Hard Fork on Anthropic’s IPO and the Week in AI

Max Read introduces an episode centered on Anthropic’s proposed initial public offering. He says the filing could help investors assess the business behind the AI boom and frames the outcomes as uncertain: the eventual valuation, the demand for AI and the consequences for the wider industry remain to be seen.

Read welcomes New York Times technology reporter Erin Griffith. She says the filing offers a rare detailed look at a major independent large-language-model company whose business sits within a much larger ecosystem of infrastructure investment and customer spending.

Griffith says some investors were encouraged by the reported operating results, while acknowledging that assessments differ. She distinguishes a reported $80 billion net loss, which she attributes largely to accounting for convertible debt, from an operating loss of roughly $600 million to $700 million in the first half of 2026. These are figures and explanations discussed on the show.

The company also faces substantial infrastructure commitments for data centers, chips and compute. Griffith describes a planning risk in both directions: contracting for capacity that demand does not absorb can leave large obligations, while insufficient capacity can mean lost business.

Read cites $6.3 billion in first-half 2026 revenue, and Griffith describes very rapid growth. Their question is not whether there has been demand, but how much will persist as competition, spending controls and the industry’s capacity plans evolve.

Griffith says the filing did not include third-quarter results and discusses reporting that growth had slowed after a rapid expansion of Claude Code. She mentions competing products and efforts by customers to control token spending. The group treats both the figures and the causes as matters for further investor scrutiny.

They distinguish a market-wide slowdown from a shift in market share: either could affect Anthropic, but a general decline in demand would have different implications for the many companies financing AI infrastructure. Griffith says she will watch how the company explains its growth during the expected investor roadshow.

Griffith explains a roadshow as meetings where a company and its bankers present their case to potential public-market investors. She expects Anthropic to discuss possible growth beyond coding, including legal work and biotechnology; these are potential applications, not demonstrated outcomes of the offering.

Read turns to the lengthy risk section of the filing. He and Griffith describe disclosures ranging from compute constraints and customer issues to AI safety. They note language about concerns over catastrophic or existential risks and discuss its careful attribution rather than treating a particular prediction as fact.

Read points to language about model safety and welfare potentially affecting when older models are discontinued. The speakers offer speculative interpretations of that wording and joke about what it might mean for investors. The filing language alone does not establish that a model is conscious.

They revisit Anthropic’s origins in a split from OpenAI and the role of AI safety in its identity. Griffith describes a governance arrangement involving multiple share classes, founder voting power and an independent trust; she presents it as an extension of a broader technology-industry trend toward founder control.

Read asks how much philosophy influences business choices. Griffith says early supporters included people associated with effective altruism, while later investors often viewed the company as an enterprise software business. They discuss two possible explanations for the same decisions: safety commitments and conventional incentives to retain control and build a profitable company.

Griffith says a public offering would expose private-market valuations to a broader set of investors evaluating financial statements and growth. She cautions that public markets can also be driven by momentum, so a share price will not resolve every debate about the value or risks of AI.

Read asks what to watch after the offering. The two discuss the gap between long-term AI risk language and the questions investors may emphasize on a roadshow. Their jokes about extreme outcomes are commentary, not probability estimates offered as investment advice.

After a break, Read is joined by Griffith, technology reporter Mike Isaac and writer Jasmine Sun. Isaac says he perceived the opening of the filing as emphasizing AI’s positive possibilities before turning to risk. The panel discusses how a company’s language can serve both its mission and its relationship with investors.

Sun discusses what she sees as tension between Anthropic’s public-benefit and safety identity and its role as a profit-seeking company. She points to founder charitable pledges and company culture as part of that discussion. The guests treat their interpretations of motives and governance as views, not findings of fact.

With US midterm elections approaching, Read asks Sun about AI and data centers as political issues. She says local opposition can be intense and bipartisan, but reports that AI and data centers generally rank below other priorities in broad polling. Her point is that negative sentiment does not automatically translate into a decisive national vote.

Sun says local projects can still matter in close races, especially where residents encounter power lines, electricity demand or a proposed facility directly. She discusses Ohio and Texas as examples and distinguishes concrete local concerns from more abstract debates about AI regulation.

Isaac expects technology companies to reconsider how they work with political leaders if the balance of power changes. The group distinguishes possible adjustments in Washington relationships from a wider cultural shift in Silicon Valley, which they say is less certain.

Sun describes a potential split among Republicans between supporters of deregulation and people more concerned about children, jobs or local data centers. The panel treats the direction of future regulation and campaigns as an open question rather than a settled election outcome.

Read asks Griffith about a California ballot proposal. She describes Proposition 40 as a proposed one-time 5% tax on billionaire wealth, including unrealized gains, and reports that a group opposed to it had raised substantial funds. Her description is of a proposal being contested before the vote.

Griffith says opponents include prominent technology investors and founders who question the proposal’s wording, sustainability and possible effects on California. She says some organizations opposed to the measure object to its design rather than necessarily rejecting every possible wealth tax.

The panel discusses the coalition opposing the measure, including healthcare organizations, and questions whether criticism of a proposal’s drafting can also serve broader political interests. These are the participants’ interpretations of campaign strategy, not a prediction of the result.

Read introduces a recently discovered PHP vulnerability, which the panel calls part of “Black October.” Isaac explains that widely used older infrastructure can contain flaws that affect many services at once. The guests do not provide a complete technical assessment or a list of affected systems.

The group debates how AI changes the pace of finding and fixing vulnerabilities. One argument is that the same tools can help defenders identify and patch existing flaws; Isaac worries that discovery and exploitation may still outpace fixes across many independent sites.

Griffith raises a reported intrusion involving Anthropic as an example of the difficulty of security even for AI-focused companies. Read cites an argument that defenders may have a window to improve security, while warning that well-resourced actors remain a separate concern. The panel presents these as competing assessments.

The conversation moves to Microsoft’s offer of buyouts to 60,000 employees. Griffith stresses that an offer is not the same as 60,000 departures and discusses an expected acceptance range. The guests connect the announcement to broader changes in technology employment while distinguishing speculation about causes from confirmed results.

They discuss how repeated workforce reductions change expectations of long careers at large technology companies. Isaac suggests infrastructure spending may influence some company decisions, while the panel acknowledges that the effect on other employers could differ.

After a break, Isaac highlights an Amazon engineering memo. He says it questions token consumption and lines of code as measures of engineering quality and instead emphasizes longer-term outcomes. The practical difficulty, the panel observes, is evaluating those outcomes on a quarterly schedule.

Sun describes an Anthropic research report on expert reliance on AI advice. She reports that many participants followed the AI’s recommendations and used its arguments while fewer acknowledged influence. She worries that people may defer to AI without recognizing it; the panel does not establish that the study generalizes to every decision.

Griffith points to Andreessen Horowitz’s change of branding from a crypto fund to a16z Frontier. She says the firm has maintained that it is not abandoning crypto, while the broader mandate reflects changing opportunities and the importance of marketing to venture investors.

The guests discuss how investor interest and startup formation can move between sectors, while noting that a broader mandate does not by itself prove an exit from a prior investment area. Read thanks Griffith, Isaac and Sun and closes the episode.