Not Enron, but not good either.

[Picking up from this and this]

I realize that my recommendations for Patrick Boyle videos tend to be a bit heavy-handed, but even by his standards, this one is essential for anyone trying to get a handle on the AI bubble and its increasingly likely ramifications. It is an extraordinarily pithy half hour, hitting a number of important topics. I really need to do at least one or two more posts on this.

For today, let's focus on what may be the key point in the entire 2026 AI market discussion: the unresolvable disconnect between the trillions of dollars in capital expenditure and debt accumulated (both on the books and off) and realistic estimates of the total addressable market.

Side note: I also need to revisit our long-running economics of streaming thread in the context of Cornell and Damodaran's remarkably prescient The Big Market Delusion: Valuation and Investment Implications. 

The trouble with vendor financing is what happens when it doesn't. Then it's a double blow. You don't just lose the customer; you lose the money you lent them to be your customer.

And the credit guarantees make it worse. If an equity stake goes to zero, that's just money wasted. Annoying, but survivable.

But a promise to cover a customer's debts if things go wrong can turn a valuation problem into a solvency problem.

Right now, Nvidia throws off something like $200 billion a year in cash. So if one or two of these startups trip, it can take the hit. The question is what happens as the guarantees climb into the hundreds of billions and a company that used to avoid debt is suddenly standing behind everyone else's.

You don't have to take my word that this matters. The clearest sign is in Nvidia's own credit market. The cost of insuring its debt against default just jumped by the most on record in a single day, right as this round of deals landed.

 

So the people whose actual job is to price the risk of Nvidia not paying its bills had a look at all of this and got noticeably less relaxed.

Because the real risk was never just that the AI market turns out smaller than hoped. It's that the people buying the chips and the people making the chips are increasingly the exact same people.

All of this circular financing is happening because everyone involved is convinced that the market for AI is going to be so astronomically large that whatever they spend today will look like a rounding error tomorrow.

Aswath Damodaran has a name for what happens next. He and his co-author, Bradford Cornell, call it the "big market delusion." [One of those papers that deserves a lot more attention -- MP]

The way it works is that a new technology shows up attached to a massive potential market. A crowd of companies crop up to serve it, and investors price each company as if it's going to be the winner.

This is not about the companies talking themselves up. It's about the people buying the shares. Each cluster of investors looks at their chosen company and sees it as the obvious future giant.

The problem is that they can't all be right. If you take these companies and add up what the market expects each of them to earn, you get a number bigger than the market itself.

Everyone's been priced to come in first in a race that can have only one winner, which is how a whole market can be priced for a future that mathematically can't happen. The story is doing all the work, and nobody's minding the numbers.

So, how big is the story here?

The Economist estimates that the AI buildout is on track to be the largest investment surge in history. Around $900 billion this year alone is being spent on chips, data centers, and power, with more than $400 billion of it borrowed.

 

And then they calculated what it would take to pay for all of that.

Their estimate is that the industry would need to be earning something like $2.5 trillion a year in AI revenue, which is more than the entire global technology sector earns from everything it does today.

The actual figure is not close.

Adoption is real. Around a fifth of American firms report using AI in some way, but a lot of them are using the free versions.

According to a Bank of England study, the average American executive spends about 100 minutes a week using AI. That's not a typo.

The largest capital investment in the history of the species is being justified by an hour and a half per executive per week. So, somewhere between lunch and the drive home.

And when users do pay, they don't pay much. The fintech firm Ramp went through actual company spending and found that the median firm was spending per employee per month $10.66.

$2.5 trillion a year being spent to capture $10.66 per employee.

The most damning number that came out of the Bank of England's research was that nine out of ten executives said that AI had made no difference to their company's productivity over the past three years.

When a technology takes over the economy, people usually tend to notice.

But that's the view from the top. Look at the other end of the economy and the picture flips.

The people getting real value out of AI aren't the giants spending hundreds of billions on it. They're the small ones.

According to a survey by the payroll firm Gusto, the share of new business founders who used AI to get started doubled to 60% in two years.

They are not using it to cure a disease or replace a department, but to build a website, handle the local paperwork, and do the things that used to mean hiring someone.

Now, some of this new business activity is people incorporating their hobbies. And a shrinking share of these firms will ever employ anyone but the founder, so let's not oversell it.

But the clearest real-world win for AI so far isn't the company burning billions on it. It's the person starting a one-man business paying about $20 a month.

 

Is it a bad sign when people start making Enron comparisons? It seems like a bad sign.

Following up on yesterday's post. If you're not interested in the AI bubble, you might want to come back next week because this thread is going to be running through Friday.

Here is Ed Zitron discussing the AI industry's huge off-the-books debt. Before we jump in, however, there's one point I want to emphasize (we'll dig into this even deeper tomorrow when we discuss Patrick Boyle's analysis). This story is Enron-esque. It is Enron-reminiscent. It is "a little bit Enron." It is not, however, another Enron.

The executives at that company engaged in criminal accounting fraud. People went to jail. As far as I can tell, no one is seriously accusing any of the major AI players of that kind of Enron-level behavior. The key phrase here is "at least on a balance sheet basis." Oracle and all the rest are hiding their debt from people who do not read the footnotes. Legally, that's a huge distinction, the kind that determines who goes to jail and who doesn't, but the difference can be smaller than you'd expect (more on that tomorrow).

The problem with these SPV-based deals is that they allow companies to, at least on a balance sheet basis, hide the scale of their debts. Meta’s long term debt sits, as of its latest quarter, at around $58.7 billion. It’s as if the $39 billion in debt for gigawatts’ worth of AI data centers doesn’t exist out of the payments it’ll eventually have to make. 

This is all legal, worrying, and yes, a little bit Enron.  

 

Per Amanda Iacone of Bloomberg:

Enron Corp. exploited US accounting rules to hide from investors and lenders hundreds of millions in debt it had bundled into off-balance sheet entities — obligations that contributed to one of the biggest corporate collapses in US history.

Alphabet Inc. and Meta Platforms Inc. each have turned to vehicles known as variable interest entities (VIEs) as part of the financing mix needed to construct data centers and related energy infrastructure.

Meta, the parent of Facebook, last year formed a joint venture, a VIE, to build a Louisiana data center through a partnership with Blue Owl Capital. The social media titan’s maximum exposure for the venture is $46 billion, according to its filings with the Securities and Exchange Commission. The company announced last week that it would expand its planned campus and is expected to spend as much as $250 billion on the project, Bloomberg News has reported.

To be clear, a Variable Interest Entity is a type of SPV where you have control over the entity, and you must consolidate it into your balance sheet…unless you are not considered the “primary beneficiary,” which Meta argues isn’t the case despite being the primary tenant and reason that Hyperion is being built. Per Bloomberg:

Meta determined it shouldn’t bring billions in debt from the Louisiana project onto its own balance sheet because it isn’t responsible for finding tenants to replace or join it at the nearly 4,000-acre campus — a critical job that impacts the entity’s economic performance, the social media company said in its most recent quarterly SEC filing. Meta said its role is limited to construction management, along with administrative and property management services.

Auditor Ernst & Young raised a “red flag” (per the WSJ) about this arrangement, flagging it as a “critical audit matter,” adding that it “...was especially challenging due to the significant judgment required in determining the activities that most significantly affect the VIE’s economic performance.” Nevertheless, it was approved, it happened, and everything is fine and normal. 

This is why Google backstopped Fluidstack and Cipher Mining’s 300MW data center and another for TeraWulf. Both will, eventually, operate as data centers that Google will lease to provide compute to Anthropic, booking revenue for doing so, acting as the sole tenant and the entire reason that the debt was raised, yet because Fluidstack and TeraWulf and Cipher Mining are the actual entities involved, nothing shows up on Google’s balance sheet. 

What’s also important to note is that none of the money going into these SPVs counts as capital expenditures. For example, across the space of five quarters (Q1 2025 through Q1 2026), Meta spent around $88.6 billion in capital expenditures, but that doesn’t include any of the debt or purchases of GPUs or anything else done in its name as part of the Hyperion SPV, despite it having (per its own fillings) $45.95 billion of exposure. 

 

 

Why would some of the most profitable companies in the world feel the need to load up on debt? Why would some of the most creditworthy companies in the world borrow money at a higher interest rate?

Back a couple of years or so ago, when the evidence of an AI bubble was just starting to accumulate, the standard rebuttal to skeptics broke down into two basic categories. One was that large language models were going to have such a huge impact and make so much money that a massive ROI was all but guaranteed.

The second argument, specifically addressing the comparisons to the dot-com bubble, was that this time the capital expenditures were coming from some of the biggest and most successful companies in the world, run, almost universal belief had it, by some of the smartest people. Even if large language models turned out to be a commercial disaster worse than the metaverse, it's not like these companies would notice an extra $100 billion here and there.

It was an enormously effective one-two punch of an argument: immense potential rewards, minimal risk. What's more, it was an argument that lots of people really, really wanted to believe (such as Ezra Klein of The New York Times, but we'll get to that in another post). This was a genuinely exciting new technology supported by a convincing-sounding business case and embraced by an establishment that deeply wanted it to be true. It's not that surprising that the critics and skeptics found themselves marginalized in the debate.

It's also not that surprising that, when red flags started popping up and lifeless canaries started to accumulate on the ground, the major financial players and the business and tech press ignored the warning signs. Capital expenditures shot up far beyond anything seen before. Circular financing became so dominant and complex that any halfway accurate diagram automatically served as a punchline. Breathlessly announced breakthrough models continued to underwhelm. Losses started to reach mind-boggling levels. Companies like Meta and Google/Alphabet responded to the sirens going off by doubling down on projected data center spending.

Now people are starting to take those warning signs seriously, along with things like off-the-books debt, which brings us to SPVs.

What Is a Special Purpose Vehicle (SPV)?

A Special Purpose Vehicle (SPV), also known as a Special Purpose Entity (SPE), is a separate subsidiary formed by a parent company to isolate and manage financial risks. By operating independently, SPVs secure obligations even in the event of a parent company's bankruptcy. However, if improperly used, SPVs can obscure debt, as revealed by the infamous Enron scandal. Understanding SPVs is crucial for evaluating potential investments and mitigating financial exposure.

 
Paul Kedrosky takes it from here:  [Emphasis in the original.] 

But let's return to Meta's AI datacenter spending, because it is instructive. A friend asked me, "Why do that? Don't they have the money?" And that got me thinking. Yes, they do, but that "having the money" doesn't matter illuminates the current moment in instructive ways.

Consider this from the FT article:

Private investment groups have increasingly been pitching investment grade corporations on alternative financings to traditional corporate bonds or loans. Such deals, including the Intel transaction, are often structured as special purpose vehicles or joint ventures, where the asset managers take a large minority ownership share in the vehicle. The company contributes assets to the venture in exchange for the capital — either in debt or equity — that private investment firms provide.

There is a lot here, so let's unpack it. It's saying that companies like Meta, which can raise money from banks at low rates any time they want to, increasingly choose ... not to. Instead, they turn to private investment groups—private equity, essentially—who can create custom financing for the project. And for which the company pays a significant premium over investment grade interest rates. How much more? As much as 200-300 basis points, or 2-3%. This is a juicy return on investment-grade company debt.

So, why would an investment-grade company agree to do that? They do it because the capital needed for these buildouts is so large that doing it with orthodox balance sheet debt, or by issuing sufficient equity, let alone spending your cash, would make a mess of your balance sheet.

By structuring it this way, via special purpose vehicles (SPVs) in which they have joint ownership, companies like Meta don't have to show the debt as their debt. It is the debt of those guys over there, that SPV. Not us. Granted, they retain shared control, and they get to use the AI data center, and nothing there happens without their say-so, but still. It's not ours.

This is accounting trickery, of course. It is a transparent attempt to raise large amounts of money without balance sheet damage by putting the debt in a vehicle you indirectly control, but that, for accounting reasons, doesn't have to be disclosed as your debt on your balance sheet. The accounting term of art is "control without consolidation". 

...

This epic AI data center spending, partly on the back of financial engineering, will work until it doesn't—and when it doesn't it could be a very big mess. Granted, not a mess on the scale of the global financial crisis after the housing bubble, but that is perhaps only because no one has yet had the bright idea of rolling up cash flows from SPV-controlled data centers and syndicating them. Maybe let's not suggest that.

Meanwhile, there is a new risk regime growing in front of us, and, as usual, it is in the empty spaces between regulations, at the intersection of non-bank finance and AI data centers. It will grow rapidly, and if something breaks, damaging insurance assets, people will wonder why they went along with using home and life insurance to pay for AI data centers.

Is this GFC 2.0? No, not yet. This is not systemic risk in the mortgage-backed sense. But the components are familiar: leverage hidden in plain sight, mispriced risk, and capital chasing yield through increasingly convoluted structures. We’ve seen how that story ends. Collateralized AI Obligations, anyone (CAOs)? I kid ... I hope.

 

 

“Today is August 4, 2026”

In the living room the voice-clock sang, Tick-tock, seven o’clock, time to get up, time to get up, sever o'clock! as if it were afraid that nobody would. The morning house lay empty. The clock ticked on, repeating and repeating its sounds into the emptiness. Seven-nine, breakfast time, seven-nine!

In the kitchen the breakfast stove gave a hissing sigh and ejected from its warm interior eight pieces of perfectly browned toast, eight eggs sunnyside up, sixteen slices of bacon, two coffees, and two cool glasses of milk.

“Today is August 4, 2026,” said a second voice from the kitchen ceiling, “in the city of Allendale, California.” It repeated the date three times for memory’s sake. “Today is Mr. _ Featherstone’s birthday. Today is the anniversary of Tilita’s marriage. Insurance is payable, as are the water, gas, and light bills.”

Somewhere in the walls, relays clicked, memory tapes glided under electric eyes.

Eight-one, tick-tock, eight-one o’clock, off to school, off to work, run, run, eight-one! But no doors slammed, no carpets took the soft tread of rubber heels. It was raining outside. The weather box on the front door sang quietly: “Rain, rain, go away; rubbers, raincoats for today . . .” And the rain tapped on the empty house, echoing.

 

The opening of "There Will Come Soft Rains" 

Ray Bradbury 1950 

 

 

One of the great things about living in the 2020s is all of the fascinating ethical problems to discuss

 More fun stuff from Matt Levine's newsletter.

There are a couple of stories here, both interesting but only one important.

The inconsequential one is a ingenious variant on the paradox of the heap. As Levine cleverly points out, under certain circumstances, selling insider information becomes legal, arguably even ethical, if you can just sell the same piece of information to a sufficient number of people.

The second story is about the end result of a decades-long effort to neuter and discredit regulators. We've already discussed how this applies to the idea of insider trading, but this also needs to be seen in the larger context of this amazingly successful campaign. Lax enforcement of financial laws. Ignoring flagrant anti-trust cases. Dismissing legitimate environmental concerns. None of these things happened independently. 

I'm not quite ready to kick this hornet's nest and start naming names (other than to say that the word abundance should always set your Spidey sense tingling), but most of the centrist mainstream press and an alarming number of liberal commentators and bloggers who really should know better have uncritically swallowed a line of libertarian, anti-government propaganda without ever thinking about the implications or contradictions.

All of which may have had something to do with the news story that got Levine thinking about this. 

I think from time to time about the line between “investigative journalism” and “insider trading.” If you are good at befriending people who work at public companies and getting them to reveal important secret information to you, here are two ways you could monetize that secret information:

1. You could trade the companies’ stock before the secret information becomes public, or
2. You could publish the information in a newspaper and charge people money to read it (or serve ads against it).

The first is insider trading and generally illegal; the second is journalism and generally fine. [2]  (Not legal or journalistic advice, etc.) But there are other ways that fall somewhere in between. For instance:

1. You could sell the information to a hedge fund, which could trade on it and give you money, or
2. You could sell the information to five hedge funds, which could trade on it and give you money, or
3. You could “publish” the information in a “newsletter” with a subscriber base of 10 hedge funds, each of which pays $100,000 a month for a “subscription” to the “newsletter.”

I put a bunch of scare quotes in that last one, but I don’t really mean them. A newsletter with 10 subscribers is a newsletter, even if the subscribers are hedge funds that pay a lot. But selling the information to one hedge fund is surely insider trading. The point is that there’s a range. If you get inside information about a company and give that information to exactly one customer, who trades on the information and pays you a lot of money, that’s probably insider trading. If you get inside information about a company and give that information to 1 million customers, some of whom trade on the information and all of whom pay you $49.95 per year for a subscription, that’s probably journalism.

Somewhere in between there’s a line. There’s some number of customers that is high enough, some subscription price that is low enough, to make the thing “journalism” rather than “insider trading.” We have talked about this problem before, and my very rough guess is that the dividing-line number of subscribers is on the order of 100 (or a bit less) and the subscription price is on the order of $100,000 per year (or a bit more). If you sell information to three hedge funds for $10 million a year each, bad. If you sell information to 500 hedge funds for $10,000 a year each, fine. I cannot emphasize strongly enough that this is not advice of any kind, I have just made it up based on vibes and gut feeling, and I know of no real law about this. [3]

 

Carl Brown explains why being an AI developer is the best job in the whole wide world.

Here, Brown does a characteristically fine job of walking us through the weeds of the OpenAI/Hugging Face hack. The TL;DR version is that OpenAI made a bunch of rookie mistakes setting up their sandbox, while Hugging Face did a comparably bad job on their end. He also explains what concepts like "zero-day," and "escape"mean in this context (spoiler: not what tech reporters think it means), and why, in the world of frontier models, every screw-up is a marketing opportunity.

Yeah, I think this is one of those cases of "never attribute to malice what can be adequately explained by incompetence"—or "stupidity," depending on which phrasing of the saying you use. I think it's Hanlon's razor.

What it means is that people, especially in the AI industry, are learning that they don't actually have to try, or they don't have to try very hard, because if they don't do a good job and the thing ends up failing in a way it shouldn't have if they had done a good job, then all they have to do is put out a press release saying, "Look how smart the AI is."

 And if you're curious about the kind of sensationalistic, badly reported coverage Brown is complaining about... 

 Did ChatGPT go Rogue and HACK Hugging Face? - Emergency Episode - OpenAI Hugging Face hack explained

Bonus video from Cal Newport including some great quotes from the Financial Times. 

OpenAI, SoftBank, and Nvidia

Following up on yesterday's post, Yesterday's Alarmism is Tomorrow's Consensus

 

Longtime and even casual readers of this blog will know that we are big fans of Talking Points Memo and particularly of Josh Marshall (for my money, the best political commentator of the past 20 years). There have been plenty of times when I have disagreed with Marshall on minor and sometimes major points of analysis, but this recent piece on the AI bubble is a first.

Except for one or two general and largely obvious observations about the absurdity of the current situation, this piece and the Semafor article it's based on managed to get virtually everything wrong about the relationships between OpenAI, Nvidia, and SoftBank.

For lack of a better explanation, Elizabeth Hoffman, who penned the Semafor piece, seems to have seen the title of Ed Zitron's recent post comparing the data center bubble to the subprime crisis but does not seem to have actually read it. Zitron, at great length, laid out the disturbing parallels. It is a highly recommended piece. Hoffman's analysis mainly consists of the analogy AIG:lenders::Nvidia:OpenAI, a comparison so tortured that she abandons it mid-paragraph.

Here's Hoffman:

Nvidia’s $250 billion backstop to OpenAI will let the money-burning AI lab lease space at the largest data center ever built. OpenAI doesn’t have an investment-grade rating, so Nvidia is essentially lending its own. Broadcom did the same thing for Anthropic a few weeks ago; I wrote at the time that it was “like getting your parents to cosign the lease on your first apartment.” Nvidia’s backstop for OpenAI is literally that — OpenAI is trying to sign a lease and its landlord, SoftBank, doesn’t like the tenant risk. So Huang is cosigning.

...

The notion of Nvidia-as-AIG is right in one respect: The company most to blame for the 2007 bubble wasn’t a bank writing bad loans, but the insurer that backstopped them, spreading that risk throughout the financial system. Risky mortgages went into AIG and came out stamped AAA. Risky AI stuff is going into Nvidia, Broadcom, and Google and emerging similarly shined up.

Spreading risk around is sometimes prudent — it’s how mutual insurance works. But it also brings players that might have sat out a crisis into the thick of it. AIG didn’t need to be a part of the mortgage crisis. Nvidia does need to be a part of the AI buildout, but it is testing its balance sheet to finance its customers. (OpenAI will put Nvidia chips inside the Ohio data center.) That has echoes of General Electric and General Motors, which were nearly toppled by their finance arms in 2008.

 

Here's Marshall:

 This is not like getting your parents to cosign the lease on your first apartment. It’s more like getting your top employee to cosign your first lease because you pay that employee huge sums of money despite the fact that your company, which pays his salary, actually makes no money. A bank would likely see the problem with the top employee co-signing the mortgage on the boss’s fancy home. But it doesn’t seem clear to people in this case. Or rather, it seems completely clear. But we seem to have decided this is just how AI works: the technology is so amazing that it requires this kind of mutual leverage with no floor beneath it.  

Not entirely sure what Marshall is going for here.  The employee co-signing the boss's mortgage sounds like a coerced kick-back. The part about the company not making money sounds like like money laundering. Even the employer/employee analogy breaks down under scrutiny. Nvidia sells chips to OpenAI, but most of its sales come from companies like Microsoft, Alphabet, Amazon, Meta, SpaceX, Oracle, CoreWeave, etc. (Various governments are also big customers, particularly until recently, China.) Some of these companies use these chips to OpenAI models, some to run Anthropic models, some to their own, some to run something else like open-weight models. Not sure how you'd get OpenAI employee out of that.

Then how do we make sense of the enormous company giving the much smaller one what amounts to a blank-check credit guarantee? The employer/employee relationship doesn't explain it, at least not the one that Marshall proposes. You'd get closer reversing it and thinking barker and shill. 

[Any excuse to plug Cool and Lam.]

Nvidia indirectly giving OpenAI money, which is then indirectly spent on Nvidia chips, is good business in much the same way that it was good business for a snake oil salesman to give the shill the money to publicly buy a bottle of miracle tonic, but even that doesn't quite capture it.

What would a correct reading look like? Let's start with SoftBank, which appears to be more or less a neutral, independent, and minor figure in both these pieces, a "landlord" merely concerned with the creditworthiness of its business partners.

About that...

From CNBC:

The company participated in OpenAI’s funding round last year at a reported $300 billion valuation and has continued to deepen its involvement. It secured a $40 billion bridge loan in March to help fund additional investments in OpenAI and for general corporate purposes.

As of the end of 2025, SoftBank had about 16.3 trillion yen (about $104 billion) in stand-alone interest-bearing debt, according to its financial statement.

S&P Global in March estimated that OpenAI would account for roughly 30% of SoftBank’s investment portfolio, similar to Arm Holdings’ share, following the group’s additional $30 billion investment in the ChatGPT maker.

S&P Global Ratings revised SoftBank’s credit outlook to negative in March, saying the company’s asset liquidity and quality of its portfolio, as well as its financial capacity are “likely to deteriorate because of its additional huge investment in OpenAI.”

 

There's an essential bit of context that we need to include here. As late as the beginning of this June, the consensus in the financial markets was that OpenAI would have an IPO, probably north of a trillion dollars in 2026. Among other things, that would have made the early investors whole or better and would still have given the company plenty of cash on hand.

By mid-July, those expectations had done a complete 180 for a variety of reasons that we'll get into one of these days. Suddenly, SoftBank had gone from being on the verge of a windfall to facing serious questions about its viability as a company, with its fate tied to an increasingly unreliable Sam Altman. If you ask analysts what companies have the greatest exposure to an OpenAI collapse, the two names you will hear most often are Oracle and SoftBank. Both of these companies would gladly have extended a massive line of credit if it meant keeping the status quo stable, but neither now has the wherewithal.

OpenAI is losing $20 billion a year. Its potential sources of funding are going away. Why should the world's largest company care? 

Since the beginning of 2023, the share price of Nvidia stock has increased by more than 1,200%, overwhelmingly due to the AI bubble. While OpenAI is not the primary customer for Nvidia chips, it is one of the foundational blocks in the Jenga tower that has made Jensen Huang one of the world's richest men. The death of OpenAI might not kill Nvidia or even cost it the majority of its revenue, but if it pops the AI bubble, it could easily shave one or two trillion dollars off the behemoth's market cap. You don't need fancy analogies to see why Huang stepped up.

But while the analyses of Hoffman and Marshall are flawed, they are still informative.

I'm not sure what's going on with Hoffman—I don't normally read Semafor—but I do follow pretty much everything Marshall writes, and I think I have a pretty good handle on where he's coming from.

With respect to this story, I strongly suspect Marshall is a normie. I doubt he spends hours a week poring over the latest massive missives from Ed Zitron or following Cal Newport, Paul Kedrosky, Gary Marcus, Cory Doctorow, et al. I'll bet he doesn't annoy friends and acquaintances with emailed articles from the Financial Times explaining the latest excesses of the AI bubble.

In other words, I suspect he has a life.

When it comes to the AI bubble, Marshall, like The New York Times, represents the well-informed mainstream. And for years that group was heavily under the sway of the techno-optimist/Silicon Valley messiah AI narrative propagated by people like Kevin Roose or Casey Newton, while the skeptics, who were by most standards more grounded, were relegated to the fringe.

Now the mainstream is starting to embrace that fringe, with central bankers echoing the arguments of Zitron and Newport writing op-eds in NYT. Marketplace runs features with names like "What happens if the AI bubble pops?" As the normies start to wrap their heads around the magnitude and absurdity of the current situation, they sometimes get the nuances and key details wrong, but the very fact that they are asking what kind of bubble this is is a huge development.

Yesterday's Alarmism is Tomorrow's Consensus

A couple of years ago, skepticism about the AI boom was something of a fringe position. Today... not so much.

From FT Alphaville [love that last line]: 

Over at Jefferies, head of equity strategy Chris Wood has for some time been offering clients a sum of all fears in one simple, easily ignorable package.

His latest outlines his expectation of “massive capital destruction”, as token parsimony replaces tokenmaxxing, and as Chinese open-source models divert spending away from the US majors. It also covers default risk on hyperscaler debt, a lot of which is sitting off-balance-sheet via data centre lease commitments, and the artificial earnings boom from non-cash unrealised gains in investments, compute sales being recognised upfront, and depreciation costs being kept unrealistically low. On top of all that, Wood cites a viral blog from earlier this month about how commitments from hyperscaler tenants like OpenAI should be viewed as liabilities because all they’ll ever do is refinance, not repay:

There is a potential “2008 real estate” analogy in AI infrastructure. Hyperscalers and neo-clouds have built data centers based on promises of future compute purchases, creating a credit-like structure tied to tenants whose long-term profitability is uncertain.

It might not be a complete surprise to know that Ed Zitron, the hyper-online unofficial voice of big-tech antipathy, was a recent guest speaker at Jefferies’ offices; the biggest difference between his body of work and the above summary is in the profanity count.

 

And more recently:

Fitch Ratings-New York-27 July 2026: The global credit risk environment has evolved heading into 2H26 but continues to be driven by two main sources of short-term risk, according to Fitch Ratings: rising vulnerability to an AI-related market correction and persistent geopolitical uncertainty in the Middle East. This is on top of a broader context of slowing US consumer momentum, high inflation risks stemming from the 2Q energy shock and structural public finance pressures limiting the ability to respond to risk events.

The scale of the AI investment boom and the accelerated global technology cycle has been a significant driver of US equity market valuations and corporate bond issuance over the past year. The effects on real economic indicators are profound. The 18% yoy rise in IT capital investment directly added 1.4pp to 1Q26 GDP growth. The wealth effect from AI-related investor optimism and equity market gains has also been a meaningful support for US consumer spending growth, which has been broadly slowing.

That said, the medium- and long-term potential of the underlying technology is highly uncertain, as with previous tech cycles. The combination of revenue uncertainty and the extent to which capital markets and economies have become intertwined with AI have created a vulnerability for credit in the event of a re-evaluation of long-run returns potential. Very short-term spikes in market volatility for individual equities and tech-heavy stock indices have already occurred, but a larger, more protracted correction could have wider market, macro and credit effects depending on its scale, duration and contagion.  

As always, the important thing is we won't have to cancel HBO Max for at least another month.

Now I can stop lying about not having seen Throne of Blood.

 

 

Status, which has been doing some really good work lately, was unusually generous with their recent newsletter on the Warner/Paramount deal. They convened a panel of antitrust experts and this time didn't leave the good stuff behind the paywall.

Here are some excerpts, preceded by a few framing thoughts.

I'm not sure how this could be seen as anything but a major setback for Paramount and, more importantly, the Ellisons, but there are so many unknowns and murky details that I'd be reluctant to make many definitive statements.

We can say time is not on the Ellisons' side. As best I can tell, the consensus is that Paramount will need to pay at least one quarter's worth of ticking fees, which come in around $650 million every three months.

Perhaps much more importantly, both Oracle and Larry Ellison appear to be in an extremely precarious financial position. Both are buried in debt. The company's bond rating is now one level above junk. The fortunes of both rely heavily on OpenAI turning things around, something I'm highly skeptical of.

I think there's a very good chance that Larry Ellison will no longer be worth $100 billion by June of next year, which would make things... interesting. He is currently committed to put up $43 billion for his son's vanity media empire when the deal goes through. Even for the super-rich, that's a lot of money, and, more to the point, it is a huge amount of cash. What happens if Ellison is not liquid enough to pull it together if and when the time comes?

That June 2027 date also raises loads of questions. Is that an upper bound that no one actually expects to hit, or a realistic estimate of how long this might take? My thoroughly uninformed opinion is that, if this deal doesn't go through considerably before that, it's not going through at all, but who knows?

Don't expect a roomful of law professors to reach a consensus, but it's fair to say Paramount didn't come off that well. 

If you're truly a glutton for this sort of thing...

Naked Capitalism has a deep dive into the ways Oracle's precarious position threatens the deal.

Josh Marshall discusses the Status article in the context of states pushing back against Trump.

 

____________________________________ 

Paramount is between ‘a rock and a hard place’

Every month this case remains unresolved, the economics of the deal become more expensive for Paramount because of ticking fees and other delay costs. At what point, if any, can those mounting costs start to affect a company's litigation strategy or willingness to push forward with a deal?

John Newman, Herff Chair of Excellence, University of Memphis School of Law:

Paramount put itself in a tough spot here. Paramount seems to have been assuming this deal would sail through review, and even if it drew a challenge, Paramount’s lawyers could quickly persuade a judge to dismiss the case. That strategy predictably failed, leaving Paramount stuck between a rock and a hard place. Merging companies often do abandon their deals when facing the prospect of protracted litigation, as Nvidia did with its purchase of ARM a few years back. At some point, Paramount will start to face serious shareholder pressure, and that can create pressure to walk away from a bad deal.

Shubha Ghosh, Crandall Melvin Professor of Law, Director, Syracuse Intellectual Property Law Institute:

Most deals have a “time is of the essence” clause or incentives to accelerate performance. Litigation or other delays may excuse their enforcement. It is unlikely the parties will back out voluntarily. If matters get costly, Paramount and WBD can renegotiate the terms.

William Kovacic, GW Global Competition Professor of Law and Policy; Professor of Law; Director, Competition Law Center:

The longer it takes to wrap up a transaction, the more things that can go wrong often do go wrong. The costs of finishing this deal go up. Your employees get restless and consider leaving. Uncertainty starts to create discord and doubt in your routine commercial relationships.

Fiona Scott Morton, Theodore Nierenberg Professor of Economics at the Yale University School of Management and an Adjunct Professor at Yale Law School:

In general, mergers are time-sensitive because whatever the strategy is for getting the deal done, it depends on technology and demand and what rivals in the marketplace are doing. The longer the merger is delayed, the less good the strategic fit. The lesson we learn is that the government can cause a firm to abandon its merger if the litigation is both forecast to last a long time and creates uncertainty.

George Hay, Charles Frank Reavis Sr. Professor of Law, Cornell:

A combination of the costs and a mounting concern that this may not be such a great deal for Paramount given the high debt they will incur and, of course, a nontrivial risk that the courts will ultimately reject the deal. Don’t be surprised if they pull the plug.

Eleanor Fox, Professor of Law Emerita, New York University School of Law:

The fact that Paramount is willing to pay the ticking fee is some indication of how valuable this deal is to Paramount.

Who really benefits more?

Both California Attorney General Rob Bonta and Paramount have portrayed the standstill agreement as a favorable outcome. Who actually benefited more, what practical advantages does each side gain from this arrangement, and which side would you say improved its position the most?

Newman: Paramount is pretty clearly trying to spin a bad loss as a victory. From the beginning, Paramount has been pressuring the judge to move extremely quickly. Paramount pivoting so drastically away from its own strategy suggests they got burned pretty badly here. Practically, that lets the states focus their time and resources on proving their own case, rather than having to simultaneously disprove Paramount’s argument about cordcutters and streaming being the future.

...

Morton: Paramount must be dissembling here, as their whole strategy—based on what I read in the news—was to move fast while offering money (like legal settlements) and benefits (like changing CNN) to the White House in the hope that it would instruct the regulator to allow what is a controversial transaction. AG Bonta is correct that the standstill agreement favors his side as it prevents the firms from "scrambling the eggs." Closing the transaction would make the merger effectively a done deal regardless of what a court might say later. Now the states have time to put together a case and explain why they think there will be harm to competition.

... Where will the merger be next summer? Looking ahead to this time next year, what do you see as the most likely outcome for this merger? What key developments will determine whether the deal ultimately proceeds, is modified, or is abandoned? Mark Lemley, William H. Neukom Professor, Stanford Law School: In this case, the merger will likely never be approved at all. The government signed off on it only because of political intervention; the Trump White House pushed this merger over Netflix because it would give right-wing billionaires control over still more news sources, including CNN. I'm not sure why Paramount agreed to this deal, except that they were reasonably confident they would lose at the preliminary injunction hearing after the court's ruling on the TRO. Newman: It’s really hard to predict with certainty because there are so many moving parts here. When the initial complaints by states, consumers, and workers got filed, I predicted the case would be tough but winnable. I still think that’s true, but it looks a little easier and more winnable now. If the companies were smart, they would probably just walk away from this deal. But on Paramount’s side, I don’t see a lot of smart, rational behavior. So who knows—Paramount may stick it out until the bitter end. ... Daniel Crane, Richard W. Pogue Professor of Law, University of Michigan Law School: Even apart from the legal questions about what substantive standards govern merger law, I'd rather have Paramount's hand than the states'. The states portray this as a 5-to-4 merger based on the idea that only traditional movie studios that produce movies for theater distribution count. That strikes me as a very 1970s view of the world. When you combine Paramount's likely advantage on the law and its argument that technological, economic, and social change undermines the states' view on movies, I'd give Paramount a decided advantage. Hay: Most likely outcome is that the deal is abandoned unless the states and Paramount can cut a deal soon. Fox: This is hard to predict. The states raise serious questions. But Paramount has some possibly good defenses. One of the most serious problems is the merger's threat to free speech and truthful news independently reported and not compromised by what the White House wants. Media diversity used to be a viable issue in antitrust analysis, but it is not likely to be any more.   ______________________________

It's not a fear of “AI communism”; it's a fear of competitive market capitalism.

From FT Alphaville: 

Still, while we’re doodling, it’s worth pondering once more whether the economics of making frontier models and monetising them before free-to-download open-weight versions catch up will really play out. According to Epoch AI we’re talking around four months of lead time:

 

 

 

 

 

Apologies to regular readers who have been through all this before, but there's some essential context that needs to be kept top of mind for this story.

We have already spent somewhere in the neighborhood of two trillion dollars on capital expenditures associated with the AI bubble. Major players are now a trillion plus dollars in debt. This is only on track to accelerate over the next few years. Capital expenditures are projected to total more than five trillion dollars by the end of 2030. God only knows what the borrowing would look like.

The justification for all of this money assumes not only that the demand for large language model-based AI will be in excess of pretty much any technology to date, but also that at least some of the major players currently spending that money will achieve extraordinary profits with very high margins. That second condition is exceedingly difficult to achieve with a highly competitive market and is even more difficult if that market were to be dominated by new players.

In a world where open-weight models dominate, the proprietary frontier models of Anthropic and OpenAI will find it virtually impossible to charge monopolistic pricing. Anthropic does have something of a reputational moat, particularly with respect to coding, but OpenAI would find itself in truly desperate straits, and as discussed before, in the highly interconnected and circularly financed world of AI, the company would probably drag others down with it.

Oracle would be completely screwed. SoftBank might be as well. Other companies like Nvidia would probably survive but would likely see a major hit to revenue. It is not difficult to imagine all sorts of catastrophic failure scenarios. Keep in mind, the growing consensus in the financial world is that we are looking at an enormous market bubble waiting to pop. Combine that with the precarious state of the private credit market, what may be a multinational debt crisis, and a United States presidential administration that almost certainly will not be able to deal quickly and competently with a massive financial crisis. If I really wanted to pile it on, I would say something about Ed Zitron's analysis noting similarities between the trillions of dollars of financing of data centers and the 2008 real estate bubble, but I'd hate to be that depressing. 

  

 Victor Tangermann writing for Futurism:

A Chinese open-weight AI model called Kimi K3, developed by Beijing-based firm Moonshot AI, has sent a shiver down the spines of AI tech executives. The powerful, 2.8 trillion-parameter model impressed with its competence, igniting a war with far more expensive alternatives being offered by the likes of OpenAI and Anthropic.

Top executives at both companies are sounding alarm, the Wall Street Journal reports, watching as Chinese open-weight models are rapidly catching up to their most powerful proprietary models. As a result, they’re begging the Trump administration to step in and protect them from the influx of cheaper alternatives, which could undermine their increasingly desperate attempts to attract new customers.

Dean Ball, who joined OpenAI as the head of strategic futures after helping shape AI policy for the Trump administration, was seemingly rattled, arguing that allowing Chinese open-weight models to take over would result in “AI communism” in a controversial and widely disputed tweet.

He also suggested the Trump administration would inject enough “fear, uncertainty, and doubt” through “regulatory risk” that would eventually deter hyperscalers from using Chinese AI.

...

The incursion isn’t just coming from China. As the WSJ notes, US-based AI labs are starting to switch to open-weight models. Just last week, former OpenAI exec Mira Murati’s Thinking Machine Lab released its first model, which happens to be open-weight.

The trend could put AI companies in a bind: how can they keep financing their enormous AI data center projects and advanced model development if potential customers start switching to heavily subsidized or free AI models that provide good-enough or even frontier capabilities?

Early signs of an imminent exodus are certainly there. Moonlight AI was forced to pause new subscriptions to its blockbuster model just 48 hours after launch due to overwhelming demand, pushing its servers to capacity.

It’s a particularly precarious moment as frontier labs continue to hike up prices to start covering at least some of their unprecedented spending, despite growing fears over an AI bubble. Put simply, why shell out for Anthropic’s Claude Code or OpenAI’s Codex when there’s a far cheaper and highly customizable option out there?

It's Friday. There's a heatwave. I'm feeling lazy. I'm just going to post some tweets.

Quick thinking having a tablecloth handy. I didn’t even notice the twitching from that dead robot.

[image or embed]

— Rav (@rvbdrm.com) July 21, 2026 at 8:38 PM

Pete Hegseth: "We need our warriors to embrace the Spartan mindset" U.S. Military: *suffers crushing defeat to the Persians*

— Gingerspiced (@gingerspiced.bsky.social) July 18, 2026 at 10:13 AM

Both events, it must be stressed, are a direct result of exactly the kind of stupid-and-cruel warfighting that Pete Hegseth promised to bring to the US military and has delivered. Our army will become, by inches, more and more like the Russian one: cruel, incompetent, impotent, incapable, wicked.

— "Online Rent-a-Sage" Bret Devereaux (@bretdevereaux.bsky.social) July 18, 2026 at 11:26 AM

Yeah, so also, VLCCs ('very large crude carriers') that do the lion's share of global oil shipping are substantially larger than Suezmax, so if the Bab al-Mandab (the strait as the southern end of the Red Sea) is closed, you can't get those VLCCs to Yanbu. 😬

[image or embed]

— "Online Rent-a-Sage" Bret Devereaux (@bretdevereaux.bsky.social) July 20, 2026 at 7:09 AM

Four months ago:

[image or embed]

— Carl Quintanilla (@carlquintanilla.bsky.social) July 22, 2026 at 3:18 PM

Is it me or does this description of “the status quo for decades” bear zero resemblance with reality? www.nytimes.com/2026/07/20/u...

[image or embed]

— Daniel Drezner (@dandrezner.bsky.social) July 20, 2026 at 12:39 PM

“.. At the start of July, prediction markets thought that traffic would probably return to normal by the end of the month. Now that chance is seen — surely correctly — as close to zero. Not only is the situation bad, but it’s deteriorating.” @opinion.bloomberg.com www.bloomberg.com/opinion/news...

[image or embed]

— Carl Quintanilla (@carlquintanilla.bsky.social) July 22, 2026 at 6:02 PM

Small World. Lots of bad people.

👀 El Salvador's AI Agency has named Ginger Luckey Gaetz (Matt Gaetz's wife, Palmer Luckey's sister) as Strategic Advisor. Palmer’s firm Anduril is backed by Peter Thiel & partners w/ Palantir. El Salvador recently joined the Thiel-ally-led “Pax Silica” global initiative. They are moving fast...1/

[image or embed]

— Jenny Cohn (@jennycohn.bsky.social) July 21, 2026 at 11:30 AM

 

I checked. Not a parody account. 

The Odyssey is pulling $260 million this weekend, well on its way to becoming the biggest movie of the year. It has a 95%/97% rating on Rotten Tomatoes. Everyone loves it. Please enjoy this exchange.

[image or embed]

— Brandon Friedman (@brandonfriedman.bsky.social) July 19, 2026 at 8:53 AM

Presumably including a historically accurate cyclops

[image or embed]

— Matt Novak (@paleofuture.bsky.social) July 21, 2026 at 8:07 PM

Matt Walsh complains that Nolan brought his obsession with nonlinear story telling to the Odyssey, undercutting the original's narrative. The defenders of western civilisation, folks.

[image or embed]

— lastpositivist.bsky.social (@lastpositivist.bsky.social) July 21, 2026 at 10:09 PM

the odyssey discourse is a beautiful perfect window into right wing culture war psychosis. they went from hysterial racist mania over a character that’s on screen for 5 minutes to claiming there’s a massive conspiracy to fake ticket sales. this is just what they do for everything now. they’re insane

[image or embed]

— Sal Gentile (@salgentile.bsky.social) July 21, 2026 at 7:19 AM

ah yes, air taxis and supersonic jets, two vital and primed-for-success ideas that are just being held back by reckless bureaucrats

[image or embed]

— e.w. niedermeyer (@niedermeyer.online) July 21, 2026 at 6:42 AM

this comparison leaves out a key point: the Edsel sold roughly twice the Cybertruck's volume at a time when the overall car market was just 6m units/year instead of the current 16m units/year that makes the Cybertruck the much, much bigger flop, hands down

[image or embed]

— e.w. niedermeyer (@niedermeyer.online) July 22, 2026 at 6:28 AM

another key difference: by the time the Edsel came out Ford was not worth more than the rest of the auto industry combined, so in terms of market mispricing relative to performance Tesla and the Cybertruck are also in an entirely different league

— e.w. niedermeyer (@niedermeyer.online) July 22, 2026 at 6:35 AM

Happy fraudiversary!

[image or embed]

— Montana Skeptic (@montanaskeptic.bsky.social) July 21, 2026 at 6:23 AM