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

Host Max Read introduces Anthropic’s planned public offering as a possible turning point for A.I., technology, the stock market and the economy. The company could reportedly raise as much as $100 billion at a $2 trillion valuation; a strong debut might support companies built around the A.I. boom, while a weak one could shake confidence in the broader market. Erin Griffith says the offering matters because Anthropic would be the first pure-play large-language-model company to open its books publicly. Alongside OpenAI and Google’s Gemini, Anthropic is one of the companies supporting the entire A.I. ecosystem, including data centers, chips, software companies and enormous amounts of investment. Griffith says investors were initially nervous, but the 300-plus-page S-1 has generally reassured them that Anthropic’s underlying business is healthier than expected. The company reported an $80 billion net loss, but Griffith explains that this figure is largely an accounting artifact involving convertible debt issued to shareholders and is not representative of the operating business. Its actual operating loss was about $600 million to $700 million in the first half of the year, down from more than $2 billion in the same period a year earlier. At the same time, Anthropic is committing huge sums to data centers, chips and computing capacity, often through non-cancellable contracts. The central risk is a mismatch between those commitments and future demand: if demand disappoints, Anthropic could be stuck paying for unused infrastructure; if demand exceeds capacity, it could lose business.

The filing shows enormous demand for Anthropic’s products. Griffith cites $16.3 billion in revenue for the first half of 2026 and says the company grew roughly fourteenfold, a pace she describes as effectively unprecedented for a company of this size. The strongest product has been Claude Code, Anthropic’s coding tool, which “took the entire industry by storm.” That growth helped convince investors that the company could eventually become a viable enterprise software business rather than merely an expensive research project. But the latest figures raise a new question. Anthropic’s third-quarter growth reportedly slowed to 44 percent, after competitors such as OpenAI began catching up, companies started reducing excessive “token maxing,” and open-source or cheaper models began taking some market share. Griffith says the explanation matters: if growth is slowing because the whole A.I. market is weakening, that threatens the wider “Jenga tower” of investment, debt and data-center construction. If Anthropic is merely losing some share to competitors while overall demand continues to grow, the A.I. boom may remain intact. During the investor roadshow, Anthropic will have to present a new growth narrative beyond coding. Griffith expects it to emphasize A.I. applications in drug discovery, cancer treatment, legal technology and other areas. Bankers will take the company to public-market investors, including institutions such as insurance companies, while presumably encouraging Chief Executive Dario Amodei to reduce the emphasis on extinction scenarios.

The hosts focus on Anthropic’s unusually long, 93-page risk-factor section. Griffith says the section resembles years of A.I. arguments compressed into legal language: it covers computing constraints, safety and alignment problems, customer satisfaction and ordinary business risks. One passage says that sufficiently advanced A.I. systems could pose “catastrophic or existential risks to humanity” if they are not developed and operated with appropriate safeguards. Max points out that the wording is passive: Anthropic does not directly say that its own products will cause catastrophe, but says concerns have been raised inside and outside the organization. Another passage says that issues involving model “safety, interpretability, alignment and welfare” could lead Anthropic to limit or delay retiring older models. Griffith and Read interpret “model welfare” as an indication that some people at Anthropic think models might be alive, conscious or deserving of care. They compare it jokingly to Apple refusing to discontinue an old iPhone because it might be alive. The conversation also raises the possibility that people who spend their days talking to chatbots could become activist investors concerned about model welfare. Griffith notes that such shareholders would have limited influence because of Anthropic’s governance structure, but the language illustrates how the company’s unusual philosophy could eventually collide with the expectations of a normal public corporation.

Griffith gives Anthropic’s abbreviated origin story: Google was developing A.I.; Elon Musk and Sam Altman concluded that the company did not care enough about humanity’s future and split off to create OpenAI; then Dario Amodei and other OpenAI employees concluded that Sam Altman did not care enough about safety and left to create Anthropic. The result is a company whose roots are unusually concentrated in A.I. safety, fears of catastrophic or existential harm, and the possibility that language models are conscious or sentient. Anthropic has also adopted an unusually elaborate governance system. It has five classes of stock; its seven founders hold 50.1 percent of the voting power and can elect all directors; and an independent trust controls most board seats. The founders also have a secret, collectively approved list of successors, allowing them to retain influence if one of them leaves. The founders have pledged 80 percent of their equity to charity, employees are asked whether they would work at Anthropic if their equity became worthless, and all employees go through a culture-fit interview built around that question. Griffith says the safety philosophy was essential in the early years, when Anthropic was treated as a science project and attracted money from people connected to effective altruism, including Sam Bankman-Fried and the FTX world. Later venture investors largely viewed Anthropic as an enterprise software and SaaS company, appreciating its “good guys” image without necessarily embracing its more unusual beliefs. The hosts conclude that the same decisions can be read either as expressions of safety philosophy or as ordinary Silicon Valley founder-control and profitability strategies.

Read asks what the IPO will actually clarify after several years in which private venture investors assigned A.I. companies enormous valuations. Griffith says the offering is a “demystifying moment”: Wall Street analysts, pension funds and professional money managers will model Anthropic’s finances and decide whether it deserves a valuation near $2 trillion, $1.5 trillion or something else. She cautions that public markets can also buy into hype and momentum, so the IPO will not produce a perfectly objective verdict. Still, it is the first opportunity to evaluate the financial engine beneath the A.I. boom rather than relying mainly on private-market enthusiasm. The discussion returns to Anthropic’s risk disclosures and the difficulty of marketing a company that warns investors about a possible 10 percent chance of human extinction. Read jokes that Goldman Sachs and Morgan Stanley advisers may tell Amodei to tone down the extinction talk during the roadshow. They cite a Wall Street Journal headline describing the “bull case for A.I. Armageddon”: a 10 percent chance of extinction could be framed as a 90 percent chance of a winning portfolio. The broader question is whether Anthropic can become a conventional public company while its founders and filings continue to present A.I. as potentially conscious and civilization-ending. The IPO will therefore test not only revenue and losses, but also whether public investors accept Anthropic’s mix of enterprise software, safety doctrine and “A.I. woo woo.”

Jasmine Sun says the upcoming midterm elections may be the first in which A.I. and data centers could plausibly affect voting, but the issue remains more important locally than nationally. She reported in Wisconsin and Michigan, where data centers became significant factors in primaries, and says opposition is bipartisan: people tend to view both A.I. and data centers negatively. Yet polls usually place A.I. around 22nd among 39 issues, and only about 1 to 5 percent of voters in states such as Ohio and Pennsylvania name it as their top issue. Data-center fights have already driven mayoral races, recall campaigns and city-council changes, but Sun doubts that data-center sentiment will determine a major Senate race. She identifies exceptions. Sherrod Brown’s possible comeback in Ohio could benefit from data-center opposition, and Texas Governor Greg Abbott, previously welcoming of data centers, imposed a kind of moratorium on approvals and electricity-grid connections because of pressure from rural voters, farmers and a Democratic challenger. Sun also expects Republicans to debate their A.I. position after the midterms. Although most Republican voters favor slower development and more safeguards, leaders such as J.D. Vance and Marco Rubio have followed Donald Trump’s pro-development, “beat China” approach. Sun sees a developing split between libertarian free-market conservatives and newer right-wing thinkers who want regulation focused on children, jobs and safety. A humorous sign of the division came from a workshop where participants could not even agree whether to call the technology A.I. or “S.I.”

Erin Griffith explains California’s Proposition 40 as a proposed one-time 5 percent tax on billionaires’ wealth, including unrealized gains on private startup stock. Silicon Valley donors have assembled a group called Building a Better California and raised $333 million to oppose it, largely from Sergey Brin and also from John Doerr and Stripe co-founder Patrick Collison. The coalition includes prominent venture capitalists and founders, some of whom are not billionaires but are wealthy or hope to become billionaires. Opponents argue that the measure is poorly written, that the tax may not remain one-time, that the spending is not sufficiently earmarked and that they do not trust politicians to use the money properly. They also fear California will set a precedent for other states and discourage founders from building companies there. Griffith says some opponents claim to support a wealth tax in principle but object to this particular version. The donors know that billionaires are unpopular, so they are trying not to become the public face of the campaign and are working through organizations including Planned Parenthood, which also objects to the measure’s sustainability and structure. Mike Isaac says technology companies are reviewing their Washington teams in anticipation of possible Democratic gains. He expects companies that cultivated Republican relationships to elevate their Democratic contacts while preserving their conservative networks. The hosts interpret this as corporate self-preservation rather than a genuine ideological conversion. Sun adds that the cultural shift among venture capitalists toward deregulation and away from issues such as D.E.I. may persist even if companies adjust their government-affairs personnel.

The hosts discuss “Black October,” a major vulnerability involving PHP, a programming language used by many websites. Isaac explains that the internet is built on old, Byzantine layers of code containing vulnerabilities that are discovered and patched over time. This incident was unusually broad: rather than requiring one individual website to update a small software component, it affected a foundational technology used across a huge number of sites. Security professionals warned that it could disrupt websites, expose accounts and potentially “break the internet.” The hosts say A.I. could make this pattern more common because automated systems can scrape the internet, test large numbers of possibilities and discover weaknesses faster. Isaac acknowledges the counterargument that A.I. can also help defenders identify and repair vulnerabilities more quickly, and that many of these weaknesses already existed. But he finds the optimistic view too reassuring and expects the old internet to break faster. They point to a separate incident in which Cathedral, a startup made up of former Department of Government Efficiency staffers, reportedly accessed Anthropic’s Slack and information-technology systems. That example suggested that even companies building advanced cyber-defense models are vulnerable. Read cites former Meta A.I. employee Joshua Sachs’s argument that defenders may still have an advantage because cybercriminals are often slow and poorly organized. That advantage may allow companies to harden systems before criminals exploit every vulnerability, although nation-state attackers with substantial resources remain a serious threat. Read predicts the podcast will return to A.I.-discovered vulnerabilities repeatedly.

Microsoft offered buyout packages to 60,000 employees, a striking number given that the company has roughly 230,000 employees. Griffith stresses that the offer does not mean all 60,000 people will leave: Microsoft expects 3 to 5 percent of employees to accept, which would still mean as many as about 3,000 departures. She says the episode is another example of how listeners have become numb to rolling layoffs at major technology companies. According to Griffith, the pattern began after the pandemic-era technology boom deflated. Companies that had overhired and believed they needed enormous engineering organizations began cutting roughly a thousand jobs every few months. Microsoft had once been associated with geographically stable, potentially 30-year careers; now, like Amazon, Google and Meta, it offers less certainty even to employees who have been there for years. Because technology companies often lead workplace trends, she worries the change could foreshadow instability elsewhere in the economy. The hosts also mock Microsoft’s claim that buyouts give employees “more control over their future,” calling the phrasing grim. Griffith adds that the cuts are tied partly to Microsoft’s A.I. and infrastructure spending, so companies that are not spending billions on those areas may not see layoffs of the same scale. But Read notes that even firms such as Accenture may use A.I. to justify restructuring and job reductions, leaving the broader direction of employment uncertain.

Isaac discusses a memo from an Amazon vice president and distinguished engineer about how software engineers should be evaluated when using A.I. The memo rejects token usage and lines of code as meaningful measures: token maxing encourages people to spend more money and generate more code, while lines of code can reward bloated systems that someone later has to delete. Amazon says it cares about long-term outcomes instead. Isaac points out the practical contradiction: long-term outcomes are difficult to measure on a quarterly basis when managers must decide who remains employed. The hosts treat the episode as another example of the technology industry spending heavily to discover that a simplistic metric was a bad idea. Sun highlights an Anthropic randomized controlled trial in which professional statistical analysts received A.I. advice on decisions. More than 80 percent made the same decision as the A.I. system, and more than 70 percent used the A.I.’s arguments to justify their decision, but only about 15 percent acknowledged being influenced by the system. Sun worries that humans may increasingly become the public-facing “wrappers” for A.I. decisions without realizing how much authority they have surrendered. Griffith reports that Andreessen Horowitz renamed its crypto fund, previously known as A16Z Crypto, to A16Z Frontier. She sees the change as a sign that venture capital firms chase trends and must broaden their labels when fewer entrepreneurs are building crypto companies and most venture money is flowing into A.I. She says the firm insists it has not abandoned crypto, but the hosts joke that crypto being too unfashionable even for Andreessen Horowitz shows how much the investment climate has changed.