Venture Returns, OpenAI Litigation and the Economics of AI Growth

Harry Stebbings opens with a question about falling public software valuations: if companies such as Figma and Datadog are repriced, can venture still make money? Rory O’Driscoll sees a sorting process rather than a blanket rejection of technology. Investors are assigning very different multiples to businesses they expect to grow quickly and those they think have matured.

O’Driscoll argues that high-growth companies can still command very high forward-revenue multiples. He describes venture as a business that follows trends, while warning that businesses perceived to be past their fastest growth can be valued far more cheaply.

Jason Lemkin worries about what Figma’s share performance implies for less exceptional software companies. He distinguishes a great product from the price investors paid for it and asks what boards should tell founders whose companies are smaller or growing more slowly.

O’Driscoll warns against anchoring to the extraordinary price immediately after an IPO. His point is not that Figma is a bad company, but that a company bought at a high growth multiple can leave investors waiting a long time if growth and the multiple both decline.

Lemkin provocatively describes venture’s task as turning high revenue multiples into cash through acquisitions or public offerings before the companies have earned comparable valuations through free cash flow. He says liquidity can vanish when multiples fall, making those windows especially important.

O’Driscoll rejects the description of the model as a scam. Investors buy a basket of potential category winners before earnings are mature because a few businesses may become extraordinarily valuable; many others will not. That is risky price discovery, in his view, not proof that every expensive investment will succeed.

The guests distinguish an impressive business from a venture-scale outcome. A software company with substantial revenue can be a remarkable entrepreneurial achievement even if its likely growth and eventual valuation do not satisfy a particular fund’s return requirements.

Stebbings asks what a non-AI software founder with $50 million to $75 million in revenue should do. O’Driscoll says growth of 75% to 100% at that scale can still lead to a significant company; Lemkin counters that the entry price and ownership an investor was able to obtain change the fund-level mathematics.

Lemkin says that, compared with earlier rounds in his career, high valuations and smaller ownership stakes put much more pressure on the outcome. Even a successful multibillion-dollar exit may produce a less compelling fund return when dilution leaves an investor with only a small stake.

Stebbings argues that the biggest funds compare ordinary software growth with unusually fast-growing AI companies. Lemkin suggests that established software businesses search for direct AI tailwinds, citing RevenueCat’s developer growth as an example; he has an investment interest in RevenueCat.

Lemkin describes a company in his portfolio that added AI analysis to a niche product after years of slower growth. He says the new capability changed its prospects, but presents the expected doubling as the company’s outlook rather than a completed result.

O’Driscoll says founders should separate the difficulty of raising attractive venture capital from the value of the company they already own. A founder can adjust spending, reduce dependence on outside capital and pursue AI features without assuming that every business must become the next foundation-model company.

He illustrates the distinction with a hypothetical: a company reaching $200 million in revenue and selling at five times revenue would be worth $1 billion. A founder with 20% could do very well even if the long, gradual journey is less attractive to a venture fund seeking extreme growth.

Stebbings turns to departures at Thinking Machines Lab, which he says had raised capital at a very high valuation. Lemkin and O’Driscoll liken it to a seed investment with unusually large numbers attached: founder and team risks remain early-stage risks regardless of the size of the financing.

Lemkin questions whether a lab leader needs research credibility to recruit and retain exceptional technical talent. This is his interpretation, not a finding about the reasons for anyone’s departure. The panel does not establish why individual employees left or the company’s current prospects.

O’Driscoll raises a hypothetical response for an early-stage investment whose original team thesis has changed: return unspent money if there is no compelling replacement plan. The guests do not claim that Thinking Machines has chosen this course; they are discussing capital allocation in an imagined case.

The investors discuss whether a talented team might once have provided acquisition value as a downside scenario. They also note that transaction terms and employee retention arrangements can change what outside investors recover, so a headline purchase price is not the same as proceeds to shareholders.

Returning part of an investment quickly can be economically different from recovering it many years later. During a fund’s investment period, the capital may be available to invest again; O’Driscoll uses a hypothetical reallocation to a successful company to show how the fund could recover from an early loss.

Lemkin argues that the strongest AI researchers often choose intellectually interesting work and trusted colleagues, not compensation alone. His comments about employment incentives are general observations, and he says he does not know whether they explain the situation at Thinking Machines.

The panel contrasts incremental work on large language models with a new research direction associated with Yann LeCun. O’Driscoll sees the latter as a higher-risk investment because the research path is less established, while also recognizing why the chance to explore it might attract scientists.

The guests ask what protections investors may negotiate if a valuable research team leaves. They also discuss the appeal of research organizations outside the United States and the potential role of sovereign AI projects, without treating either a commercial outcome or a talent advantage as assured.

Lemkin says most ordinary software startups cannot simply hire a frontier AI research team, even if they can offer high compensation. They may need a technically capable founder or leader who can make practical use of the available models instead.

Stebbings moves to Elon Musk’s litigation against OpenAI and its leaders. The guests describe competing positions and speculate about motives and outcomes; they explicitly acknowledge that they are not legal experts, and none of their predictions resolves the disputed claims.

O’Driscoll recounts his understanding of OpenAI’s nonprofit origins and subsequent financing structure. He presents the original participants’ public-interest intentions as his interpretation and describes the disagreement over how a capital-intensive AI organization should be financed.

The panel summarizes Musk’s position as an allegation that he was misled about the organization’s intended direction, and the opposing position as the view that his money was a charitable donation. These are competing characterizations in a lawsuit, not established findings of fraud.

O’Driscoll discusses large possible damages and dilution as part of the litigation’s stakes. The figures are his description and scenario, not an adjudicated award; the panel is uncertain how claims would be calculated and how any ultimate result might affect investors.

The guests speculate about incentives for a wealthy litigant and potential disruption to a competitor. They cannot know Musk’s private motives, and their characterization should be read as opinion. They also point out that litigation can expose old emails and other records to public scrutiny.

O’Driscoll says discovery can produce embarrassing material even when its legal meaning remains disputed. He refers to documents and depositions discussed in the program, but the participants disagree over what any particular record proves about the original intent of the founders.

He describes a judge allowing the case to proceed as leaving questions to be decided, not as proof that either party is right. The speakers then consider whether a protracted dispute could complicate fundraising through uncertainty about future dilution.

Lemkin expects a prolonged and unpredictable fight and compares a hypothetical settlement with other founder disputes. His suggested settlement percentages are speculation, not an offer or a reported agreement, and the panel notes that a jury trial can turn on contested evidence.

Lemkin offers a personal theory about how Greg Brockman might have felt after leaving Stripe, explicitly calling it a theory. O’Driscoll declines to claim knowledge of Brockman’s feelings; the program supplies no independent basis for treating the imagined motive as fact.

O’Driscoll says he personally doubts the allegation that OpenAI’s founders intended from the outset to deceive a donor. He acknowledges that the dispute would still have to be argued and that his view is a prediction rather than a legal determination.

The panel discusses the burdens of public testimony and a trial, including the possibility of reputational damage even without a legal victory. Its predictions about the jury and the duration of the case are speculative and the guests differ on how likely Musk is to prevail.

Stebbings asks whether a settlement could remove distraction from the business. The others point out that settlement depends on both parties, and they do not know whether any private negotiations occurred or what terms either side would accept.

O’Driscoll models litigation risk as a possible future dilution weighted by a probability, rather than a certain loss of the entire company. He says an investor would need to form a view on both the magnitude and likelihood, neither of which the panel can know from the conversation.

The guests debate how much a lawsuit might affect fundraising while OpenAI also faces competitive and operational pressures. O’Driscoll argues that litigation should be managed by counsel rather than allowed to absorb the entire company’s attention; the others emphasize the potential distraction.

Stebbings asks whether introducing advertising could weaken ChatGPT at a time of competition. O’Driscoll argues that serving a large free user base is costly and that consumer subscriptions alone may not cover it, presenting advertising as one possible business-model response.

O’Driscoll compares the question with the early advertising decisions at Google and Facebook. His estimates of typical free-to-paid conversion are illustrative, not disclosed ChatGPT figures, and the discussion does not establish what an advertising rollout will earn.

Lemkin says clearly labeled, relevant ads could be useful when a user is looking for a product. He uses vendor discovery as an example and argues that a small placement accompanying an answer might help a user as well as an advertiser, depending on execution.

O’Driscoll agrees that a limited number of relevant ads could add information, while warning that ad-heavy interfaces can deteriorate over time. The value of an auction to the buyer of an ad does not by itself establish that its recommendation is the best one for a user.

Lemkin says he increasingly uses language models to discover products rather than relying on conventional search. This is his own experience; model answers can be incomplete or wrong, and he does not provide an independent comparison of the alternatives.

The guests speculate about the speed with which AI advertising might grow. A billion-dollar example and later revenue forecasts are scenarios, not reported results; O’Driscoll notes that even substantial ad revenue must be compared with the much larger costs of developing and serving models.

O’Driscoll says product discovery may be commercially valuable because a conversational query can express purchase intent. He describes using an AI system to narrow a television purchase while acknowledging that model answers can also be incorrect.

The discussion turns to answer-engine optimization: helping a company appear in AI-generated answers. Stebbings discloses an interest in Peak, while Lemkin questions whether products he has seen deliver more than basic advice. Their enthusiasm and skepticism reflect experience and investment interests, not a controlled test.

Lemkin considers whether ad-technology companies might benefit if AI platforms let outside systems buy placements. O’Driscoll is skeptical that a platform would need such intermediaries for its own paid ads and distinguishes that market from improving a brand’s visibility in unpaid answers.

The panel discusses Adobe’s acquisition of Semrush and possible AI-search applications. Their explanation of Adobe’s intentions is an inference, not a statement of an announced plan. They contrast buying a substantial customer base with acquiring a small, more experimental startup.

Both guests describe prior connections to businesses acquired by Adobe. They suggest that distribution and enough organizational scale can help an acquired product survive inside a large company, although whether any particular acquisition succeeds depends on execution.

Stebbings asks about high valuations for ClickHouse and Replit. Lemkin describes ClickHouse as an open-source database technology with roots at Yandex that later developed a hosted business; he argues that existing products can find new demand when AI changes what customers need.

The guests discuss the challenge of converting a popular open-source project into paid cloud revenue. Their comments about specific customers, recent growth and the size of the round are based on what they had heard and are not audited financial disclosures.

O’Driscoll says an investor paying a very high revenue multiple is underwriting persistent growth. For that bet to work, the company needs a large enough category, a strong competitive position and several more years of performance; otherwise valuation can overwhelm a good operating business.

He distinguishes technical and founder risk, go-to-market risk and valuation risk. As earlier risks diminish, price and assumptions about the duration of growth can become the central question for a later-stage investor.

The guests describe ClickHouse’s analytical database use case and the importance of querying large volumes of data. O’Driscoll compares analytical workloads with transactional systems and argues that demand for analysis could increase as organizations use more AI.

Whether ClickHouse can justify its price depends partly on the eventual size of its market and its place alongside other data platforms. The panel’s comparisons with Snowflake and Databricks are rough investment hypotheses rather than forecasts of an actual exit.

On competitive investing, O’Driscoll argues that an investor with a small late-stage stake and limited access to confidential information may face different conflicts from an early-stage board investor. He considers buying competing private companies at that stage closer to owning public growth stocks.

Stebbings asks about Replit’s reported valuation. Lemkin says its product improved substantially in his own use, describing a game he built over roughly 100 hours. That anecdote is not a benchmark, and the revenue figures and future growth discussed are hypothetical or unverified.

Lemkin recalls earlier attempts to build applications that he could not complete and contrasts them with his more recent experience. He argues that a more capable product and sustained revenue growth could make a higher price less risky, though the valuation still relies on future performance.

O’Driscoll says the optimistic investment case assumes rapid growth continues. The guests’ illustrative year-end revenue estimates should not be mistaken for actual financial results, and neither presents a verified forecast from Replit.

O’Driscoll describes trying Replit himself: the agent-centered interface encourages users to describe what they want rather than inspect every line of code. He finds it powerful but says that users wishing to understand the codebase may need a different workflow.

The conversation returns to large venture firms investing across stages. O’Driscoll describes late-stage private investing as a way to gain exposure to companies that remain private longer, while Stebbings questions whether competitive early rounds currently offer adequate risk-adjusted returns.

Stebbings says he prefers either meaningful ownership very early or a small position in a company with established traction, and discloses an investment in Airwallex. O’Driscoll counters that, over time, lower risk normally brings lower expected returns and that temporary pricing imbalances need not make an entire stage permanently unattractive.

The guests debate how far a fund can shift strategy while remaining credible to its limited partners. Stebbings cites Thrive as an example of a firm operating across stages; O’Driscoll agrees that some firms have built both early and late capabilities successfully.

O’Driscoll concludes that companies staying private for longer have shifted several years of growth from public to private markets. Limited partners seeking exposure may therefore pay venture-style fees for assets they once might have accessed in public markets. The panel treats that shift as an economic tradeoff, not a guaranteed opportunity.