AI Agents, Venture Pricing and the Tradeoffs Behind the Funding Boom

Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss AI assistants, venture pricing, coding tools and the infrastructure behind them. The conversation includes reported financing and IPO plans, personal investment opinions and a dispute involving Airwallex; those claims are not independently verified in the episode.

Stebbings opens with a reported delay of a possible Anthropic IPO and asks whether it signals market weakness. O’Driscoll offers another possible explanation: the company may want a cleaner account of recent quarterly performance. He presents this as his interpretation, not knowledge of the company's decision.

The panel considers the tradeoff in waiting. More current numbers might improve the investor story, but another month exposes a company to changes in market conditions. O’Driscoll says a receptive market can close unexpectedly; neither the proposed date nor a valuation is established in the discussion.

Stebbings wonders whether the suggested delay could also reflect pressure in conversations with prospective investors. O’Driscoll says management may simply believe it has time to wait for a cleaner account, while acknowledging the downside if the market turns. Neither speaker claims access to the company's internal conversations.

Stebbings asks whether frontier-model companies can list if they face potential liability and have difficulty obtaining insurance. O’Driscoll believes the absence of insurance need not prevent a listing and argues that financial scale and disclosure are relevant. This is his view of a hypothetical situation, not a legal conclusion.

O’Driscoll says securities offerings disclose risks rather than eliminate them, and suggests a large company may have ways to bear some potential liability itself. The discussion also references public statements on AI safety. Those views do not resolve insurance availability, securities compliance or any actual liability claim.

Lemkin expects well-funded companies to maintain substantial legal teams as they address potential claims. O’Driscoll emphasizes risk disclosure as a feature of securities offerings. The exchange does not resolve the legal duties or liability of any particular AI provider, nor does it establish the probability of an extreme AI outcome.

The group then discusses a reported OpenAI forecast for very large future cash burn and capital needs. O’Driscoll describes possible revenue growth, spending and third-party infrastructure commitments as figures from that forecast. They are unverified estimates raised by the speakers, not audited financial statements or established future results.

His broader point is that frontier AI is more capital-intensive than conventional software. Some infrastructure spending may sit on other companies' balance sheets rather than a model provider's own. The panel jokes that projections can be exceeded; it does not know the company's eventual cash requirements.

Stebbings highlights Meta's Muse as his story of the week. Lemkin calls it a strong consumer product and a potential competitor to ChatGPT, praising both its assistant interface and agent features. These are his impressions from use, not an independent comparison of all products or current pricing.

Lemkin argues that a consumer assistant can combine agent actions with ordinary conversational questions in one interface. He sees that combination as commercially significant, particularly if it is easy to access. His comparison reflects a user's early experience and may not reflect the products' future capabilities or economics.

O’Driscoll agrees that the launch fits Meta's distribution through its existing consumer platforms. He considers the product an encouraging change from earlier skepticism about Meta's AI investment. He also discusses market capitalization moves, but attributes their significance to his interpretation rather than proving a single cause.

O’Driscoll says the ability to distribute a working product through existing platforms could give Meta an advantage. At the same time, he says he does not know the product's economics. A favorable reaction to a launch should not be confused with proof of durable adoption or profitability.

Lemkin says the tool built him a personal CRM that tracks sponsors and emails. He describes it as useful but limited for collaboration across a sales team. O’Driscoll notes that such an application combines a model, a database, email access and a sandbox, and may consume substantial compute if used widely.

The discussion turns to agentic shopping. Stebbings describes Amazon restricting access while Shopify partners with the product. O’Driscoll thinks each response could reflect different incentives: Amazon has an advertising business and control over its customer experience, while Shopify represents merchants that may welcome additional demand.

He also suggests an agent that purchases a specific item might reduce the opportunities for advertising and additional items in a shopping basket. Amazon's fulfillment network and scale give it negotiating leverage, in his view. Shopify may care more about merchants gaining sales and its role in payments.

The proposed explanation for Amazon's response depends on more than a payment transaction: customers visiting a marketplace may encounter recommendations or sponsored listings that an agent bypasses. O’Driscoll suggests a merchant-focused platform faces a different calculus. He presents this as analysis of incentives, not confirmation of private negotiations.

O’Driscoll argues that real consumer demand forces businesses to settle questions about APIs and agent payments faster than abstract planning alone. Lemkin predicts agents will increasingly route around intermediaries where alternatives exist. Both agree that the eventual arrangements will depend on negotiation, not just technical capability.

Lemkin does not claim every intermediary will disappear; he argues that even a modest hit to advertising, upselling or growth could be significant. He describes an agent proposing a plan to migrate away from a vendor after a price increase. O’Driscoll sees pressure on information-based middlemen but remains less certain the effect will be fatal.

O’Driscoll notes that a persistent agent can compare many sites at low marginal effort, potentially putting pressure on businesses that aggregate information. He is more cautious than Lemkin about predicting the scale of the effect. Physical fulfillment and exclusive supply can still matter to a consumer platform.

O’Driscoll identifies physical infrastructure and delivery as potential defenses for Amazon. Both panelists expect lower search costs to shift some bargaining power, while acknowledging that large platforms can restrict access. How they respond when many agents seek access, he says, may prove more complicated than blocking one provider.

The panel introduces Jev, a developer-oriented tool for fast classification and decision tasks. Lemkin says his own experience was positive after he reworked prompts: he used it to decide which people in the SaaStr community should meet. His account of speed and cost is specific to his experiment, not a controlled industry benchmark.

Lemkin's first attempts at using Jev were unsuccessful, he says, because the prompts and use case needed adjustment. Once he narrowed the task, he found the resulting classification useful. He stresses that a simple yes-or-no decision does not require the same behavior as a full conversational answer.

He says the tool answers a narrow subset of questions and does not replace a model that writes text or performs complex reasoning. In his application, he estimates it could handle only a portion of the model calls. He sees the opportunity in avoiding expensive models for simple classifications, not in replacing every AI product.

O’Driscoll describes quick true-or-false answers, rankings and scores as a different category from generating long text. He proposes an illustrative market calculation: if a portion of spending shifts to a much cheaper service, the supplier may gain a business while total revenue for that workload falls. The numbers are scenarios, not verified forecasts.

O’Driscoll's market illustration assumes only a fraction of large-model calls can migrate to a simpler service. Lemkin disputes the precise cost comparison and believes the savings in his own use case are much greater. Both agree that lower unit prices mean the dollars flowing to the new category need not equal the spending displaced.

The speakers distinguish unbundling a developer API from replacing the consumer-facing ChatGPT app. A developer may route a simple decision to a specialized service while still using a larger model for complex work. O’Driscoll expects the needs of consumer interfaces and software developers to diverge further.

Lemkin says getting the benefit requires testing each use case. His own first day of experiments failed until he revised the prompts. He found that a system that performs well on one simple matching question might still make a poor decision on a more nuanced hiring or partnership question.

Model routing adds another layer of operational work, he says. After trying an automatic router, he observed degraded results for his application and changed the configuration. He expects ongoing evaluation and quality assurance to grow more important as teams choose among many models, even when routing promises lower costs.

For Lemkin, the practical question is whether a routing system can reliably select the right model for an individual task. A wrong choice in a critical coding workflow can cost time even when it saves tokens. He predicts more teams will need continuous evaluations rather than relying on a single benchmark.

Asked how OpenAI and Anthropic might respond to low-cost tools, the speakers offer competing speculation about incentives and product strategy. Lemkin criticizes smaller models based on his own tests, while acknowledging exaggeration. Their comments are opinions about particular experiences, not independent evidence of providers' internal decisions or universal model performance.

Stebbings says larger seed rounds have made traditional seed investing harder and cites a reported $40 million Jev round. Lemkin points to fellowships and pre-company programs as ways investors try to meet founders earlier. Neither claims that conventional seed opportunities have disappeared entirely.

Lemkin compares newer pre-company investment programs with earlier founder fellowships. His argument is that investors seeking very early access may need relationships before a company has a formal product or funding round. Stebbings and Lemkin disagree on which programs attract the strongest founders.

Lemkin says investors who continue writing smaller checks may need to accept lower ownership and larger eventual outcomes to make the portfolio work. O’Driscoll adds that inflation and economic growth have increased the nominal amount of money needed to hire and operate compared with 2010.

O’Driscoll argues that a check denominated in today's dollars commands fewer resources than the same nominal amount many years ago. He uses a simplified economic comparison to explain part of the increase in seed sizes, while recognizing that market enthusiasm and competition can add to that effect.

Stebbings asks whether higher entry checks are justified by much larger exits. O’Driscoll cautions against comparing today's high exit values with new, more expensive investments: successful exits often began with checks written years earlier at different prices. Current investors need to consider what future exit valuations would have to be.

The panel discusses a Menlo investor's warning about investing near the top of a cycle. O’Driscoll interprets the argument as a mix of funds with substantial gains investing aggressively and others fearing they will miss out. Some increase in prices may be explainable, he says, but some may reflect sentiment that can reverse quickly.

O’Driscoll describes the Menlo argument as separating investors with successful earlier positions from those trying to catch up. Stebbings challenges the warning because Menlo has itself participated at high prices. The disagreement is about how to interpret market behavior, not an allegation that a specific deal is improper.

Asked how to act on that concern, O’Driscoll recommends recognizing portfolio-level exposure rather than trying to time the market perfectly. An investor can lean into opportunities but should consider whether too many expensive positions would threaten the fund if conditions worsen. He presents balance and pacing as judgment calls.

Stebbings and Lemkin push back that a generic warning to be careful is not very actionable, particularly when successful investors themselves have paid high prices. O’Driscoll replies that a transaction resembling a past success can offer a very different risk-return profile if its price is several times higher.

O’Driscoll replies that an apparently obvious risk can still be forgotten during a strong market. In his example, buying a similar asset at several times the old entry price may reduce the potential multiple even if the company performs comparably. Stebbings maintains that investors need a more concrete response than simply recognizing this.

Lemkin asks what a cautious investor would actually do. He recalls making only one investment during 2021 and suggests being selective is possible, though funds still have commitments to deploy capital. Stebbings proposes another response: reviewing which existing positions could be sold to create liquidity.

Lemkin worries about the exit market for slower-growing 'quiet compounders' even if their underlying businesses are sound. O’Driscoll is less certain that the market for profitable, steadily growing companies has disappeared permanently. Both treat current buyer interest as a condition that could change.

Stebbings questions whether a fund focused on the traditional triple-triple-double-double growth trajectory can compete for limited-partner capital with managers posting much faster paper gains. O’Driscoll says comparing an early entry price with a later round in the same company can obscure how much risk and expected return have changed.

Lemkin suggests a manager could consider both exceptional fast-growth opportunities and capital-efficient businesses if each can produce the required returns. He cautions that very high annualized returns in a single vintage should not be assumed to persist indefinitely. The panel does not offer a guaranteed return target.

The panel agrees that strong investment selection is essential whether a fund chooses a slower compounding business or a headline-grabbing AI company. Lemkin suggests managers can pursue both when the economics work. He also says that exits, not just reported valuations, matter to the actual return.

O’Driscoll stresses that company selection matters under either strategy. Paying more for the best opportunity can be rational, but he says price can still matter if an investor's valuation assumptions are wrong by a large multiple. The possibility of a cyclical valuation reversal is a risk to consider, not a prediction that every AI deal will fail.

For limited partners selecting seed and Series A managers, O’Driscoll favors investors with an edge in the strongest new opportunities. He would prefer higher ownership but says consistently accessing excellent deals can matter more than rejecting every small initial stake. He distinguishes early-stage access from much larger later rounds.

O’Driscoll distinguishes the risk of a small seed check from a much larger investment at a far higher valuation only months later. Even where the underlying company is attractive, he says, investors should consider how much downside protection remains at the later price. A strategic acquisition is only a possible outcome.

Lemkin says choosing emerging managers is difficult when LPs are tempted to chase the most recent returns and the next successful fund may look unconventional in advance. O’Driscoll argues past success can persist more in venture than in public-market fund selection, but is not a guarantee of future results.

Lemkin says a manager who performed exceptionally well in one period may be difficult for LPs to access later, while a future successful manager might have a strategy that looks unusual today. O’Driscoll suggests access and reputation can reinforce success in private markets. Both acknowledge that past performance cannot settle every allocation decision.

The show's mock investment committee turns to Factory, an enterprise coding-agent company that Stebbings says is raising at a $5 billion valuation. He discloses that it is his investment. Lemkin likes the direction of the market and says he would hypothetically back it, while noting that Stebbings knows the company better.

Lemkin says enterprise buyers he met care about data sovereignty and model choice, and expresses his own concern about sharing confidential information with model providers. O’Driscoll agrees that buyers may value separation between a coding product and its model suppliers. Their concerns and predictions do not establish that any named provider trains on or discloses a customer's data.

O’Driscoll calls coding a major source of present AI value and favors exposure across coding, testing, review and related workflows. He sees coding tools as potentially useful even if the broader AI market slows. This is his investment thesis and a hypothetical committee discussion, not evidence of a particular company's future returns.

O’Driscoll's view is that enterprises may want both technical help adopting coding agents and a supplier that gives them model choice. He thinks that demand could continue through a broader slowdown. Stebbings says Factory is already in his portfolio, so his perspective is not that of an uninvolved observer.

Lemkin says he does not know how much of Factory's offering is air-gapped, on-premises or in a private cloud. He recalls the difficulty of introducing external code tools in a previous enterprise role and expects data protection to remain an important customer question. Specific contractual safeguards would need to be assessed separately.

Stebbings turns to legal AI and cites reported revenue and a possible valuation for Legora. Lemkin says he likes both Legora and Harvey but is concerned by a reported negative gross-margin figure for Harvey, explicitly noting that he does not know whether the report is accurate or representative of a sustained period.

Lemkin says a reported negative margin, if accurate and sustained, would make him more cautious about leading a new round. O’Driscoll offers the alternative interpretation that heavy customer usage might drive high short-term costs, but says he lacks the facts to decide on the company. Neither verifies the reported figure.

O’Driscoll discloses that his firm invests in GCAI, another legal-AI business. He says heavy usage could have a positive interpretation even if it temporarily pressures margins, while acknowledging he lacks facts needed to evaluate particular valuations. He later says he would not lead the hypothetical round at the stated price.

The mock committee considers Crusoe, described by Stebbings as a data-center, GPU and managed-inference business raising at a reported $30.9 billion valuation. He cites a large contracted order book. Those transaction and contract figures are introduced by the host and are not independently established in the episode.

Lemkin sees infrastructure access, customer commitments and the ability to build facilities as relevant to the investment case. He says he would pass at the hypothetical price after his own rough valuation exercise. His committee role is part of the show's format, not a real investment decision disclosed here.

Lemkin emphasizes that data-center investing involves projected cash flows, access to power and equipment, customer agreements and financing. His skeptical conclusion on the mock valuation depends on those assumptions. A large reported order book alone does not establish the profitability of a specific project.

O’Driscoll uses the example to make cycle risk more concrete. He argues that leveraged data-center projects may be more exposed to a slowdown in the growth of AI demand than less leveraged software businesses. He has not modeled Crusoe specifically, and his concern is about portfolio exposure to the broader infrastructure trade.

He would examine the quality and duration of customer contracts as well as visibility into debt financing and runway. A lower price and enough financing to withstand a downturn could matter more than optimizing the valuation in a hot market, he says. He also observes that market sentiment can swing rapidly within a single day.

The panel addresses an online dispute involving Keith Rabois, Joe Lonsdale and Airwallex. Stebbings discloses that he is an Airwallex investor and a friend of Rabois, disputes allegations about the company, and says claims changed over time. The episode does not independently establish the disputed allegations or the company's ownership structure.

Lemkin says he dislikes the online attacks but believes some companies and buyers treat exposure to Chinese technology as a business concern. He recounts his own understanding of other transaction discussions. These anecdotes do not establish a connection to Airwallex or verify the claims made about it.

Stebbings reiterates his rejection of a claimed ownership figure and says he has seen the capitalization table. Lemkin says he is not disputing Stebbings's account, while maintaining that perceived geopolitical exposure can influence deals. Their statements are attributed to them; the episode does not provide independent documentation for either position.

O’Driscoll argues that government should establish clear, consistent rules for commerce and security rather than leaving investors to litigate broad policy questions on social media. He acknowledges continuing trade between the United States and China and sees current policy as uneven. This is his policy opinion, not a determination of legal compliance.

Lemkin takes a communications lesson from the dispute and Jev's launch. He praises coordinated product publicity and argues that founders of large companies need credible advocates who can respond to public criticism. He is discussing communication strategy, not verifying the accuracy of allegations against Airwallex.

Lemkin says an effective launch can involve a network of supporters explaining a product publicly. In a dispute, he believes independent advocates can answer claims rather than leaving a founder to respond alone. Stebbings notes that an investor base without a large online presence may make that harder to organize.

Stebbings agrees that advocates matter but notes that some of Airwallex's major investors have little social-media presence. Lemkin replies that founders can develop support beyond their existing investor base. The exchange becomes humorous, but the underlying point is that a founder's network can influence how claims are answered in public.

The panel closes after joking about possible spokespeople and thanking one another. The episode's company figures, reported transactions, legal concerns and investment predictions remain the participants' accounts and should not be taken as independently verified facts or investment advice.