Friends, I believe we’re approaching the end of this era. Both OpenAI and Anthropic have filed the paperwork to go public, starting a race for exit liquidity for two companies that burn billions of dollars a year and have no path to profitability.
Both of these companies are dogs. No matter how much financial engineering or how many oafish suggestions about the government taking 50% of every AI firm they make, the underlying economics of AI labs simply do not make sense, which is likely why Clammy Sam Altman is so vague about IPO timing, per The Information:
OpenAI CEO Sam Altman told staff in a Slack message on Monday that he expects OpenAI to go public “within the next year” and that “many things could cause it to be sooner or later in that range, but filing now gives us optionality if we want to go sooner.” Another OpenAI leader also teased an upcoming new AI model that the company is preparing to release.
So, yes, OpenAI is expected to go public within the next year, or sooner, or later, at some point it’ll go public, but when it does, well, I dunno. I really don’t know, actually. I have it on good authority that the underlying financials of OpenAI look like the horrible dog from John Carpenter’s The Thing, and any dithering on Altman’s part is an attempt to delay the inevitable, by which I mean “OpenAI needs $865 billion in the next four years to meet its commitments, and the only way to keep raising money is via the public markets”:
However, he said, the magnitude of capital OpenAI needs for its compute and infrastructure buildout could cause it to accelerate IPO plans. (The Information on Tuesday reported on OpenAI’s discussions to lease a proposed Ohio data center campus that would require it to raise or get financing for hundreds of billions of dollars of Nvidia chips, in addition to making substantial lease payments.)
Altman isn’t alone. Anthropic President and Co-Founder Daniela Amodei said at a recent conference that “it’s a very capital-intensive business to train AI models,” adding that the public market is “very well-suited to that.”
As ever, Anthropic is saying one thing and doing another. Last week, Anthropic rankled investors in bonds associated with its $35 billion deal with Broadcom and Google by, to quote Semafor, “resisting sharing financial information” around its section of the bonds:
Some of the lenders being pitched to buy a slice of the $4.6 billion [editor’s note: it was $4.4 billion in the end] notes that don't have a backstop from Broadcom — meaning they are pure exposure to Anthropic — say they haven't received a detailed look at the AI company's numbers, causing some to pass on the deal, the people said. Such disclosures are standard in lending deals.
Hey, quick question: do you think Anthropic is neglecting to share its financials with lenders because they’re good or because they’re bad? As Semafor noted, the standard in basically any lending deal is that you have to share something more robust than the non-GAAPslop investor decks that Anthropic uses to con venture capitalists, but then again, this is private credit, baby! Anthropic can share as much or as little as it wants as long as there are willing marks.
To be explicit, Anthropic is the one on the hook for every payment of this $35 billion debt deal.
According to the Financial Times, asset managers Apollo and Blackstone are finalizing a $35 billion private credit deal to “finance Anthropic’s growth plans,” specifically using the money to buy Google’s TPUs from Broadcom, at which point Google will install them in a data center and rent them back to Anthropic:
The $35bn financing package is the initial step to fund about 1 gigawatt of Broadcom’s planned AI computing capacity, which is expected to expand to 20 gigawatts through 2028, according to a joint statement by Broadcom, Apollo and Blackstone on Tuesday.
A special purpose vehicle formed by Apollo’s Atlas SP Partners named “Compute SPV” would issue the debt, and Anthropic’s five-year lease payments for the chips underpinned the value of the transaction, said people briefed on the matter.
Apollo and Blackstone structured the loan across three tranches, with interest payments on the two senior segments backstopped by Broadcom. The chipmaker is making the so-called tensor processing units, or TPUs, with Google. Its agreement to provide support if Anthropic misses an interest payment helped vastly reduce the costs on the debt.
That’s right — everything sits off balance sheet in a Special Purpose Vehicle (SPV), legally shielding everybody involved…other than Broadcom, which (per Investing dot com) backstopped $30 billion of the debt.
To be specific, if Anthropic defaults on its payments, Broadcom has to step up and pay off bondholders with something called a residual value guarantee, which means that in the event of default, the TPUs would be sold off and Broadcom would cover whatever the difference was between what they cost and what they sold for.
This is some incredibly dodgy shit, but also suggests that Anthropic has abysmal credit, which makes me wonder how Ms. Amodei thinks raising capital in the public markets will go for a company that now very publicly holds $35 billion in off-balance sheet debt to go with its $2.5 billion revolving credit line.
In truth, the Amodei siblings have complete and utter contempt for the media, the markets and their investors. They know that quite literally anything they say will be taken with complete seriousness, to the point that a long winded and specious slog of a blog about “AI that builds itself” is quoted by the media as if recursive self-improvement were anything other than a pipedream.
Yet this still-theoretical concept is being used as a potential excuse for OpenAI to delay its IPO, per The Information:
Altman said that if the company’s technology advances at a rapid pace to the point where the AI itself is able to create new AI—known as recursive self-improvement—that would lessen the chance of a quicker public listing. “The faster the potential RSI takeoff looks like it could be, the more it could be advantageous to delay an IPO,” because the “technology and the world may change in surprising ways, and there might be good reasons to be a private company during that time,” Altman said.
RSI is the wet dream of an AI industry that’s incapable of working out a sustainable or profitable business model. Nobody — not Altman, not Amodei, not Pichai, not Dean, not Hassabis, not Zuckerberg and most certainly not Musk — has managed to work out a viable business model based on large language models, and despite having an effective monopoly over all tech talent and venture capital, the best idea these fucknuts have is “what if we made the LLM work out how to train itself?”
The fact that the media is taking this with any degree of seriousness is one of the loudest bubble indicators we’ve had in a while. If these companies had anything approaching AI that trained itself…they’d be using it. The AI would be training itself. We’d know, because they wouldn’t shut up about it, but instead we have to deal with yet another agonizing, ten-thousand-word-long blog from Dario Amodei (hey, that’s my job!). Ironically, this may be the first time that somebody has ever ripped off Mark Zuckerberg, who wrote his own blog in the middle of 2025 that claimed that Meta had “...seen glimpses of [its] AI systems improving themselves,” which was, of course, a blatant lie that was nevertheless repeated by the media ad verbatim.
RSI is also, I’d argue, a sign that they’re kind of giving up. Instead of talking about things that the thousands of overpaid academics farting around Anthropic or OpenAI are doing, both companies appear to be leaning on the idea that their models are so special that the people themselves don’t matter. RSI is as theoretical as AGI (artificial general intelligence, or a conscious AI), but feels far more tangible because, at least in theory, it’s just a model that’s doing model stuff without a human being.
If I had to guess, the reason that both OpenAI and Anthropic’s representative coding televangelists are talking about creating “loops” where LLMs prompt themselves is to try and claim that they’re on the verge of RSI:
I expect that “loops” will become the next thing that journalists pick up on and start oinking about. To be clear, “loops” already exist, in that you can make an LLM decide to keep taking actions whether or not a user prompts it for as long as you’d like. Whether the output works at the end isn’t Peter or Boris’ problem, as both of them are allowed to burn anywhere from $130,000 to $1.3 million a month in tokens.
Loops are, of course, literally having a hallucination-prone LLM prompt itself or another LLM, with all the chaos and mistakes that you’d expect to follow. Neither Cherny nor Steinberger give a fuck about how much any of this costs as long as it allows their representative CEOs to keep feeding them endless tokens, even if in doing so they inspire a brief and painful bubble of wasteful token spend in the pursuit of “AI that builds itself.”
There’s a very real possibility that the RSI bubble is the last phase of the larger AI bubble. Recursive Superintelligence raised $500 million a month ago without a product or, well, anything other than a vague theory that (to quote VentureBurn) “human intervention is the bottleneck for AI progress”:
“We are building a system that doesn’t just process information; it processes its own logic,” said one source close to the founders. “The goal is an AI stack that designs its own next-generation architecture. If successful, the leap from one version of the model to the next won’t take eighteen months; it could take eighteen hours.”
That’s a load-bearing “if,” buddy. That “if” refers to the idea that they’ll magically create the literal future of computer science — a self-training AI model that, in theory, could sit around and innovate all on its own, which also begs the question as to what all the researchers would do when that happens.
Nevertheless, I expect RSI to become the next — and hopefully the last — hot topic in AI as everybody gives up on coming up with ideas other than “what if the AI came up with the ideas for us.”
The AI Bubble Is Part Of The Death of Silicon Valley
The RSI bubble neatly fits into an idea I’ve been working on for a while — that the AI bubble is, in fact, multiple bubbles wrapped into one, led by the largest one of all — The Rot-Com Bubble, my theory that everything we see today is the result of Silicon Valley running out of hypergrowth ideas.
The frenzied, reality-defying hype around Anthropic, OpenAI and Large Language Models is a direct symptom of a tech industry with nowhere else to go. There are no other industries that have any sign of becoming the next Google Search or Facebook or Smartphone, which is why everybody — the media, the markets, and every hyperscaler — is conspiring to try and keep the bubble inflated through accepting effectively any viable narrative and blessing even the most vulgar of circular financing arrangements. If anybody for even a second breaks the kayfabe that AI is the biggest, most hugest, most important bubble of all time, everybody has to accept that none of this makes sense…
…and that there’s nothing else on the other side.
The many bubbles that make up the larger AI bubble all represent different aspects of the same desperation. Microsoft, Google, Amazon and Meta are buying all those GPUs and shoving AI in every crevice of their experience because they know that their core businesses will eventually slow down, with nothing else to replace them. Oracle is spending $340 billion or more on capex entirely for OpenAI because its core business lines are plateauing or collapsing. SoftBank is mortgaging its entire future on OpenAI because it desperately needs another Alibaba or ARM to keep up with its ruinous debt. Broadcom needs to backstop $30 billion in bonds for Anthropic, an unprofitable and unsustainable venture capital welfare recipient, because it knows that its other business units can’t keep up with the Rot Economy’s demands for eternal growth.
AI feels, on some level, like the final stage of the modern era of technology. It’s flattened effectively every startup and tech company into some sort of aberration of Large Language Models, turned every semiconductor firm into an AI firm, made every venture capitalist an AI investor, made every tech journalist an AI journalist, and crowded out effectively every other subject in favor of a larger argument about whether one specific technology is the future.
These many bubbles always come back to a singular point: that the people building modern technology have effectively abandoned innovation, twisted by the Rot Economy’s growth-at-all-costs mindset. The result is that the majority of venture capital goes to latter-stage companies and established founders, turning venture capital into a cult of personality more interested in Twitter clout and view counts than anything to do with the future.
Venture capital is now dominated by people that don’t build anything but worship at the altar of what they imagine a “builder” looks like. As a result, these people flock to founders that confirm their biases — those who are usually men, usually white, usually software engineers from big schools, usually building things that look and sound like everybody else. Seed stage investment is dead, and all that remains is a follower culture.
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.
The myth of the Valley was always that investors were smart and took risks. In reality, investors follow other investors based on whatever people are excited about in a group chat or on Twitter or TBPN or any other hype-slop they can get their hands on. Modern venture capitalists hate thinking and introspection, invest in basically the same things at the same time, and haven’t done a real job since the early 2000s.
The result is that Silicon Valley stopped building the future, and started investing in its own destruction.
This series will cover the many parts that make up the larger Silicon Valley bubble, and the many collapses that will lead to the end.
This will, as with my What If? Series, be a two-parter, with the option to extend to three.
And man, am I gonna have some fun with it.
Coming Up On This Week’s Where’s Your Ed At Premium
- Silicon Valley — The Mother of All Bubbles
- The Sam Altman (and OpenAI) Bubble
- The Anthropic Bubble
- The Tokenonomics Bubble