 23%.

Riddle me this, Batman: what do you call it when a company sells a service at a price that does not cover its costs, absorbing said costs to retain the customer? I’m sure he has some sort of mental gymnastics routine where he considers that spend to be “marketing costs,” but we all know what actually happened.

Sidenote: Have you ever noticed how fucking dishonest this entire industry is? Perhaps this’ll be next week’s free newsletter, but it feels like AI is just a series of different scams — the obfuscated economics, the weasel wording around everything from revenues to how much actual capacity companies have, subsidized subscriptions that literally sell customers on a lie, it’s all so fucking grotesque! I hate it! Everyone involved should be ashamed of themselves!

In any case, I 100% believe Masad when he says that rugpulls are coming. A recent update to OpenAI's ChatGPT and Codex usage limits page revealed that starting in December, users’ use of chat and coding functionality will draw from the same rate limits (which had been largely separated to this point) and that said usage will “be measured in tokens,” suggesting that OpenAI’s subsidized subscriptions are coming to an end.

And I imagine this problem will flow throughout the industry as venture capital dries up.

It Ends When The Money Runs Out, And AI Is Priced For Perfection

I think I get asked five times a week when this all ends, and I truly cannot answer you and, as some of you have noticed, get criticized for having tried to guess in the past.

What I can tell you is that the bubble bursts when the money runs out.

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.
80% of OpenAI and Anthropic’s revenues come from venture-backed AI startups, which means that the underlying revenue growth of these companies is contingent on venture capital’s continual subsidies which, at some point, will run out.
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.
- Per Westside, data center construction starts fell by more than half in September 2026 month-over-month ($9.4bn, down from $25.8bn in August), as the median cost of a new build grew by 31% year-over-year.
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.

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.

And now I’m even hearing low rumblings of a compute glut, with AI startups buying compute on the off-chance they might need it, assuming wrongheadedly that if they don’t they can sell it at a premium.

You know, kind of like how people said in 2006 that if they’d just refinance their mortgage.

After all, the price of housing never goes down.