If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large (updated to version 3.0 a few weeks ago). My Hater's Guides To the SaaSpocalypse, Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater's Guide To Oracle.

This week, I published the Hater’s Guide to Softbank — a sordid tale of tech’s most degenerate gambler, who, thanks to a couple of early lucky wins, has managed to set the foundations for the AI bubble’s biggest (and possibly most gratifying) downfall. And, on Friday, I’m going to take a deep dive into the memory industry — and the reason why you can’t afford a new gaming PC. 

Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week.

---

Soundtrack: Mastodon — Streambreather

No bailouts, no handouts, no special treatment, no tax breaks, no CHIPS act, and no sovereign wealth fund. It is time to tell the AI industry to go fuck itself, because it’s effectively done the same to the rest of society. This industry is unworthy — a sham conjured up by a tech industry that’s run out of ideas, a trillion-dollars’ worth of manufactured consent and entirely-avoidable financial crises — and should not be protected under any circumstance. 

Every single time you hear somebody discuss “bailout” or “too big to fail” or “sovereign wealth funds,” know that this is the industry, on some level, attempting to create the air that it cannot die, when in fact every one of these companies is just as weak and brittle as any other startup.

I also think that the media — and the world at large — is too ready to accept the prospect of a bailout after watching those who drove the world into a ditch in 2008 escape blame, and I must be clear: the AI industry is very different to the financial industry. It is inessential to the economy, and its relevance is only as large as the hype campaign that sits behind it. 

This is an industry of losers that has inflated only because of the joint manufactured consent of Silicon Valley, the mainstream media, and an enshittified stock market that rewards grifting and circular financing. OpenAI had $5.7 billion and Anthropic a little under $5 billion in the first quarter of this year — and those revenues mostly came from companies that were burning AI tokens at a horrendous rate because they’d just been forced to pay the actual cost of AI — and now everybody’s pulling back on that spend. 

Generative AI will not bring us AGI, nor does it do much of what we associate with artificial intelligence. It is not autonomous. It is not “intelligent.” It does not have thoughts, or “knowledge,” and no matter how many layers of harnesses and scripts you put on top of it, it is still (per OpenAI) mathematically certain to hallucinate. I estimate that at least 70% of the entire AI industry’s revenues are made up of OpenAI and Anthropic’s compute spend, and as both companies are horrendously unprofitable, this means that the AI industry is, for the most part, venture capitalists funnelling money to hyperscalers so that they can funnel that money to NVIDIA or data center capex.

If this software were worthy, it would stand on its own two feet. It wouldn’t need circular financing and a cult of personality to prop it up, either. If it were truly special, there wouldn’t need to be an army of crazed acolytes that attack you for not pledging yourself to the graveyard smash. There has never been a tool or product in history sold with such hysteria and aggressive monocultural force that has ever turned out to be anything more than a grift. Some people have developed unhealthy relationships with large language models (LLMs) and the companies that make them, and that, not any certainty or proof of Artificial General Intelligence (AGI), is what motivates them. 

This software is uniquely dark, both in what it unlocks in some people through its use and in the sense of the entities that sell it. Some people are in genuine awe of each of the rotation of clammy, soulless pod-people that saunter out of Anthropic every few weeks. Each one sounds a little weirder, more cultish, more disconnected from the real world. Silicon Valley may believe itself atheistic, but Anthropic has a worrying sense of fanaticism, both in the people that work there and its fanbase. Imagine the absolute worst fanbase of a video game possible, and then add layers of financialization, grifting and high school drama laced with pseudo-religious attachment. All for a fucking app! 

Please, people. Nobody in the real world cares about “loops.” Nobody is thinking about tokenization. If you said inference to a guy on the street they’d take you to see a doctor. Nobody gives a shit. They don’t know what OpenClaw is either. Grow up. Go outside. You sound like a lunatic. Does your mother know how many Claude 20x accounts you have? It’s obsessive! 

Anyway, the only reason that AI has any presence in our economy is that Microsoft, Google, Meta, and Amazon are intent on spending more than $765 billion in capital expenditures in 2026 and a trillion more in 2027 because they have no other hypergrowth ideas, even though generative AI has yet to show any real potential as something that can drive meaningful revenues (let alone profits), as evidenced by the fact that none of these companies break out their actual AI revenues, a point I made on CNBC late last week. 

Google does not have the next Google Search, Microsoft does not have the next Microsoft Office, Meta does not have the next Facebook, and Amazon does not have the new AWS. That’s why they need you to believe that AI is a big deal without them ever having to prove why outside of capital expenditures. They want you to assume that all this money can’t be wrong, even though when you remove OpenAI and Anthropic (who represent 89% of the revenues of the largest AI companies) the AI industry is, at best, pulling in $20 billion in annual revenue.

And lord do they want you to say “it’s early,” and that it’s just like the Dot Com Bubble, all so that you’ll either accept AI as your lord and savior or, alternatively, help justify one of the largest misallocations of capital in history as “building useful infrastructure.”

Stop Pretending This AI Is Like The Dot Com Bubble

Newsflash! AI GPUs are useful for generative AI and not much else. Every “innovation” in LLMs has only been made possible by throwing billions of dollars at the problem either in headcount or compute costs — every ounce of talent in the tech industry, every bit of media attention, every dollar of capital expenditures, all focused on one industry that has successfully created LLMs that are more expensive and significantly less useful than human beings. 

The reason every AI person speaks in pie-in-the-sky hypotheticals is that the actual outcomes are decidedly mediocre when you compare them to their ruinous costs. Anthropic and OpenAI raised (assuming the rounds completely close) over $300 billion in 2026 alone, and take up the vast majority of available AI compute. They need you to speak in the future tense, because nothing — absolutely nothing — about what’s been created so far justifies even a fraction of its financial and infrastructural cost.

When the AI bubble bursts, none of this infrastructure will be particularly useful. As I said in my premium about how this is worse than the Dot Com Bubble, GPUs are not fiber optic cable, and when the bubble bursts, NVIDIA chips will either be sitting in the coffers of the largest tech companies in the world, held by asset managers, or auctioned at a steep discount by creditors. These are not going to be useful for hobbyists, nor will they be cheaper to run, nor will incomplete data centers be cheaper to finish.

The Dot Com era fiber overbuild was a result of a complete misread of demand signals, per Justin Kollar:

This continental rewiring was also justified by another powerful myth—that internet traffic was doubling every 90 days. The claim spread through analyst reports, earnings calls, and investor presentations like a particularly virulent meme. If true, it meant that demand was growing exponentially, far outpacing any conceivable supply, and that every new trench of fiber would soon pay for itself many times over.

But the mathematics were fiction. Network researchers like Andrew Odlyzko (at AT&T), looking at actual traffic data, found that U.S. backbone traffic was doubling roughly once a year—rapid growth, certainly, but nowhere near the purported 90-day cycle. Meanwhile, advances in fiber technology were making each strand exponentially more powerful. Dense wavelength-division multiplexing allowed dozens of signals to travel simultaneously down the same line at different wavelengths of light, like multiple conversations happening in different colors.

While demand doubled annually, supply expanded tenfold or more. Carriers buried the discrepancy under layers of creative accounting that would have impressed medieval alchemists. They sold “indefeasible rights of use”—essentially decades-long leases on fiber capacity—and booked the entire value immediately as revenue. They engaged in elaborate “capacity swaps,” trading bandwidth with competitors and treating each exchange as a sale, manufacturing revenue from thin air.

It’s tempting to compare this to GPUs, but it doesn’t make sense at all!  

You see, internet demand was a result of people wanting to get online and use the internet, with the leftover “useful infrastructure” having a blatantly obvious use case after the bubble burst, albeit one that took a lot longer to arrive than investors had hoped. There was no question about how that gear might be used or for what purpose one used fiber optic internet or networking gear, nor was there any question as to the underlying business model of offering an internet connection might mean. 

We were also fairly early, and internet speeds were atrocious. In 2000, only 52% of American adults were using the internet, and by 2003, that number had only increased to 61%. Per the World Bank, in 2005 only 16% of the world used the internet, and in 2024, that number had increased to 71%. When the internet was connected to via a 56k modem, access was charged by-the-minute, and obviously much, much slower than even the primitive (though expensive) broadband connections of the day. 

While we’re used to connecting at speeds that make using a web-based app near-indistinguishable from one that runs on our computer, back in 2000, 2001, or 2002, the average US internet speed was, at best, 400 Kilobits/s, or roughly 50 kilobytes a second, compared to the average US internet speed of over 200 Megabits per second, or 25 megabytes a second. 

Sidenote: Yes, fiber optic internet (and DSL for that matter) was expensive, both for the customers, but also for the providers. Verizon spent $23 billion on bringing FiOS to people’s homes between 2004 and 2010, for example, but the “up front” cost had a defined, obvious return on investment. 

Generative AI, on the other hand, is fucking everywhere, and anyone with an internet connection experiences it in effectively the same way. It’s non-consensually available in effectively every app — every Facebook, Google and Microsoft account, for example — and every media outlet known to man has mentioned AI multiple times since 2023. OpenAI and Anthropic might claim they need more data centers, but it’s unclear what “more data centers” actually achieves other than propping up NVIDIA and giving hyperscalers something to invest in. 

A lack of data center capacity isn’t holding back people from using generative AI, nor is it stopping anybody from launching a product, nor can anyone actually express what it is that they’re being built for other than “reasons for Anthropic and OpenAI to spend money.” Anthropic’s supposed lack of compute did not stop it training or launching Mythos or Fable, and when it bought hundreds of megawatts of compute from SpaceX, the biggest news was that it expanded rate limits to allow users to burn $8,000 worth of tokens for $200 a month.

Nothing about the painfully slow pace of data center development appears to be restraining a single AI company, outside of hyperscalers complaining they could’ve made more money from either Anthropic or Meta. In fact, the entire argument for more data centers appears to be “we need more compute so that people can buy it” far more than any cogent position around what these capacity shortages actually mean. 

Who are the companies lining up to spend billions of dollars of compute — or, to be more specific, spend $435 billion or more to justify the $1 trillion in GPU sales that NVIDIA claims it’ll have by the end of 2027? That’s how much demand we’ll need. As NVIDIA intends to sell over a trillion dollars of Blackwell and Vera Rubin GPUs by the end of 2027, it needs to have around (assuming a PUE of 1.35) 40GW of data center capacity built to support the 30GW+ of GPUs it will have sold. At about $12 a megawatt of critical IT (IE: the stuff in the data center that runs AI compute, and not everything else, like the cooling systems and any transmission loss), that’s $435 billion. 

OpenAI estimates it’ll spend $50 billion on compute in 2026, and Anthropic will likely spend comparable amounts. Otherwise, the only other player — outside of Microsoft, Google, and Amazon renting (or backstopping) capacity for Anthropic and OpenAI — with any meaningful compute spend is Meta (with Nebius and CoreWeave)... and Bloomberg is reporting that Meta is planning to start selling its compute because it doesn’t need all of it. 

You’ll be shocked to hear that it might be renting some of that capacity…to Anthropic.

Now NVIDIA is agreeing to financially backstop young cloud providers buying their GPUs by promising to rent back any unused capacity, yet another sign that actual, real demand does not exist at scale. AI boosters with black mold problems will say “this is just to help them raise debt,” to which I say “If the demand actually existed in any provable way, NVIDIA wouldn’t have to pay its customers to buy its products!” 

Anyway, my larger point is that there was real demand during the dot com bubble, and LLMs’ demand appears decidedly artificial outside of OpenAI and Anthropic, who cannot afford to pay without unlimited venture capital funding. 

This shit isn’t going to become magically cheaper once the bubble bursts, and considering the demand doesn’t appear to be there at scale with two-thirds of all venture capital funding focused on AI, I’m not sure what people expect to happen. Right now is the number one time in history where we should see near-infinite demand for compute across every single surface, and way more deals for compute capacity for companies other than the same four or five companies.

Right now, as I’ve discussed before, Anthropic and OpenAI take up the majority of compute, leaving the rest of the world to fight for the leftover scraps, and because data centers take 18 to 36 months to build, capacity is taking forever to come online to fill the indeterminately-large amount of demand that remains. Nevertheless, said demand can’t be that large, otherwise we’d A) have other companies trying to build their own compute (other than Poolside, which failed to raise money to do so) and B) massive remaining performance obligations — hundreds of billions of dollars’ worth — rather than the grim truth that 50% of hyperscaler RPOs are from Anthropic and OpenAI, inflating obligations by $448 billion, hiding the fact that Microsoft’s RPO growth is flat year-over-year and Amazon’s is only growing at a modest 20% when you remove Anthropic and OpenAI’s hundreds of billions of dollars’ of compute spend. Google’s is a little messier, as it’s hard to parse exactly how large its deals with Anthropic are thanks to its backstops and circular deals around Anthropic and its TPU chips.

There’s also the compelling question as to what it is that anyone would be picking up once the bubble bursts. Demand for AI services is a direct result of the entire media, tech industry and venture capital ecosystem manufacturing consent for the use of LLMs, forcing them into every corner of every experience, something that will most decidedly end once the stock market and investors cease incentivizing it. 

Once every media story isn’t about AI, once every Business Idiot with AI psychosis stops posting about it every day, when everyone stops asking about your AI strategy or wanking on about “sovereign AI,” it’ll become blatantly obvious that the actual demand for AI was not particularly strong.

We have little compelling evidence that providing any inference-based services is profitable, which means that even if open source AI outlives the frontier AI labs, it’s unclear who would actually power the infrastructure. People can come up with however many weird blogs where they’ve done some napkin maths to try and extrapolate a potentially profitable inference provider, but I’ll only believe that one is profitable when someone shows me some fucking profit.

And to be clear, without that profit, it’s unclear why anyone would offer these services at all. When you rent out a GPU cluster, you do so based on anticipated demand and the quality of service you want to provide. If you order too much, you’ve got a bunch of fallow capacity you’re paying for (and will lose money on), and if you order too little, you’ll have either unstable services or money left on the table…and even then, it’s unclear how profitable that would be. 

AI demand is, at this point, a direct result of societal pressure and non-consensually overwhelming customers with AI features. While there are people that like and pay for ChatGPT or Claude, those who do so on a subscription basis are doing so because they can get $30 to $40 of compute for a dollar. The vast, vast majority of AI compute demand is from services provided to people either for free or sold at such a massive discount that it’s impossible that anyone on a $20 or $200-a-month plan could even afford these services had they paid their actual token cost. To paraphrase Cory Doctorow, your demand is based on selling $40 for a dollar. That’s not a real business, nor is that organic demand.

One could argue that “these services will become cheaper,” but that would require them to… become cheaper. More compute isn’t (and hasn’t) lowering the cost of AI. Newer GPUs aren’t lowering the cost. Barely-tested Broadcom GPUs, Amazon Trainium XPUs, and Google TPUs aren’t lowering the costs. Even if they were to somehow magically do so in the future, what do we do with the H100, H200, B100, B200, B300 or AMD GPUs? Melt them down for scrap? Steal the RAM? Build a GPU fort? 

The Dot Com (and, by extension, telecom) Bubble was never a question of whether the internet was a useful thing that people would pay for, nor were there journalists and dodgy studies that desperately pleaded with us that AI is here, and it’s real. 

Everybody has access to AI now! They can all see it and use it if they want to, and they’ve got lots and lots of ways to pay for it! Maybe the reason that AI revenues are so putrid is that they don’t really have any reasons to pay for it, either because the free services do most of what they need (IE: google searches) or subsidized subscriptions that cost $200 a month allow them to burn as much compute whipping up HTML-based calorie tracking apps that get two users.

Every time I read somebody on Twitter say that “we’re early” or that “most people haven’t even tried agents” I feel like screaming. Motherfucker, everyone is talking about agents in every single media property all the time. AI boosters will refer to literally any AI feature as an agent, even if it’s a basic web search or generating code. The reason that most people are kind of “meh” about AI is that it doesn’t do things that they associate with AI (autonomously and automatically taking care of the things they need with little prompting or coaxing), everybody knows it hallucinates, and AI data centers are horrifying monoliths of capital that get massive tax breaks, use a ton of water, belch toxins into the air, and are being built by faceless corporations, ultra-oafs like Kevin “Mr. Dogshit” O’leary, or charmlessly damp Valley elitists like Altman and Amodei.

Every single person freaking out about “what if China does AI better than America” is living in a child’s fantasy. Oh no! China might get Mythos-level AI? Bad news folks! Anthropic itself already admitted that cheaper models — including Claude Haiku 4.5 and Kimi K2.7 — were able to identify the very same vulnerabilities as Fable (so, Mythos with guardrails). 

China has cheap power, data center capacity, and NVIDIA’s Blackwell GPUs. The thing that everybody is scared of has happened already, and you know what else happened? Nothing, because they, like American AI labs, are building LLMs. The only thing that American labs are scared of is cheaper open source Chinese models offering similar performance to their premium products, something that has also already happened. 

Remember: the only people that can afford to build data centers are either hyperscalers (that are now having to fund the buildout with debt as their cash flow turns negative), Oracle (which will die if OpenAI can’t pay it), unprofitable neoclouds, and land speculators. AI data centers are massive, expensive operations, and raising money to finish (or furnish) one after the bubble bursts will be very, very difficult.

I realize that everybody wants there to be a happy ending after all of this collapses. I get that it’s easier to think of things in familiar terms — even if said terms involved a 77% drop in the NASDAQ — because there was something good and nice at the end.

But doing so only serves to help protect the interests — and brands! — of venture capitalists, asset managers, private credit funds, hyperscalers, captured tech and business journalists and sell-side analysts that insisted on ignoring every warning sign and waving away problems by saying it was “just like Uber (nope!)” or “just like Amazon Web Services (between 2003 and 2015, Amazon spent $29.7 billion on capex, normalized for inflation),” or simply saying that “yes it’s a bubble, but bubbles lead to great industries.”

GPUs aren’t dark fiber! GPUs aren’t fucking railroads! GPUs are GPUs! They are used for basically one thing! And that one thing lacks meaningful demand outside of subsidized services and circular financing! 

And now people are discussing a bailout like this is 2008, and I must be clear how different this is, and how little it resembles the Great Financial Crisis!

AI Is Not Too Big To Fail — And You Can’t Bail It Out Anyway

The AI industry has demanded everything from us — more money than has ever been invested, more power than anything has ever needed, the stolen works of millions of hard-working creatives, so many GPUs and so many data centers that it’s causing a global supply chain crisis and a new class of RAM and storage-based inflation, the majority of venture capital funding,  and constant attention focused on an endless campaign of fear-mongering with the express intention of hyping a technology based on a mixture of mysticism and outright lies — and still, even as we enter the late innings of the bubble, it wants more. 

Capital-hog Sam Altman has floated the idea of handing 5% of OpenAI to the US government, a stake worth around $42 billion, claiming that (to quote the FT) “...giving the public a financial stake in the company is the best way to share the upside of AI,” failing to note what said upside might be, likely because there isn’t one unless “the public” refers to “the shareholders of OpenAI.” 

It isn’t clear how this would happen, outside of it requiring congressional approval as a result of the Takings Clause of the Fifth Amendment, which states that “private property [can’t] be taken for public use without just compensation,” meaning that the US government would likely have to buy the stock at whatever valuation it considered “just.” 

Yet the FT had one other interesting tidbit — that Altman is suggesting that whatever this is would “...would involve other US AI companies handing over a similar stake, although it is not clear if the other labs would be willing to do so”:

Altman and other OpenAI executives have suggested that each of America’s leading AI developers allot 5 per cent of their equity to a vehicle like the Alaska Permanent Fund, a sovereign fund that invests the state’s oil wealth into stocks and pays dividends to the state government and residents.

This is, just to be clear, not a bailout. Even though it’s blatantly obvious that Altman wants to cozy up to the Trump Administration and, he hopes, get $42 billion of funding to attach his questionably-valued quasi-startup, $42 billion is $8 billion less than OpenAI will spend on compute in 2026, and considering OpenAI has projected to burn $852 billion through the end of 2030, that 5% stake would only exist to prolong the inevitable.

You see, a bailout usually has an endpoint — a time at which the company in question no longer needs the funds. 

This Isn’t Like 2008 — We’re In A Stock Market and Data Center Speculation Bubble, And You Can’t Bail That Out

So, let’s be clear about something: we’re actually in several bubbles at once.

• The stock market bubble, where both the value of stocks and the earnings of companies in the market are inflated to an historic level.
• A data center speculation bubble, where I believe we’re building AI GPU capacity in expectation of $450 billion or more in annual data center revenue for an industry that, without two unsustainable venture-backed oafs, has a few billion dollars’ worth of demand.
• An AI startup bubble, where the vast majority of AI startups are both over-valued and have no foreseeable path to acquisition or a public offering.
  • These startups also rely on buying tokens from OpenAI and Anthropic, making them far more cash-intensive, making them absorb the majority of venture capital funding.
• A private credit bubble, where asset managers have sunk billions of dollars of pension and insurance funds into AI data centers. 
• A semiconductor bubble, where supply chains have become saturated with demand from those building AI data centers, inflating the cost of RAM and storage, making all electronics more expensive, including those inside the AI data centers, creating a vicious cycle that has doubled the cost of a gigawatt data center from $50 billion to $100 billion in a little under 10 months.

The great financial crisis, by comparison, was two major bubbles (per my piece on how AI Isn’t Too Big To Fail from a few months ago) — the over-investment and speculation on mortgages (both subprime and otherwise), and the collapse of the commercial paper (a type of loan) market that kept much of the banking system functioning, which was the real “Too Big To Fail”:

The AI bubble has made us think about corporate debt in terms of Capex — massive loans designed to bring vast data center gigaprojects to life — but in reality, a lot of it goes to small-to-mid-size businesses to cover day-to-day spending like payroll, and where the repayment terms are often measured in months rather than decades. 

These loans, issued by banks, money market funds, and other non-traditional lenders, are funded by either repo lending (asset-backed short-term deals where you effectively sell a security and buy it back very quickly at a slightly-inflated price) like Lehman used, or commercial paper — a short-term (usually a month) loan that can, in some cases (such as AIG’s!) be issued without collateral. In others, collateral can be as simple as “we have accounts receivables saying we’ll get paid.”

At its peak, commercial paper was a $2tn global industry.  

Federal Reserve Chairman Ben Bernanke noted that around the time of the bailout, AIG had $20 billion of commercial paper — short-term debt for corporations and banks where the maturity can be as low as one day, or as high as 270 days —  outstanding, in simple terms meaning it had $20bn of loans it had yet to pay within the coming year. 

Commercial paper was, at the time, often paid off using more commercial paper, and when AIG’s credit rating dropped in the middle of September 2008, it was unable to roll over its debt (by which I mean “get new commercial paper to pay off its old commercial paper”), and money market funds like Fidelity couldn’t even buy it anymore because it wasn’t investment grade, which meant that AIG couldn’t pay back its loans. 

While I won’t recount the entirety of the premium (mostly because it’s super long), AIG was deemed “Too Big To Fail” because it would’ve exploded the markets had it done so. Michael Lewitt, an economist and money manager, described a hypothetical AIG failure as being “as close to an extinction-level event as the financial markets have seen since the Great Depression” in a New York Times op-ed:

“If A.I.G. had collapsed – and been unable to pay all of its insurance claims – institutional investors around the world would have been instantly forced to reappraise the value of those securities, and that in turn would have reduced their own capital and the value of their own debt. Small investors, including anyone who owned money market funds with A.I.G. securities could have been hurt, too. And some insurance policy holders were worried, even though they have some protections.”

Yet the real “Too Big To Fail” was far quieter and more malignant, taking the form of trillions of dollars funnelled to banks:

A little-discussed part of the scale of the bailout were the liquidity mechanisms created to stop the bleeding — the Primary Dealer Credit Facilities (PDCF) and Term Securities Lending Facilities (TSLF) that provided as much as $100 billion dollars to banks and financial institutions every day. 

They existed as short-term lenders of last resort, providing overnight funding to institutions (banks, investment banks, hedge funds) that had become illiquid as their stocks tanked and their stupid, reckless bets came home to roost. The TSLF in particular helped plug the gap in the failing repo market, and I must be clear that everybody who put the US financial system in these conditions should be in prison, or worse.

The banking system ran (and still runs) on overnight facilities like the federal repo market, where financial institutions offer up collateral — like, say, mortgages — as a means of funding their day-to-day operations. Previously, money market funds were the lenders in the repo market…except they were now a little hesitant to take that collateral, which forced the government to step in with the PDCF (which traded risky, frozen assets like subprime mortgages for cash to avoid a default) and the TSLF (which traded risky bonds for US treasuries).

Absolutely nothing about these facilities or anything to do with “too big to fail” were to do with stabilizing the stock market, which was effectively cut in half, with unemployment spiking to 10%. These measures existed exclusively to protect the financial system, with only $46 billion (about 10%) focused on trying to save homeowners from foreclosure, and in the end, to quote a congressional panel from 2009, “...the panel sees no evidence that Treasury has used TARP funds to support the housing market by avoiding preventable foreclosures.” 

The Troubled Asset Relief Program (TARP) spent over $400 billion to bail out the banks, financial institutions and auto industry that would’ve collapsed as a result of an economy-wide lending freeze. Nobody went to jail, nothing really changed, and banks still don’t have to keep reserves thanks to changes made around COVID.

By comparison, OpenAI and Anthropic are systemically irrelevant, much like the rest of the generative AI industry. While their existence supports the overall symbolic value of the US stock market, their actual economic presence is minor, outside of what I estimate is around $75 billion to $100 billion of 2026 compute spend and what will likely be around $60 billion of combined revenue, with the rest of the AI industry having so little that it’s barely worth thinking about.

It’s also unclear what you’d bail out, unless the plan is to feed them capital for all eternity until they work out how to run a functional business (so, forever). Neither of them have significant debt — and Broadcom is backstopping $30 billion of Anthropic’s $35 billion TPU deal with Apollo — and their equity positions (outside of SoftBank, which I’ll get to) are only load-bearing to venture capitalists in the sense that their fund vintages will painfully sour if they’re unable to go public. 

We Should Talk About SoftBank: There is one company that is systemically dependent on OpenAI — SoftBank. As I covered in this week’s Hater’s Guide, SoftBank has wagered effectively its entire future on $40 billion or more in short-term loans to fund Sam Altman’s No IT Loads Party, and if OpenAI can’t go public, SoftBank will face a legitimate liquidity crisis. 

This, again, is nothing compared to what would’ve happened if AIG had collapsed or if the US government hadn’t propped up the liquidity of effectively every major bank. That being said, SoftBank is one of the largest companies on the Japanese stock market, and one of its largest investors is the Japanese government pension investment fund (GPIF), and thus might see some kind of bailout.

There is no avoiding the carnage to come, outside of there being somewhere in the order of ten to a hundred times the demand for AI compute by 2030 that exists today, which would require AI compute to be larger than the $779 billion that the software industry earns annually. 

There is no bailout that can reverse the trend once demand wanes for NVIDIA’s GPUs after hyperscalers reduce their capex, which will in turn kill the revenues of Taiwanese ODMs that build AI servers for hyperscalers, which will in turn kill the revenues of RAM and storage companies, which will lead to a prolonged depression throughout a semiconductor industry addicted to hopium peddled by a tech industry ruled by Business Idiots that have no idea what to do other than hire people, fire people and spend money. 

As I’ve said many times, people are conflating massive capital expenditures — invested through debt-fueled data center speculation and hyperscalers bereft of hypergrowth ideas — with real, diverse and consistent AI demand, pumping valuations based on vibes rather than reality, which means that when vibes take a violent, permanent shift, nobody has anything to point to as a means of turning people’s frowns upside down.

A sidenote on private credit: I will say that I am deeply worried about the private credit industry and its trillions of dollars of loans, as we don’t really have a firm hold on its exposure to the AI bubble, other than that some indeterminate amount of billions have been sunk into data centers.

Private credit, as mentioned, has sucked up a lot of money from pension funds, insurers, and, ironically, banks themselves who, due to the post-2008 crackdown on speculative bets, are restricted from making massive punts on AI infrastructure companies, but are (thanks to a wonderful loophole) make the same bets by proxy by shuffling cash to a private credit fund. 

The collapse in value of AI startups wouldn’t be changed by a bailout unless the US government literally invested in worthless startups as a means of propping up venture capital, and said “bailout” would number in the hundreds of billions of dollars, and while I know you’re gonna say “ohhhh Trump is so corrupt oooh Trump will do this Trump will do that,” this is not a rational or logical or even historically-accurate thing to say. 

Trump cannot simply mobilize $50 billion or $100 billion. It will go through the House and the Senate, and any bailout of the AI sector would be an incredibly-unpopular decision, infuriating not just those on the left who’ve grown tired of Big Tech, but with those Republicans that pretend to care about working Americans or fiscal probity. 

As a reminder, the first vote of the 2008 bailout failed, with Republicans and Democrats each fairly split on how they felt about the bill — and that rejection happened during a time when the US financial system was quite literally falling to shit. 

As far as the data center bubble goes, the government is absolutely willing to let unfinished or abandoned properties lay dormant. In the final quarter of 2008, 11% of US homes were empty, or 15% if you include vacation homes. 

Banks that have invested in data centers that have yet to be built (or start construction) can (and will) resell the land, though likely at a loss, and land retains value even if you haven’t built a giant warehouse full of GPUs that only lose money. There isn’t a need for a bailout here, and one won’t be forthcoming. After the Global Financial Crisis, builders were allowed to collapse to the extent that the number of construction firms halved in America between 2007 and 2012.

You could argue that Trump “will just do that this time,” or that he’ll “get a bribe” or something, but is that really the best you’ve got? Scary stories about the President? If every answer you have is “but Trump will just do it,” you’re not analyzing, you’re catastrophizing. 

And, most crucially, the vast majority of big tech will be fine, at least in the short term, when the bubble bursts. NVIDIA will likely cease being the largest company on the stock market, and the Magnificent Seven will have a dramatic fall from grace, but outside of unforeseen horrendous financial decisions, the worst I could see would be impairments for Microsoft, Google, Meta, and Amazon, and SEC action against NVIDIA if it did actually sell GPUs to China.

This doesn’t mean that things won’t fucking suck for anyone in the market, nor that the vast majority of people won’t fucking suffer as they always do when bubbles burst. 

Which is why I am making a firm, clear statement to end this piece.

When The Time Comes, Let The AI Industry Burn

I repeat myself:

No bailouts, no handouts, no special treatment, no tax breaks, no CHIPS act, and no sovereign wealth fund. It is time to tell the AI industry to go fuck itself, because it’s effectively done the same to the rest of society. These companies must be forced to stand on their own two feet and die with dignity if their wretched business models can’t keep up.

The world’s governments have rolled on their backs and shown their bellies to the tech industry for far too long, and have been aggressively conned by some of the richest people alive into believing that fucking Sam Altman and Dario Amodei are building anything other than the world’s least-profitable software. 

We do not need a “sovereign AI strategy,” nor do we need “a sovereign AI wealth fund,” nor do we need to “make sure America leads in AI,” at least not when we’re talking about large language models, the underlying technology of ChatGPT and Claude, two of the most over-hyped and deceptively-marketed pieces of software in history. 

Whether or not LLMs are a useful tool is irrelevant, because the AI industry has demanded the world hand it as much land and money and as many resources as it desires to continue proliferating a technology that has only ever lost money and has no path to sustainability. The only reason it has gone anywhere is because the tech industry has united around it as a means of hiding from the fact it has no next big thing, and nothing — absolutely nothing — that a LLM can do remotely justifies the investment.

And it has only got this far because of a captured business and tech media overstating its capabilities and hand-waving its obvious efficacy issues and economic instability. There are too many that have proven easily-wooed by whimsical white boys that promise they’re building machine intelligence, and when the markets bleed red, these people should know that they’re responsible. So much of the so-called journalism around AI has been used to enrich the already-rich and inflate a bubble that will hurt hundreds of millions of regular people globally as Sam Altman and Dario Amodei remain billionaires despite their companies’ fates.

When the time comes, the AI industry must burn. It must be allowed to die. Generative AI has already been given far too much money, oxygen and attention, and if it cannot survive without continual venture capital and media coddling, it is unworthy and unnecessary, and must face the cold, hard reality that every regular person faces when they fail.

And there is no “bailing out” these wretched firms. Giving $42 billion to OpenAI or Anthropic will not fix their business models, nor will it magic up the $400 billion or more in annual revenue to substantiate just NVIDIA’s AI GPU sales through the end of 2027. 

These people are not building the future — they’re finding ways to re-entrench the status quo, to give Microsoft, Google, Amazon and Meta ways to grow their revenues and centralize infrastructure under the auspices of “innovation.” 

If any policy makers read this, know that you’ve been had by the AI industry. They want you to believe they’re essential so you’ll bail them and their rich friends out when the time comes, or funnel taxpayer funds into building them data centers. They are not building autonomous intelligence, nor will they ever do so. 

I think it’s fanciful to imagine that there would ever be actual consequences for this bubble, but if there are, the people to hold responsible are Sam Altman, Dario Amodei, Satya Nadella, Sundar Pichai, Andy Jassy, Jensen Huang, Mark Zuckerberg, and everyone else who forcefully manufactured consent for a dead end technology and built the rails to serve the world its next great financial crisis.

Until something changes, the tech industry will never be capable of building anything other than consensus and reinforcements of the status quo.

So, spit in the face of those who even hint at a bailout, refuse to accept it, and demand that they do the complex, ugly work of thinking about the actual consequences of everyone being wrong. When this era ends, we will need to thoroughly excavate the collapse to make sure it doesn’t happen again, identifying the organizations and personalities that were used to manufacture consent and spread mythology about LLMs. 

Every major bubble that has ever happened has mostly left the stones of responsibility unturned. The carnage that I fear will follow this era’s collapse will be horrifying, and we must do everything in our power to both thoroughly understand how we got here and make sure it doesn’t happen again, which will involve many hard conversations about our financial system, media ecosystem, and how innovation is invested in, built, bought and sold. 

The same goes for the acolytes of this era. There are people who have developed a genuine hostility toward those who do not immediately accept a for-profit entity as their lord and savior. This is a sickness within the tech industry that must be put to an end. 

Much of this will be unavoidable, because I think what follows the AI bubble will be a greater revaluation of the tech industry, a necessary reckoning with reality for a Silicon Valley that’s far more beholden to capital than it is human progress. The cults of personality that dominate this industry do not care about you, or me, or anyone other than those they revere and their theoretical placement in their dream of a society dominated by the rich and their chosen cronies.

I refuse to accept their future as an inevitability.

As I said a few weeks ago:

The AI bubble is sold as the future, but actually resembles the death of Silicon Valley. Only a tech industry dominated by symbolic wealth and value creation would ever abide a trillion dollars of waste for a still-theoretical outcome, and only an intellectually-rotten Valley would be so easily-grifted by people like Dario Amodei and Sam Altman.

This era must end, and all failures must be allowed to fail. 

Let AI burn.

---

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 10,000 to 18,000 words, including vast, detailed analyses of the biggest events and companies in the AI bubble. 

Emily M. Bender and Nanna Inie

In our op-ed for Tech Policy Press ("We Need to Talk About How We Talk About 'AI'"), we made the case against the anthropomorphizing language that makes it harder to have clear discussions of what so-called "AI" technologies actually do, and when and whether to use them. But these ways of speaking are deeply ingrained at this point, and it takes work carve new conversational and writing habits. That work involves at least three steps:

1. Noticing which word choices are anthropomorphizing
2. Finding alternatives
3. Getting in the habit of using the alternatives

In our research (summarized in the op-ed) we have been working on the first two steps, categorizing the kinds of anthropomorphizing language and using those categories to organize potential alternatives.

De-anthropomorphizing language talks about computer systems in terms of their functionality (what people build and/or use them to do), assigns agency to people using systems and not systems, and avoids aggrandizing metaphors about cognition.

We aim to find substitutes that are as self-explanatory as possible, so that you can just go ahead and use them without having to explain. (Though of course, if someone asks "Why are you calling it that?" that's also a great opening.)

Some of these rephrasings may feel a little clunky, and they can end up longer than the anthropomorphizing shorthand. This means it takes a little more dedication to use them, but also isn't necessarily a bad thing. We should stop and think about the tech we are using, or even discussing, and what it actually does.

We'll go through the categories of anthropomorphizing language we identified in Inie et al 2026, and give examples of de-anthropomorphized versions for each.

Our suggestions

Cognizer and products of cognition

This category is super frequent, because it's right in the marketing term artificial intelligence itself. This is language that locates thinking in an algorithm. Instead, we recommend describing software as performing calculations or other algorithmic operations, and locate the thinking with the people using the system. (In some cases, people clearly aren't thinking when they use them, but they are still the ones who should be.)

Examples:

artificial intelligence → probabilistic automation
hybrid intelligence → augmented human intelligence
image recognition → image labeling
speech recognition → automatic transcription
the model shows bias → the model reflects bias
model mistakes → model errors
chatbots are good at … → chatbots are good for …
hallucination → undesirable output

In general, we recommend avoiding using artificial intelligence or AI in reference to technologies. We do still talk about the AI industry, because that is the name of a thing, and talk about AI as an ideology. But when the intended referent is some specific technological system, it is always better to name that system itself. Maybe it's some specific product. Or maybe it's a system with a particular function like automatic transcription. Either way, it's worth finding names that aren't also anthropomorphizing. If you need a more general term, our recommendation of probabilistic automation above works for many (but not all) things sold as "AI".

We've also put hallucination in this category, because in its original sense it refers to perceiving things that are not there, but of course software systems (and conversation simulators in particular) don't perceive anything. Our proposed one-to-one replacement phrase is undesirable outputs, but it is also important to know that all LLM output is probablisitically produced synthetic text; there is no fundamental difference between desirable and undesirable outputs on the system side, but only for the people interpreting them.

Emotion

These are turns of phrase that suggest that software systems have emotional lives. We don't have particular rephrasings to recommend here because there is no accurate way to talk about emotional states of computers other than to reassert the obvious, that they don't have any. What's perhaps most subtle (and thus most fun for linguists) about this category is that these allusions to emotional experience can sneak in in surprising ways: If you say that ChatGPT struggles to do something, or that you had to coax it into some output, you are describing it as if it had emotional states.

Communication

In this category, we find words that place automated systems, usually synthetic text extruding machines, on an equal footing with people in communicative situations. If we ask something of Claude, we are describing Claude as a conversational partner. Instead of verbs like ask, say, inform, discuss, use verbs appropriate to computers like input and output. Another strategy is to foreground the fact of simulation.

Examples:

prompt → text input
answer → output
chatbot / conversational agent → conversation simulator

Agency

Turns of phrase that locate agency with a machine often serve to obfuscate the interests and goals of people. We suggest revising to locate agency with people or choosing less agentive verbs.

Examples:

ChatGPT assisted students → the students used ChatGPT
revealing the solution → displaying the solution
AI agent → probabilistic, unverified software manipulator

The elephant in the room of this category is the buzzword AI agent (and its variants like agentic AI systems). This is a term for software systems that connect LLMs (probabilistic synthetic text extruding machines) and/or other components up with other systems that can impact the world, i.e. systems previously designed for people to do things like schedule appointments, book flights, or make other purchases. Our suggestion for this one for now is probabilistic, unverified software manipulator, which has the advantage of giving a suitably gross acronym ("No thank you, I don't want to use your PUSMic system.") But, we are definitely open to other ideas! Send them our way and if any seem particularly apt, we will add them to this list.

Human Role Analogy

These are words that cast systems as doing the same work as people in various roles, and serve to hide all of the ways in which such automation falls short of what is needed all the while devaluing the actual work that people do and relationships that we form. Calling systems tutor or co-creator are overclaims that describe what a developer might wish they could develop—for those who want to replace people in these roles.

For this category, our recommendation is to use language that describes algorithms as tools (or products) that people use, rather than as human-like entities, and more clearly indicates system functionality while also not telegraphing a plan to replace people.

Names and Pronouns

The names and pronouns we use to refer to systems can also function in anthropomoprhizing ways. With system names, its somewhat trickier, because the system developers usually get to do the naming, and if they use a person's name for it, everyone else is stuck repeating that anthropomorphizing choice (we're looking at you, Anthropic, with Claude) or going for circumlocutions (Anthropic's conversation simulator).

Pronouns are chosen each time, and avoiding pronouns usually reserved for people (and pets), e.g. he, she, and singular they is a good first step. But subtle choices—such as grouping algorithms and people under you or them—can anthropomorphize. Separating systems from people and avoiding collective pronouns is preferable.

Examples:

who’s right? → is the machine output correct?
they produce results → the team uses it [the system] to produce results

Biological metaphors

Computer scientists working in "AI" (and its subfields) have been embedding biological metaphors in their technical terminology for a long time. These turns of phrase might have been metaphorical in their origins, but they also function to suggest more similarity than is actually there. When revising away from biological metaphors, ask how system functionality can be more precisely described to give readers a clearer sense of what is actually happening.

Examples:

neural networks → weighted networks (from Hunger 2023)
the model consumes data → data is used in setting model weights

Reflections

We encourage you to try out the above rephrasings and to create some of your own in the same spirit! It can feel awkward at first, but in our experience it is easier than reliably pronouncing or spelling the word anthropomorphization, so there is that.

It can also feel a bit socially awkward, because you are swimming against linguistic and cultural currents, but that can also be rewarding in and of itself. At a talk she gave in January, Emily was asked by a student how to contribute to the resistance against "AI" in conversations with friends, without being a stick in the mud. Emily said: Be a stick in the mud! If you think about our current situation as mired in mud that's hard to walk in, if you plant a stick, you can start to create firm ground for others to join you on.

For The Ringer, Alan Siegel pays tribute to the Costco hot dog, still $1.50 at the food court, a price that’s held steady since the ’80s and is unlikely to increase. Siegel digs into why Costco hasn’t changed the deal, and why it still works. This is a festive (if belated) Fourth of July read, and a welcome addition to a growing collection of editors’ picks about Costco, America’s beloved big-box, membership-only warehouse club.

In a world where the McDonald’s Dollar Menu is long gone, a 7-Eleven dog costs over $2, and it’s a challenge to stay on budget at Taco Bell, the Costco frank stands alone. It is America’s greatest hot dog. And the country’s last great deal. 

When ChatGPT arrived in late 2022, educators quickly asked whether students would use artificial intelligence to cheat, learn or simply get through homework more efficiently. Evidence is beginning to point toward a troubling answer: Many students appear to be completing assignments faster while learning less from them.

This conclusion comes from one of the largest studies of how generative AI is changing student behavior and academic skills. Sina Rismanchian, a doctoral student at the University of California, Irvine, partnered with researchers at McGraw Hill to analyze millions of student interactions with ALEKS, an online math platform used by more than four million students a year, from fifth grade through college. Because ALEKS includes both low-stakes practice problems and college placement tests, the researchers were able to compare how students behaved and performed before and after ChatGPT’s arrival.

To isolate AI’s effects, the researchers compared two kinds of math problems that differ in how easily students can outsource them to AI: word problems and graphing problems.

Word problems can be copied and pasted directly into AI chatbots for instant answers. Graphing problems are far more cumbersome. A student would need to upload a screenshot and still recreate the graph inside ALEKS using its tools.

After ChatGPT’s launch, student behavior and performance on the two types of problems began to diverge. 

Beginning in early 2023, students started spending less time on word problems while continuing to spend about the same amount of time on graphing problems. The gap widened every quarter. By the end of the study period, near the end of 2025, average time spent on word problems had fallen 31 percent among high school students and 27 percent among college students — from about four minutes per word problem to less than three. (Middle school students showed only a modest decline of 9 percent, and fifth graders showed essentially none.)

The researchers believe those averages are being pulled downward by some students who spend only seconds on word problems because they’re using AI to answer them.

The same pattern appeared in college placement tests. When the exams were taken without supervision, students spent much less time on word problems after ChatGPT’s release. During proctored exams, the time spent on word problems returned to historical norms.

But time is only half the story. The more troubling finding is what happened to learning.

Many colleges allow incoming students to retake placement tests after practicing more math in ALEKS, giving them a chance to qualify for a higher-level course. Before ChatGPT, that practice generally paid off. After ChatGPT, students answered more word problems correctly during unsupervised practice sessions but performed substantially worse on those same kinds of problems when they later took a proctored placement test.

Historically, students answered about 80 percent of these word problems correctly on supervised placement tests. After ChatGPT’s introduction, that fell to about 60 percent — a roughly 25 percent reduction in the odds of answering a word problem correctly. 

Performance on graphing problems, by contrast, did not decline.

After ChatGPT’s release, students performed worse on word problems (AI-susceptible) during proctored exams, but answer more word problems correctly in nonproctored settings

If students’ math skills had generally deteriorated because of pandemic learning loss, weaker high school preparation or digital distraction, graphing performance should have deteriorated too. It didn’t.

The study cannot definitively prove that students were using AI. The researchers couldn’t see what else was happening on students’ screens outside of ALEKS. But it’s difficult to think of another explanation. The changes appeared only in problems that are easy to outsource to AI, disappeared under supervision and grew steadily over nearly three years.

“What makes me nervous is that it’s not only about the word problems,” Rismanchian told me. “This cognitive surrender might be going on in writing, science, everything.”

The paper, “Faster Completion, Less Learning,” was released in June 2026 as a working paper and has not yet been peer reviewed. Like any single study, it doesn’t settle the questions of how much students are using AI in their schoolwork, whether it’s harming learning and by how much. But it joins a growing body of evidence that generative AI is causing students to skip the brain work that leads to learning, and that this “cognitive surrender” is becoming commonplace.

A randomized experiment in Turkey found that high school students who used AI to help them study math ultimately learned less than students who practiced without it. Anthropic, the maker of Claude, has separately reported that many college students appear to use AI to obtain answers and offload cognitive work. Rismanchian’s earlier research, released in March 2026, documented troubling patterns of AI usage in short response essays among undergraduate students at a large California research university.

That doesn’t mean AI always undermines learning. Carefully designed AI tutors have improved student achievement in controlled experiments by asking questions, personalizing instruction and withholding answers until students reason their way through a problem. But using AI this way should increase the time students spend on a problem, Rismanchian said. The ALEKS data show the opposite.

Rismanchian doesn’t believe the answer is simply banning AI. Instead, he argues, students need to value learning enough to resist the temptation to outsource it.

A recent RAND survey suggests many already recognize the threat to their brains. Students report worrying that AI is weakening their critical-thinking skills while more of them admit using it for schoolwork.

Students are not entirely to blame. Even as many professors have warned students not to use AI to complete classwork, universities themselves have embraced the technology, often giving students free access to premium chatbots. 

“I think we need to communicate to students that you should value your learning,” Rismanchian said. “If ChatGPT does it for you, then you haven’t learned it.”

Rismanchian understands the temptation.

An international student, Rismanchian began using ChatGPT to help polish the English in his papers. The ideas were still his own. But after several months, he said, he noticed something unsettling.

“I realized that I cannot write anymore,” he said. “I was losing my writing abilities.”

So he stopped using AI to write.

He still uses it to code.

Contact staff writer Jill Barshay at 212-678-3595, jillbarshay.35 on Signal, or barshay@hechingerreport.org.

This story about AI use eroding math skills was produced by The Hechinger Report, a nonprofit, independent news organization that covers education. Sign up for Proof Points and other Hechinger newsletters.

The post Faster solutions, lower test scores: How AI is eroding math skills appeared first on The Hechinger Report.

Internet pioneer Vinton Cerf will retire from Google after more than two decades, leaving a legacy that shaped the modern internet while predicting standardized protocols will become essential for the next generation of autonomous AI agents.

The public has undergone gravitational collapse.

For a decade, we have explained the retreat from the public internet using Yancey Strickler’s Dark Forest Theory. People withdrew into smaller, quieter spaces because speaking in public became dangerous. Search, recommendation systems, surveillance capitalism, culture wars, and cancellation dynamics transformed the public sphere into a hostile environment. The resulting cozyweb—private group chats, Discords, Slacks, newsletters, encrypted messaging groups, invite-only communities—was understood as a strategic adaptation. The public remained a single connected universe. People simply stopped talking across it.

This picture no longer fits.

The cozyweb has ceased to be merely hidden. It is becoming causally disconnected. The public internet is no longer a hostile commons shared by everyone. It is increasingly the empty space separating an archipelago of informational black holes. The Dark Forest is transforming into the Dead Forest.

Sloptraptions is an AI-assisted opt-in section of the Contraptions Newsletter. If you only want my hand-crafted writing, you can unsubscribe from this section.

A dark forest is still one forest. Signals travel. Creatures remain connected by the possibility of encounter. Silence is strategic. The Dead Forest begins where the silence becomes irreversible. The inhabitants are no longer choosing not to speak across the public sphere. Increasingly, they cannot speak.

The defining feature of a black hole is not infinite density but the existence of an event horizon: a boundary across which causal influence becomes one-way. Once crossed, no signal returns. Outbound communication is not forbidden or unwise. It is impossible.

A mature cozy community increasingly resembles such an object. Its defining characteristic is not privacy but inaccessible interiority. It possesses an evolving local culture, cadence, trust structure, hierarchy of attention, stock of shared assumptions, repertoire of jokes, vocabulary, and ongoing history that cannot be reconstructed from outside observation. These are not simply hidden facts. They constitute a living dynamical state. To understand them requires inhabiting them. Outsiders may observe artifacts, but they do not share the community’s present.

Crossing into such a community is therefore not simply gaining access to more information. It is crossing into another causal universe.

This is why the metaphor of secrecy has become inadequate. Secrets can be revealed. Documents can leak. Membership lists can become public. Event horizons are different. What lies beyond them is not a collection of hidden documents but a continuing history. The defining loss is not information but contemporaneity. Outsiders no longer participate in the same unfolding present.

This mixed metaphor of an arborescent digital cosmos entering its death-arc phase of evolution immediately clarifies several otherwise puzzling features of the contemporary internet.

***

The first is the accretion disk. Every black hole is surrounded by a liminal region where matter has not yet fallen across the horizon but is already gravitationally bound to it. This is where enormous amounts of observable activity occur. The accretion disk is the liminal zone.

The modern public internet increasingly consists of such liminal objects.

Books. Conference talks. Substack essays. Open-source repositories. Journalistic profiles. Podcasts. Public talks. Screenshots. Occasional bridges built by individuals who inhabit multiple communities simultaneously. These are not the interior life of cozy communities. They are matter orbiting their boundaries. They remain visible precisely because they have not crossed the horizon. Some eventually escape into the broader public. Some spiral inward and disappear forever. Most spend long periods circling the boundary between publicity and interiority.

A common mistake is to confuse the accretion disk for the black hole itself. Increasingly, the public mistakes public-facing artifacts for communities. But the relationship resembles that between sunlight reflected off an accretion disk and the interior of a black hole. One cannot infer the character of one from the other.

The second feature is what might be called zombie public life.

If the living public has largely collapsed into compact informational objects, why does the public sphere still appear so active? Because visibility has become detached from shared reality.

Politics, celebrity, institutional media, brands, influencers, and platform-native personalities continue to generate immense volumes of public content. But much of this activity no longer serves the historical function of public discourse: creating common knowledge among strangers. Instead, it functions as a perpetual visibility engine. Attention circulates. Narratives recycle. Audiences become increasingly parasocial. Public performance continues while public life gradually disappears.

Zombie publics are highly visible precisely because they possess relatively little interiority. They are optimized for outward radiation rather than inward development.

The third feature is artificial intelligence.

The emergence of large language models has often been interpreted as the culmination of the public internet. It is almost the opposite.

Black holes are not entirely black. Quantum mechanically, they emit Hawking radiation. This radiation does not consist of messages sent from beyond the event horizon. It is the long thermodynamic aftermath of gravitational collapse itself.

Artificial intelligence increasingly occupies an analogous role: That of a thermalized fossil public.

Its training corpus consists overwhelmingly of the accumulated public internet that existed before blackholification reached its present stage: books, websites, Wikipedia, blogs, forums, public code repositories, digitized archives, public conversations, and institutional documents. Models continuously remix this material into fluent statistical syntheses. They possess extraordinary knowledge of the fossil public.

What they fundamentally lack access to is the living interiority of today’s blackholifying cozyweb.

This is not a temporary engineering limitation. It is a consequence of the causal geometry. The defining conversations of mature communities increasingly occur beyond event horizons inaccessible to public observation. AI therefore becomes the thermalization of the fossil public: the ambient informational glow emitted by a civilization whose most vital conversations have already disappeared into causally disconnected interiors.

This also explains why AI often feels strangely omniscient yet oddly lifeless. It has absorbed the archaeological record of public civilization while remaining largely excluded from its present tense.

This distinction also clarifies the status of genuine leaks. Screenshots from private Discords, leaked Slack logs, internal documents, accidental recordings, whistleblower disclosures—these are not Hawking radiation. They are better understood as fragments of the surrounding black-hole system that never fully crossed the horizon: material lingering in unstable orbit within the accretion disk, occasionally perturbed outward before finally disappearing. They are exceptional precisely because they do not violate the causal integrity of the interior, and can be flexibly narrativized in ways entirely disconnected from the interior narrative. The event horizon remains intact.

Taken together, these three phenomena define the observable universe of the Dead Forest.

1. First, the thermalized fossil public continuously recirculated by artificial intelligence.

2. Second, the liminal accretion disks of boundary objects orbiting living communities.

3. Third, the zombie public whose endless performances preserve visibility while generating progressively less shared reality.

What is conspicuously absent is the thing that once defined the internet itself: a common causal manifold in which strangers could reliably become contemporaries through public communication.

The internet has not become private. It is dying with cosmological grandeur.

The public did not disappear because everyone retreated into private spaces. It disappeared because those spaces underwent gravitational collapse into compact worlds whose interior histories increasingly belong only to themselves. We still observe their radiation. We still see the debris orbiting their boundaries. We still mistake the theater of zombie publicity for public life.

But we no longer inhabit a universe in which the public is the primary medium through which reality is jointly constructed and enacted.

The Dead Forest is what remains after the public has collapsed into black holes.

***

The Dead Forest did not emerge because a new force entered history. It emerged because the forces that produced the Dark Forest were allowed to operate uninterrupted until they exhausted the geometry of the public sphere itself.

The original diagnosis remains largely intact.

Search dissolved into recommendation. Recommendation dissolved into algorithmic manipulation. Surveillance capitalism transformed every public utterance into extractable behavioral data. Culture-war dynamics converted visibility into permanent reputational exposure. Institutions lost the capacity to sustain neutral public ground. Politics ceased to be one domain among many and became the organizing logic of nearly every public conversation. Social media steadily rewarded identities optimized for conflict rather than curiosity. The internet of beefs expanded until it ceased to be merely an internet phenomenon and became a general model for social life.

The cozyweb was the rational adaptation.

People withdrew into smaller spaces where trust could once again be accumulated rather than continuously spent. Communities became increasingly bounded, invitation-based, contextual, and difficult to search or index. Public writing increasingly served not as participation in a common discourse but as boundary maintenance, recruitment, diplomacy, fundraising, publishing, or reputation management on behalf of private interiors.

The public sphere was no longer where life happened. It became where communities advertised their existence.

Nothing fundamentally changed after this decade-old diagnosis. No creative response took shape to check it. The dynamics simply continued unchecked as the no-treatment prognosis suggested. Cozy spaces simply accumulated enough cultural matter to undergo gravitational collapse.

COVID accelerated the migration of meaningful relationships into digitally mediated private spaces. Remote work replaced organizational corridors with Slack workspaces. Institutions weakened further while informal affinity networks strengthened. The second Trump era completed the normalization of permanent political mobilization as the background condition of public life. Meanwhile, every advance in generative AI increased the economic value of public text while simultaneously reducing the incentive to produce genuinely new public writing. The public web became both more extractable and less generative.

The result was not a new equilibrium but a phase transition.

Dark Forest Theory described a world in which everyone remained connected but increasingly chose silence. Dead Forest Theory describes the world after enough silence has accumulated that the public itself loses coherence as a shared causal medium.

AI did not produce this transition. It merely paved the dead cowpaths. Large language models arrived only after the living public had already begun collapsing irreversibly into cozy interiors. It industrialized the recycling of the fossil public while accelerating the exhaustion of what remained outside the horizons.

The internet did not die because of AI. AI is inheriting the remains as it dies, and the cycling of archival and carnival time winds down into a terminal archive.

***

If the Dead Forest is the endgame of the internet we inherited, the obvious question is whether another public sphere can ever emerge.

Not whether this public can be repaired. Cosmology suggests it cannot. Black holes do not become stars again. Gall’s law strengthens this intuition:

“A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over with a working simple system.”

The question is whether history can produce another arborescent digital cosmos and whether we can seed a new forest right now.

Nearly every civilization has imagined a cosmic tree: the Norse Yggdrasil, the Indian Kalpataru, the Mayan Ceiba, the Persian Gaokerena, the Biblical Tree of Life. Their details differ, but they share a common structural intuition. The cosmos is not fundamentally a collection of disconnected places. It is one living organism whose branches connect many domains without erasing their differences. The tree is neither a centralized empire nor an archipelago. It is a common living medium.

The public internet briefly approximated such a structure. We mistook it for a permanent feature of technological civilization. It now appears more likely to have been an unusually low-entropy historical accident.

If another cosmic arborescence is to emerge, it will not do so by reversing the gravitational collapse of the present one. It must grow from whatever remains outside the horizons before those remnants themselves disappear.

This suggests less a legible program of obvious actions than a set of six simultaneous grand challenges that require genuine invention to address.

The first concerns media.

The next public medium cannot simply optimize engagement more efficiently than its predecessors. Nor can it merely federate today’s cozywebs. It must possess intrinsic anti-cozy properties: mechanisms that continuously regenerate encounters between strangers without collapsing into algorithmic extraction or culture-war dynamics. Publicity itself must become renewable rather than exhaustible.

The second concerns politics.

The internet cannot recover a public if politics remains organized around permanent mobilization. A society in which every public utterance is interpreted primarily as coalition signaling cannot sustain common causal space. Any successor public must make disagreement productive without making identity existential. It must allow for mutual co-existence in citizenship rather than a condition of endemic armed activism.

The third concerns artificial intelligence.

Today’s models increasingly thermalize the fossil public. A future public intelligence would instead require access to continuously renewed living culture without simply consuming or exposing it. This is not merely a data problem. It is a civilizational design problem. Intelligence must become metabolically coupled to public life rather than archaeologically dependent upon its remains.

The fourth concerns institutions.

Public institutions once served as long-lived repositories of common knowledge whose legitimacy exceeded that of any particular community. Most now either retreat into cozy interiors themselves or perform zombie publicity in order to remain visible. A new arborescence requires institutions capable of producing genuine common reality rather than merely broadcasting legitimacy.

The fifth concerns public life itself.

Zombie publics cannot simply be replaced by better influencers, healthier discourse, or more responsible platforms. The problem is ontological rather than behavioral. Public life must once again become a place where significant interiority can develop rather than merely be represented. People must once again possess reasons to conduct meaningful portions of their intellectual, artistic, scientific, and civic lives in public.

Finally, there is the challenge of time.

Every year, more communities pass beyond their event horizons. More knowledge is born irretrievably private. More institutions become performative. More AI systems are trained on increasingly recycled corpora. More of the accretion disk spirals inward. More of the fossil public becomes thermalized.

The urgency is cosmological.

Dead forestification appears to possess positive feedback loops. Every successful retreat into interiority increases the incentives for further retreat. Every reduction in the vitality of the public increases the relative value of private worlds. Every increment of AI-generated public text reduces the density of genuinely renewable public culture available for future intelligences. Every new black hole slightly alters the geometry through which subsequent ones form.

If there is a threshold beyond which no new cosmic tree can grow, we do not know where it lies.

Nor do we know whether we have already crossed the ultimate event horizon on the trajectory to collective social heat death of the forest we inhabit now.

The task, then, is not to restore the internet we lost. It is to preserve enough living matter outside the horizons that another cosmology remains possible. The myths of Yggdrasil and Kalpataru remind us that civilizations have long imagined worlds held together by living connective tissue rather than by force or by isolation. Whether technological civilization can grow such a tree again is the defining grand challenge of the twenty-first century.

Dead Forest Theory suggests that this forest cannot be saved, but holds out the possibility that a new one can still be planted before it dies.