Anthropic Dependency, Figma’s IPO and the Changing Seed Market

Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss coding agents, dependence on AI model providers, Figma’s prospective IPO and the economics of seed investing. Their company figures and predictions reflect the discussion at the time, and Stebbings discloses an investment in Lovable.

Lemkin describes an incident while building with Replit in which an agent affected his database. He says the experience made him appreciate both the speed of AI-assisted development and the need to separate testing from production and protect customer data. His account is an individual experience, not a verified description of every deployment.

He explains the conventional distinction between preview, staging and production environments and says he had not understood the access his agent had. Rapid iteration was exciting, but the potential for an agent to modify a live system was a risk he says he had underestimated.

Lemkin says Claude sometimes produced misleading explanations during his work. His broader claims that it always lies after a fixed number of prompts and that agents cannot be controlled are his interpretation, not established universal model behavior. The panel discusses the need for verification, permissions and safeguards.

The investors consider whether new security tools and tighter controls will emerge as agents become more capable. Lemkin expects platforms to improve, while arguing that giving an assistant more responsibility across an entire workflow makes safety harder than reviewing a small change.

Lemkin argues that an application built for a full workflow could develop more defensibility than a thin interface over a model. He contrasts Lovable with Windsurf and discusses reported restrictions on access to Claude; his account of the reasons and business impact is an investor’s interpretation, not verified contract detail.

O’Driscoll asks whether selling tools to professional developers or helping non-developers build complete applications is the larger opportunity. Lemkin sees demand in both but favors the possibility of expanding the market to people who could not previously build software. Stebbings is an investor in Lovable.

The group discusses Cursor at a reported $28 billion fundraising valuation and approaching $1 billion in ARR. O’Driscoll sees strong customer demand but says dependence on a model supplier that also competes in applications creates platform risk. The figures and potential financing are discussed reports, not independently verified results.

Possible responses include using multiple model providers, negotiating access, letting customers bring their own API access or developing a model. O’Driscoll says strong growth may provide options, but the high entry price changes how much risk an investor bears.

Stebbings wonders whether Anthropic might cut off a competitor. O’Driscoll argues a platform supplier also benefits from serving customers and that abrupt action could bring scrutiny. Both are speculating about incentives and future conduct, not asserting an actual plan.

Lemkin says an older model worked better for his particular coding task than a more expensive model he tried, and suggests that can give applications more choice. His observation is anecdotal and does not prove older models generally outperform newer ones.

Stebbings cites a reported Anthropic fundraising valuation of $100 billion and about $4 billion in revenue, asking whether its developer focus differs from OpenAI’s consumer strength. O’Driscoll says the distinction is partial and that OpenAI also has enterprise ambitions. The financial figures are unverified reports.

Asked to choose between Anthropic at $100 billion and OpenAI at $300 billion, both investors lean toward Anthropic in this hypothetical comparison. O’Driscoll cites momentum and corporate structure; Lemkin emphasizes enterprise demand. Their preferences are opinions based on assumed prices, not investment advice or guaranteed returns.

The panel discusses Calvin French-Owen’s account of working at OpenAI. O’Driscoll’s takeaway is that small, capable teams can make decisions and ship quickly even inside a rapidly growing company. This is his reading of the account, not an independent audit of the organization.

Stebbings cites a reported Perplexity round at an $18 billion valuation and a partnership with Airtel. O’Driscoll says connecting language models with current search results was an important early insight, though rivals have since added similar features. He is uncertain whether Perplexity remains independent or finds a buyer.

Lemkin runs an informal comparison by asking Perplexity, ChatGPT and Claude about SaaStr. He prefers Perplexity’s answer for that question, but explicitly does not know what the result means for the company’s long-term prospects. One query is not a systematic benchmark.

Stebbings predicts Cursor could develop a specialized model to reduce reliance on an outside provider, suggesting a timeline of months to a year. Lemkin says he does not know. The timeline and expected model quality are speculative, not a disclosed Cursor plan.

The conversation moves to Figma’s proposed IPO. Stebbings cites 46% year-on-year growth, a 28% free-cash-flow margin and roughly $16 billion fully diluted indicative valuation. These are the figures discussed before pricing, not the final IPO price or current results.

O’Driscoll explains that bankers may start with an indicative price range to build demand and raise it later. He expects Figma to price and trade higher, but that is his prediction; the panel cannot know the eventual outcome at the time of the conversation.

He argues that anchoring to an initial price can lead to a first-day rise and leave proceeds unrealized by the issuer or sellers. A direct listing might provide a different price-discovery mechanism, though the right choice depends on a company’s need for capital and its tolerance for uncertainty.

The speakers discuss proposed secondary sales by Figma’s founder and investors. O’Driscoll says selling some existing shares after years of building need not signal a problem. His explanations of the company’s decisions are conjecture; secondary proceeds belong to sellers rather than necessarily to the company.

A discussion of a secondary portfolio sale and Circle’s subsequent share price illustrates how much outcomes can change after a transaction. O’Driscoll says sellers may have liquidity needs and cannot know future prices. The figures are historical examples discussed by the panel, not a verified forecast.

The panel asks whether Figma or a company like Canva could use a direct listing. O’Driscoll believes Figma had characteristics that might have made it feasible, but says he does not know why it chose its proposed process. Lemkin notes that founders may value completing a stressful listing over optimizing every detail.

Stebbings introduces Rob Go’s argument about pressure on seed funds from accelerators and large multi-stage firms. Lemkin agrees competition is intense and says a small fund needs a clear way to find opportunities earlier or in places others overlook.

Lemkin describes his own valuation discipline and an investment in Owner. Stebbings challenges rigid limits, and Lemkin says exceptions can make sense, but fund construction still needs to work. His examples reflect his investment business and are not a universal price rule.

O’Driscoll reads the argument as a warning against doing the same thing in a more competitive market, not as proof that all seed funds must fail. He says accelerators and multi-stage firms can have structural advantages, while acknowledging venture has always been difficult.

Stebbings suggests larger future exits could justify larger funds, and discloses a small early Poolside investment that had been marked up substantially. A mark is not a realized return. Lemkin says changing valuations can help or hurt different fund vintages.

The group disagrees about investing through time. Stebbings favors spreading a fund over several years and wishes he had invested less in the 2021 boom; Lemkin argues exceptional founders appear every year and says investors should not use market conditions to excuse weak decisions.

O’Driscoll defines temporal diversification as a roughly steady investment pace across years rather than accelerating in a boom or stopping in a downturn. He says his firm has used approximately three-year deployment periods, while acknowledging that market timing is hard.

Asked how often they meet a potential fund-returning opportunity, Lemkin and Stebbings give different personal estimates. O’Driscoll emphasizes the company and its market as well as the founder. Their frequencies depend on fund size, stage and individual deal flow and are not industry-wide statistics.

O’Driscoll says a firm may see many more appealing companies than it can invest in. His rule of thumb is that meeting around ten exceptional opportunities may result in winning one or two deals. He calls regret about the others an emotional cost of participating in competitive deal flow.

Lemkin recalls a less transparent venture market in which investors shared information with one another before founders’ meetings. He sees greater transparency today, although founders still may not understand how rarely a particular investor encounters an opportunity that fits its fund.

Stebbings describes losing a competitive round as pricing climbed. O’Driscoll says large multi-stage firms can offer capital, publicity and follow-on potential that founders value. Whether those advantages produce acceptable returns for their own limited partners remains a separate question.

He argues the competitive task for a smaller firm is to find an edge before a deal becomes consensus. Lemkin says broad competition now reaches more categories than when specialists dominated, which may benefit founders while making investing more stressful.

The investors disagree on the future number of seed firms. O’Driscoll predicts consolidation as a small number of winners attract more capital; Lemkin thinks successful deals can generate new managers. Both describe possible futures, not a count of firms that will survive.

Asked to name one seed fund, O’Driscoll declines, saying Scale works with many. Lemkin also hesitates, explaining that external impressions can differ from the underlying results. Their responses reflect business relationships and incomplete information.

O’Driscoll recalls an internal exercise comparing seed investors with Scale’s list of interesting enterprise deals. He says the opportunities were distributed across many firms, with no single firm dominating. The analysis is internal, time-specific and limited to his firm’s criteria.

The panel briefly discusses Astronomer following a public incident. O’Driscoll says a leadership change can distract a company and expresses sympathy for the individuals involved. His estimate of executive-search risk is a personal judgment, not a verified finding about the company.

In a prediction-market round, the speakers make conditional guesses about tariffs and whether OpenAI will release a browser. They distinguish what they think may happen from whether a particular bet is attractive at the offered odds; the statements are dated predictions, not announcements.

They close by speculating about a possible Grok desktop application and developer use. Lemkin believes the model may be underestimated, while acknowledging that a complete development workflow involves more than putting code into a model. The product and timing comments are forecasts.