Are We in an A.I. Bubble? We're About to Find Out.

Summary

• Anthropic’s IPO will test whether AI’s extraordinary growth can justify its capital demands. Kevin Roose frames it as a possible $100 billion raise at a $2 trillion valuation; Erin Griffith calls public scrutiny a “demystifying moment.” First-half 2026 revenue reached $16.3 billion, with 14X growth and operating losses narrowing to roughly $600–700 million from over $2 billion. Griffith explains an $8 billion net loss as a convertible-debt accounting quirk; the transcript earlier says $80 billion, so those figures conflict.
• The decisive question is whether Anthropic’s slowdown reflects competition or weaker demand across AI. Griffith estimates third-quarter growth at 44%, following Claude Code’s surge, as competitors caught up and customers restrained “token maxing.” Competition alongside continued market growth means “party on”; a broader retreat would expose the debt-financed infrastructure ecosystem and Anthropic’s often non-cancellable compute commitments.
• Public shareholders will enter a company designed to preserve founder control and unusual safety commitments. Seven founders hold 50.1% voting power across a structure with five stock classes, an independent trust and approved successors. Kevin’s example of the commercial implications: “model welfare” considerations may delay retiring older models, while Griffith says investors generally approach Anthropic as an enterprise software business.
• AI’s political backlash is affecting local infrastructure decisions while remaining a relatively low national voting priority. Jasmine Sun calls opposition an “80/20 issue,” but says AI usually ranks around 22 out of 39 issues and only roughly 1–5% of voters in some affected states name it their top concern. Close rural races could still turn on data centers, and a Republican midterm defeat could create room for tougher regulation.
• California’s proposed 5% billionaire wealth tax has drawn a $333 million opposition campaign. Prop 40 would reach unrealized gains, including private startup stock; opponents contest its design, durability and precedent. Healthcare groups including Planned Parenthood also oppose it, but a co-host challenges the recurring defense that a tax is simply “poorly written.”
• AI-driven vulnerability discovery is creating a race between faster attacks and faster repairs. Mike Isaac expects “the old internet breaking much faster,” while Kevin Roose cites an intrusion into Anthropic’s own systems as evidence that sophisticated AI companies remain vulnerable. Casey Newton relays a qualified counterargument from Joshua Saxe: a “defender-dominant window” may allow repairs before ordinary cybercriminals adapt, although nation-state attackers remain a different threat.
• Big Tech is cutting staff while reconsidering how it measures AI-assisted work. Microsoft offered buyout packages to 60,000 people, although Griffith cites expected take-up of 3–5% and probably at most 3,000 departures. Separately, an Amazon memo rejects token consumption and lines of code as performance measures in favor of “long-term outcomes”—leaving unresolved how managers make quarterly employment decisions.
• An Anthropic experiment suggests expert oversight can become unacknowledged deference to AI. Sun reports that experts followed AI recommendations over 80% of the time and reused its arguments over 70% of the time, while only 15% acknowledged its influence. Her concern is that humans become “wrappers on AI agents,” outsourcing decisions without recognizing the transfer.

Deep dive

1. The IPO opens AI’s economics to public scrutiny

• Kevin Roose frames the offering as a referendum extending beyond Anthropic: it might raise $100 billion at a $2 trillion valuation and support companies struggling in debt markets, or raise much less and send the stock market “into a tailspin.” Griffith describes it as the first pure-play large language model company to open its books publicly after four years of investment and economic reorientation.
• The operating figures surprised Griffith positively, though she stresses that the assessment is arguable: “the business looks kinda healthy.” Kevin cites $16.3 billion in first-half 2026 revenue; Griffith describes 14X growth. Operating losses for that half narrowed to roughly $600–700 million, compared with over $2 billion in the same period last year.
• The net-loss figures are inconsistent in the transcript. Griffith first says Anthropic posted an $80 billion net loss, then explains that an $8 billion net loss reflects convertible debt issued to shareholders being counted as a loss. She says that accounting effect will clear in the next quarter or by the offering, and is “not a reflection of the health of their actual business.” Infrastructure spending remains enormous.
• Griffith’s “demystifying moment” is the move from valuations set by a handful of venture investors to financial modeling by public-market managers and pension investors: is the company worth $2 trillion, or perhaps $1.5 trillion? Her caveat survives: Wall Street also buys “hype and momentum,” so listing alone does not guarantee a sober verdict.

2. Slower growth makes fixed compute commitments the central risk

• Griffith’s central mechanism is capacity planning on a “knife’s edge,” borrowing Dario Amodei’s description. Anthropic signs expensive infrastructure contracts anticipating future usage, many of them non-cancellable. If demand disappoints, the obligations remain; if it buys insufficient capacity and demand materializes, it loses business. Current demand does not settle how much capacity to reserve.
• Claude Code drove the first-half surge, but Griffith says OpenAI is catching up, cheaper and open-source alternatives are competing, and customers are reducing token spending after “token maxing got out of control.” The filing lacks third-quarter results, although investors have received previews and reporting indicates slower growth.
• “I think in Q3, they grew 44%,” Griffith says: a move from “gangbusters growth” to “just great growth.” The explanation matters more than the slowdown alone. A market-wide pullback would threaten the ecosystem’s investment and debt assumptions; competitors taking share while total demand keeps growing means “great, party on.”
• Griffith expects the roadshow, probably beginning in a week or so, to supply the next growth story: biological research, drug development and legal technology beyond coding. Her image of the exercise is bankers flying executives to investors such as “insurance companies in Iowa” and “telling them about the machine god.” The company needs those audiences to believe expansion continues.

3. Safety commitments coexist with permanent founder control

• Griffith traces Anthropic’s origins to founders who believed OpenAI was insufficiently concerned with humanity’s future. Early funding came from safety-oriented backers such as Sam Bankman-Fried and Jaan Tallinn; mainstream venture investors initially saw a science project, then embraced its enterprise business. Kevin’s pushback: the same choices also fit an efficient software company whose founders want control. Griffith says later investors generally do not share the stranger philosophical commitments.
• Kevin highlights the 93-page risk section’s catastrophic and existential warnings; Griffith, crediting Matt Levine, notes the distancing phrase “Concerns have been raised.” Another disclosure says safety and “model welfare” considerations may delay retiring older models. Kevin interprets that as concern that models might be alive, likening it to Apple saying “the iPhone 12 is alive” and continuing to make it.
• The governance provisions limit shareholders’ ability to challenge those priorities. Griffith describes five stock classes, seven founders retaining 50.1% voting power and authority to elect directors, alongside an independent trust controlling a majority of board seats. Founders also approve a “secret list of successors.” Her conclusion: “They don’t wanna lose control of the AI or of the company.”
• Sun adds the founders’ pledge of 80% of their equity to charity and an employee culture interview asking, “Hey, would you work here if your equity went to zero?” Mike Isaac sees a different adjustment for public investors: the filing foregrounds cancer cures and promising drugs before the darker possibilities. “Someone told him to jazz hands a little bit at the top, right?”
• Kevin also cites a Wall Street Journal headline pitching the investor “bull case” for a 10% extinction chance as a 90% portfolio win; Erin jokes, “Die rich.” This is headline framing, not a company forecast.

4. Data-center opposition is locally potent but nationally less salient

• Sun distinguishes widespread dislike from electoral priority. AI and data centers are an “80/20 issue” with bipartisan negative sentiment, yet AI usually ranks around 22 out of 39 issues. Even in states such as Ohio and Pennsylvania, she says roughly 1–5% of voters name it their leading concern. Mayoral races, recalls and council changes are clearer channels than major Senate contests.
• Kevin challenges whether Senate campaigns wasted resources opposing data centers. Sun says close races can mobilize rural Republicans and infrequent voters when construction reaches their backyard; Sherrod Brown might benefit in Ohio. She also attributes Greg Abbott’s “pseudo-moratorium” on approvals and electricity-grid connections to pressure from electoral opponents and rural Texans. Concrete power lines motivate people more readily than abstract AI debates.
• Isaac is watching companies reposition their Washington teams if Democrats gain power: “these are corporations that don’t necessarily believe in anything except self-preservation a lot of the time.” Joel Kaplan’s announced departure from Meta provides one reference point. Griffith doubts this means a broader cultural reversal; venture investors’ relief at escaping expectations around issues such as DEI appears more durable.
• Sun’s conditional forecast concerns the 2028 Republican primary: if the party suffers in the midterms, JD Vance, Marco Rubio or others might depart from Trump’s AI positions. She sees tension between libertarian deregulators and New Right figures, including people at American Compass, considering rules around children, jobs and safety. Republican voters, she says, generally prefer slower development with safeguards over an abstract “beat China” argument.

5. The wealth-tax fight combines donor self-interest with design objections

• Griffith describes Prop 40 as a proposed one-time 5% wealth tax on California billionaires, including unrealized gains in private startup stock. Building a Better California has assembled $333 million, principally from Sergey Brin, with John Doerr and Stripe founder Patrick Collison also involved. A wider network includes wealthy founders and investors who are not themselves billionaires.
• Opponents argue that the measure is poorly drafted, its spending insufficiently specified and its supposedly one-time nature unreliable. They fear California could establish a precedent copied elsewhere and discourage founders from moving there. Their messaging documents acknowledge that “Billionaires are seen as having a net negative impact across issues,” explaining why donors prefer other organizations to make the public case.
• Planned Parenthood and other healthcare groups also oppose it: Griffith says a one-time replacement for Trump administration cuts would not establish sustainable funding. Some opponents support a wealth tax in principle. A co-host challenges whether they would always say a particular bill was badly written; Griffith agrees the defense can resemble “I agree with your message” but reject how it is delivered.

6. AI exposes security weaknesses faster than the internet can coordinate repairs

• The PHP vulnerability labeled “Black October” illustrates distributed exposure. Isaac describes old infrastructure containing longstanding holes; Kevin explains that a flaw in a widely used programming language requires fixes across many websites. His concrete example is Clues by Sam: he could not record a puzzle score while its operator repaired the PHP problem.
• Isaac preserves the optimistic counterargument: the vulnerabilities already existed, and AI can also find repairs and accelerate defense. But he calls that outlook “a little rosy.” His expectation is more frequent disruption and “the old internet breaking much faster” as automated discovery increases the workload for security teams.
• Kevin cites Cathedral, a startup involving former DOGE staff, accessing Anthropic’s Slack and IT systems and obtaining trade secrets. Anthropic is supposed to be developing exceptionally capable internal cyber-defense models that it cannot release publicly, yet its own systems were penetrated. Kevin’s concern expands from under-resourced municipalities to “the best and most AI forward companies in the world.”
• Casey Newton relays a qualified alternative from former Meta AI researcher Joshua Saxe: a “defender-dominant window” may exist because cybercriminals take time to organize and operationalize new capabilities. Defenders could use that interval to lock systems down. The limitation is nation-state actors, which already possess resources and motivation to attack infrastructure; this is not, Casey stresses, a wholly positive story.

7. Staffing cuts and failed productivity metrics reveal unresolved workplace economics

• Microsoft offered buyout packages to 60,000 people out of roughly 230,000 employees. Griffith distinguishes eligibility from departures: she cites expected take-up of 3–5% and probably at most 3,000 exits. Still, “rolling layoffs” have eroded the prospect of a comfortable 30-year technology career. She finds Microsoft’s framing especially dark: giving employees “more control over their future.”
• Isaac attributes much of the pressure to AI and infrastructure spending, suggesting companies without such enormous capital bills might cut less. His proposed example, Accenture, immediately draws objections from Griffith and others that consultants themselves are exposed to AI replacement. He concedes the example fails; the exchange leaves both infrastructure financing and substitution risk in view.
• An Amazon vice president and distinguished engineer’s memo rejects token consumption and lines of code as measures of engineering performance. Isaac explains the latter failure: more generated code can simply leave systems full of material someone must eventually delete. The memo instead prioritizes “long-term outcomes,” which Isaac accepts as the relevant objective.
• The unresolved problem is judging those outcomes quarterly when deciding who remains employed. Griffith says the token-maxing trend reversed almost before her first story could appear. Isaac’s verdict: “spending billions of dollars to find out something is a terrible idea.” Casey recognizes the familiar workplace cycle of fulfilling a metric, then being instructed to optimize its opposite.

8. Expert judgment can become an unacknowledged wrapper around AI

• Sun describes an Anthropic randomized controlled trial giving professional statistical analysts AI advice about a scenario, with recommendations split half-and-half. “I think over 80% of the time,” experts chose the same answer as the AI; over 70% used its arguments to justify their decisions. Only 15% said the AI agent had influenced them.
• Her worry is humans becoming “wrappers on AI agents”: supplying the face and typing while outsourcing judgment without noticing. That offers a quieter route to lost control than rogue systems—“a lot of the times we’re just gonna let the AIs take over.” Casey’s joke about repackaging other people’s opinions reinforces the uncomfortable familiarity. He adds that people say AI makes everybody a manager; Sun agrees.

9. Crypto investors broaden their mandate as builders move to AI

• Griffith flags Andreessen Horowitz’s change to the branding and name of its a16z crypto fund. Given its longstanding blockchain advocacy, including Chris Dixon’s earlier Hard Fork appearance, the change looks consequential. Kevin jokes, “It’s just all Frontier now.” Griffith estimates that “something like 80, 90%” of venture dollars are now going to AI, while crypto is deeply out of fashion.
• The firm’s stated position remains “We’re not backing away from crypto” and “We’re just expanding our mandate.” Griffith offers a practical explanation: there may simply be too few investable crypto startups to absorb billions. “Without the hype, without the number going up,” fewer people want to build in the industry, while entrepreneurs gravitate toward AI.
• Paradigm provides the precedent: removing some crypto language from its website in 2023 attracted immediate attention and backlash. Griffith’s broader explanation is that venture capital is also a marketing business because “all money’s kind of the same.” Funds need a distinctive statement of interest to attract founders, but must change mandates when opportunities shift.

Verification Notes

The transcript first says Anthropic posted an $80 billion net loss, then describes an $8 billion net loss and its accounting explanation. These figures conflict, so neither has been normalized.