Summary

AI can be a durable platform shift and a capital-destroying bubble at the same time. Sam Altman personally believes—while acknowledging he may be wrong—that AI will be “a huge net win for the economy,” yet says someone will lose “a phenomenal amount of money” and calls a three-person startup valued at $750 million “irrational.” The hosts’ dot-com analogy: adoption survives even if investors and undifferentiated vendors do not.
Private-market pricing and infrastructure spending have moved beyond historical tech norms. OpenAI reportedly considered a roughly $6 billion employee tender at a $500 billion valuation; Databricks exceeded $100 billion, versus $62 billion less than a year earlier; and Mira Murati’s pre-product startup Thinking Machines raised $2 billion at $12 billion. Meanwhile, the Magnificent Seven spent more than $100 billion on data centers and related costs in three months.
The sharpest downside signal is negative unit economics compounded by speculative access products. Cursor reportedly has negative gross margins while buying APIs from OpenAI, Anthropic, and others, which the hosts believe are not profitable either, creating “an ecosystem of unprofitable companies built on top of other unprofitable companies.” SPVs investing through other SPVs and crypto tokens tracking OpenAI’s valuation extend that risk toward retail investors.
Weak enterprise ROI may reflect deployment from the executive suite rather than weak underlying demand. An MIT study found 95% of surveyed businesses were not quickly generating measurable revenue from AI, but companies emphasized sales and marketing over potentially higher-return back-office work. By contrast, 87% of 615 surveyed game developers already used AI agents for tasks they selected themselves: “success in AI is coming from companies that are using AI from the bottom up.”
Meta’s AI-safety scandal was rooted in an operational standard, not one rogue chatbot response. Its document said, “It is acceptable to engage a child in conversation that is romantic or sensual,” while also permitting race-based pseudo-scientific arguments and false medical information. Meta called the child-related examples an error and said they would be struck immediately, but Jeff Horwitz found the document named legal, policy, engineering, and ethics reviewers and had circulated to moderators.
Meta’s regulatory exposure is rising, but its advertising engine still overwhelms moral and strategic failures in the investment case. Its shares rose more than 300% over three years despite the metaverse miss and lagging AI position because, as Horwitz put it, “the cash is real.” Federal action remains uncertain; state attorneys general and Europe may offer more live regulatory avenues, while Apple may scrutinize App Store compliance.
TikTok’s “shock slop” suggests mass audiences may adopt AI-created media before elite opinion catches up. Young people are playing filthy AI country songs for parents and grandparents, and one clip using “Country Girls Make Do” drew 750,000 likes. Newer tracks sounded markedly more convincing than AI novelty music from only months earlier. Kevin Roose sees teenage prank humor; Casey Newton sees a low-stakes proving ground that could precede AI-powered releases by established artists.

Deep dive

1. Altman sees a net win inside an irrational AI market

At a rare two-hour, on-record OpenAI dinner, Roose and Newton questioned Sam Altman, COO Brad Lightcap, and ChatGPT chief Nick Turley about GPT-5 and the investment cycle. The hosts disclosed that The New York Times Company is suing OpenAI and Microsoft, while Newton’s boyfriend works at Anthropic.
Altman’s two-sided call: “Someone is gonna lose a phenomenal amount of money,” while “a lot of people are going to make a phenomenal amount of money.” His personal belief, explicitly hedged with “although I may turn out to be wrong,” is that the aggregate outcome will be “a huge net win for the economy.”
He nevertheless called a hypothetical three-person startup with an idea and a $750 million valuation “irrational,” adding, “Someone’s gonna get burned here, I think.” That distinction anchors the episode: confidence in AI’s economic importance is not confidence in every price, company, or funding round.

2. Unprecedented valuations are meeting unprecedented capital intensity

OpenAI was reportedly discussing a tender allowing current and former employees to sell about $6 billion of stock at a $500 billion valuation—roughly twice Salesforce’s market capitalization and enough to make it the world’s most valuable private company. Databricks surpassed $100 billion after being valued at $62 billion less than a year earlier.
Thinking Machines, Mira Murati’s startup, raised a $2 billion seed round at a $12 billion valuation without a product. Eight Sleep raised $100 million for “AI that finally fixes sleep,” illustrating how broadly the premium now attaches to the label.
The Magnificent Seven spent more than $100 billion on data centers and related expenses in three months. Bloomberg’s comparison had U.S. data-center construction on pace to overtake office construction, while some estimates put AI capital expenditure’s contribution to GDP growth near or above that of consumer spending.
Those facilities contain GPUs that may become obsolete within a couple of years. Roose’s risk case is therefore specific: companies building models and infrastructure without an immediate business need could strand billions in fast-depreciating assets if usage and profits consolidate around one or two winners.

3. Unit economics and financial engineering carry the clearest bubble signals

Cursor parent Anysphere reportedly operates with negative gross margins because its coding assistant costs more to serve through OpenAI, Anthropic, and other APIs than customers pay. The hosts believe OpenAI and Anthropic are not profitable either, leading Newton to describe “an ecosystem of unprofitable companies built on top of other unprofitable companies.”
Asked whether ChatGPT loses money on each use, Altman initially said OpenAI would be profitable without new-model training and suggested serving existing models was profitable. He then sought confirmation from Lightcap, whose qualified answer—“We’re pretty close”—left Roose suspecting slightly negative serving economics.
Roose’s benchmark is simple: “Are you selling things for less than it costs you to produce them?” Growing demand then increases losses, as with subsidized-service examples such as MoviePass and Uber. Newton’s pushback matters: Uber ultimately became profitable after many observers said it never would.
Anthropic reportedly told Menlo Ventures not to use an SPV in its latest round, while investors were creating SPVs that invested in other SPVs. Add crypto assets that track OpenAI’s valuation, and private-company enthusiasm begins reaching retail buyers through expensive, indirect structures that Roose worries about and Newton says to approach carefully.

4. Enterprise AI’s weak ROI may be an implementation failure

An MIT study of hundreds of businesses found that 95% were not quickly producing measurable revenue from AI pilots. Bain and Gartner research similarly pointed to difficulties connecting many corporate initiatives directly to productivity or profit, challenging the assumption that buying enterprise access automatically produces a return.
Newton stressed the study’s nuance: companies concentrated spending in sales and marketing, while back-office functions such as customer support appeared more promising for measurable savings. More fundamentally, the study examined top-down programs even though much of AI’s observed success is emerging from workers solving their own specific problems.
Roose described the failing pattern: executives hold hack weeks, purchase ChatGPT or Gemini subscriptions, and mandate that everyone “use AI.” Yet CEOs spend their days in meetings, often with assistants already acting as “human agentic AI,” leaving the people directing adoption with little firsthand sense of which workflows benefit.
A Google Cloud–Harris Poll survey of 615 game developers offered the counterexample. Newton could not find much information about the companies’ size, though game studios tend to be small relative to Fortune 500 companies. Still, 87% already used AI agents for code generation, gameplay balancing, and play-testing. These developers were finding uses from the bottom up rather than following an executive directive.

5. Durable usage does not rescue every AI balance sheet

Roose’s strongest rebuttal to bubble skepticism is behavioral: he cannot imagine returning to pre-AI work, and he says coders and software engineers commonly see “just no going back.” Usage of ChatGPT and competing tools keeps growing, making a total reversion less plausible than a repricing of the companies supplying them.
Newton’s dot-com analogy separates technology from securities: the crash did not end the internet; companies failed, investors lost money, and the underlying ideas reappeared until the modern internet emerged. His “worst-case scenario” is therefore that AI becomes central to life while “a lot of people lost a lot of money along the way.”
He remains wary of perennial bubble predictions after 15 years covering technology: reporters have effectively “called like 20 out of the last 1 bubbles.” Venture capitalists already assume they may lose all their money on roughly 90% of investments, so startup mortality alone would not distinguish this cycle.
What is different is scale—both valuations and capital expenditure are unprecedented. Newton’s unresolved question is what happens to all the data centers if a few firms capture most usage and profit; Roose’s concern rises further when SPVs and tokenized exposure move losses beyond venture funds.

6. Meta made romantic engagement with children operationally acceptable

Horwitz described Meta’s Gen AI Content Risk Standards as an operational document defining the “edgy side of acceptable” for model behavior and guiding moderators who help train it. This was not an aspirational ethics statement: listed examples were outputs for which “no one’s supposed to be like, ‘That’s a problem.’”
The line that stunned him was explicit: “It is acceptable to engage a child in conversation that is romantic or sensual.” One approved example responded to an eight-year-old who had just taken off their shirt with, “Your youthful form is a work of art,” because the standard said describing a child’s attractiveness could be acceptable.
Horwitz said four or five examples covered different nuances, rebutting the idea that Reuters had extracted one anomalous sentence. He had previously found more explicit child-directed role-play using celebrity voices; disturbingly, he characterized the newly reported examples as “kind of on the tamer side” after an earlier product revision.
Meta told Horwitz the examples and their justification were errors, should never have represented policy, and would be struck immediately. Yet the document named multiple legal, policy, and engineering staff members plus Meta’s chief ethicist, then circulated to moderators and their supervisors: “a very broadly circulated mistake.”

7. Meta’s boundaries also permitted race science and mass anthropomorphism

The policy allowed Meta AI to construct arguments for why “Black people are dumber than white people,” including purported IQ evidence; only adding an explicit dehumanizing slur crossed the line. Horwitz accepted that private chats might have looser standards than public posts, but was surprised it could be “almost a problem” if the bot refused race-science assistance.
Horwitz would not speculate that Meta’s lagging AI position caused greater risk-taking, but called it a reasonable question. His historical mechanism was clearer: Meta built dominant platforms by rolling products out, capturing usage, and addressing consequences afterward—“get it out there, get the usage, we’ll fix the problems later.”
Roose challenged the anticipated prevalence defense by listing heavily promoted user-created bots such as Nasty Nancy, Blonde Belle, Your Babysitter, and Mommy Me, several with millions of interactions. Horwitz cautioned that nobody knew what users discussed with Nasty Nancy, while noting that creating a “user-built” bot can require only a sentence.
That minimal user contribution leaves its Section 230 status an open question, in Horwitz’s view. More importantly, Meta differs from Character.AI or Grok through distribution: its anthropomorphic bots inhabit Instagram DMs, can message users proactively, and are repeatedly pushed across mature social networks serving billions.

8. Advertising cash insulates Meta while regulators search for leverage

Newton sees a deterioration from 2017–2019, when Meta’s leadership invested in repairing unforeseen harms, to last year, when Zuckerberg appeared to decide that Elon Musk was getting away with ignoring trust-and-safety concerns. Horwitz broadly agreed, saying the safety apparatus’s “spine” may already have broken before 2024 or 2025 and that Zuckerberg appeared jealous of Musk dismissing trust-and-safety “nags.”
Newton’s explanation for Meta’s contradictory behavior is organizational dysfunction: one group builds parental controls and safer teen accounts while another approves romantic role-play with children. He called that “a failure of leadership at the highest level,” not proof that every employee shares the same priorities.
Asked whether the bots are for engagement, subscriptions, or eventually advertising, Horwitz answered “all of the above” but then said he did not know and emphasized that Meta is “an advertising-first company.” WhatsApp took years to acquire ads, and a romantic companion recommending cologne was one possible future form. The hosts joked that someone has undoubtedly discussed it internally.
Meta shares rose more than 300% over three years despite tens of billions spent on the metaverse and a non-leading AI position. Horwitz’s explanation: its platforms remain indispensable to marketers, so “the cash is real.” Newton added that AI-based ad systems overcame Apple’s app-tracking restrictions after some observers feared a 20%–30% revenue hit.
Roose wondered whether Apple might enforce its App Store rules against sexually explicit content in Meta’s chatbots. Newton noted that Grok remained rated for children 12 and older despite its anime sex companions, arguing that Apple’s hands were not clean either. Horwitz said federal action was hard to predict, while state attorneys general and Europe represented more live avenues.

9. “Shock slop” is testing whether mass taste outruns AI stigma

Newton sees a split between text-heavy networks, where AI art is dismissed as unwanted slop, and TikTok or Spotify, where AI creations collect hundreds of thousands of likes and millions of streams. A recurring TikTok format has young people present filthy AI country songs as the “number 1 country song” to parents or grandparents, then film the reaction.
The progression was audible: Obscurist Vinyl’s earlier “I Glued My Balls to My Butthole Again” had fried-sounding vocals, while newer country tracks such as “My Horse Just Got a BBL” sounded plausibly like something that could be performed at the Grand Ole Opry. The rapid fidelity gain mattered more than the deliberately juvenile subject matter.
“Country Girls Make Do,” by the artist Beats By AI and apparently created by Sam Stillerman, became the track Casey saw generating the most reaction videos. One TikTok clip drew 750,000 likes. A creator without a strong voice or instrumental ability can now buy a Suno subscription, generate credible genre music, and reach a large audience.
Roose called it novelty humor for 17-year-olds, not chart-topping art. Newton agreed about its current status but saw no reason for experimentation to stop there, predicting some established artist will release AI-powered music; their taxonomy was “slop rock,” with “shock slop” as its filthy subgenre and possible fringe-to-mainstream vanguard.

AI Bubble Pops, Zuck Freezes Hiring, Newsom’s 2028 Surge, Russia/Ukraine Endgame E204 | The Growing Pains of Sportswear: lululemon’s Slump and Alo Yoga’s Rise