Harry Stebbings, Jason Lemkin and Rory O’Driscoll on AI valuations, agent economics and venture exits

Harry Stebbings opens with reports that OpenAI’s proposed fundraising valuation and round size had been reduced. He asks Jason Lemkin and Rory O’Driscoll whether this means the AI investment boom is cracking.

One response is that investors are comparing a smaller, loss-making private company with a larger, faster-growing company approaching a public listing. A lower private valuation can reflect that comparison without resolving the broader AI-demand question.

Lemkin says market volatility still warrants caution. He points to falling AI-related stocks, financing costs and what he sees as slower incremental growth at Anthropic, while acknowledging that the timing of individual events differs.

O’Driscoll distinguishes price discovery and financing risk from the demand side. If demand for AI capacity remains strong, today’s financing worries may fade; if demand weakens, those same signals could look more consequential.

The panel considers CoreWeave as an example of infrastructure financing exposure. They argue that a highly leveraged specialist can be more vulnerable to weak demand than a large cloud provider with a diversified balance sheet.

Stebbings raises Alphabet’s sharp stock decline alongside strong cloud growth, slower search growth and negative free cash flow. The investors discuss concerns about capital expenditure and the possibility that AI products could cannibalize search.

The panel says high cloud growth and market anxiety can coexist. Investors may react sharply to a small miss when expectations are high, even if the company’s underlying business is still expanding.

The discussion turns to Anthropic’s potential public offering. Lemkin expects investors to scrutinize sequential growth and model pricing, arguing that rival offerings and open-weight models may put pressure on premiums.

O’Driscoll walks through approximate quarterly revenue figures and explains the distinction between absolute additions and percentage growth. Adding a similar dollar amount to a larger base naturally produces a lower percentage growth rate.

From one perspective, sustained rapid quarterly percentage growth remains extraordinary; from another, a decline in the growth rate raises questions about what comes next. The same starting numbers do not settle the forecast.

Lemkin focuses on incremental bookings and says deceleration would concern him in a smaller startup. O’Driscoll clarifies the difference between the rate of growth slowing and the company’s total revenue falling.

The panel considers whether slower growth would reflect market saturation or pricing pressure. Lemkin favors the pricing explanation but presents it as a thesis rather than a conclusion established by the data.

They compare possible annual growth scenarios for Anthropic. A company can be adding enormous revenue while its growth multiple falls substantially; investors must decide what future trajectory is already reflected in the valuation.

Lemkin offers an anecdote from an application he is building: after changing models, he says the cost of a workflow fell by roughly 70% while quality improved. He treats it as one example, not a general price index.

The investors discuss how fast-changing model quality and prices make multi-year forecasts difficult. Lower unit costs may coexist with more agent usage, leaving total spending and margins uncertain.

Lemkin says he stopped using Canva for one personal workflow and speculates about pressure on established software products. He acknowledges uncertainty about reports concerning the company’s growth and possible changes.

O’Driscoll argues that a software company that stops adapting can continue serving an installed base for years but may face gradual obsolescence. The panel asks whether incumbents can adopt AI quickly enough to keep growing.

Lemkin says he expected mainstream customers to be less affected by new AI tools than heavy agent users, but believes some established products are already feeling pressure. He describes that assessment as a surprise and a source of concern.

The conversation moves to consumer agents and the cost of persistent virtual machines. The panel questions whether a subscription priced at tens of dollars per month can support heavy, continuously active users.

Stebbings suggests average costs may be lower than the cost of power users because many customers use a service only occasionally. The others note that the answer depends on the user mix and how infrastructure is managed.

One investor compares today’s economics with earlier digital products that grew before their infrastructure costs were attractive. A company may bet that falling compute costs will improve margins after it has acquired users.

Lemkin counters that subsidies cannot continue indefinitely for every provider. The panel discusses whether companies can develop cheaper models or optimize infrastructure before available capital runs out.

O’Driscoll says venture-backed competitors can launch before unit economics work, capture demand and rely on a future cost curve. Waiting until economics are already perfect may mean entering after the market has been taken.

The panel returns to limits on heavily subsidized AI services. They debate the opportunity cost of providing compute to consumers when the same capacity could be sold to enterprise customers.

Stebbings introduces DoorDash’s approximately $17 billion agreement to acquire Instacart. The proposed transaction is subject to closing conditions; the speakers discuss consolidation and the price relative to Instacart’s earlier private valuation.

The investors compare the deal with other platform acquisitions. They debate whether combining delivery networks can add meaningful value and note that market share and antitrust scrutiny matter to the outcome.

The conversation turns to DoorDash’s statement that relatively few orders came through agents compared with its overall delivery volume. The panel asks whether the small number reflects an enduring preference or an early stage of adoption.

One argument is that agents could compare delivery providers, prices and times, weakening the platform’s direct relationship with the customer. Another is that existing apps are already convenient and users enjoy browsing food.

Lemkin describes a possible alternative: an agent could identify a restaurant’s own delivery or pickup option and avoid some platform fees. He sees disintermediation as a potential competitive threat, not an established outcome.

The panel notes differences between US and UK delivery fees and argues that platform economics depend on convenience, network effects and the division of costs between consumers and restaurants.

O’Driscoll is more skeptical that an agent will improve every transaction. For taxis and food discovery, he says a visual app may be easier; recurring or scheduled orders are a clearer automation case.

Stebbings moves to reported financing rounds for Kalshi and Cohere. The panel contrasts a fast-growing prediction-market business facing legal challenges with a model company positioned around sovereign AI.

The investors debate the social purpose and regulatory treatment of prediction markets, including sports-related trading. Their remarks describe opinions about the business and ongoing legal issues rather than findings of wrongdoing.

For Cohere, the panel discusses Canadian government backing and the appeal of domestic AI capacity. Canada announced a staged equity commitment of up to C$1.6 billion; the speakers also question whether government support will translate into sustained private-sector demand.

They distinguish a government’s strategic motivation from an investor’s financial return. Public support may keep a provider in the market, but the panel says enterprise adoption and long-term economics remain important.

The investors discuss reported losses in the crypto sector and compare the ability of exchanges to absorb a shock. They acknowledge uncertainty about private balance sheets and debate the distinction between customer assets and company assets.

The discussion shifts to AI-assisted penetration testing. Lemkin describes a tool intended to test software defenses and worries that the same capabilities can be redirected against other organizations.

The panel frames the issue as dual use: security teams need to test their own systems, while attackers may use similar tools at scale. They speculate about future disruption to essential services rather than identifying a specific verified future attack.

Stebbings asks the investors to choose among three proposed opportunities: a coding-agent company, an inference provider and an enterprise service application. The choices have very different valuations, business models and capital requirements.

Lemkin initially makes a case for Cognition, citing the strength of demand for coding agents and enterprise adoption. He also raises the difficulty of earning a large multiple from a very high entry valuation.

O’Driscoll explains why comparing three deals is more informative than evaluating a single company in isolation. A coding agent, an inference provider and an enterprise application expose investors to different markets and competitive pressures.

Stebbings favors Serval at the lower stated valuation, arguing that the potential return multiple looks more attractive than at a much higher entry price. He notes a fund-size constraint: a large fund may need to deploy more capital than a smaller round can absorb.

Lemkin asks whether a service-management product can expand the total addressable market or only take a share of an incumbent’s existing budget. He says that distinction is crucial when evaluating vertical AI applications.

O’Driscoll suggests a test: compare what a similar customer pays an incumbent with what it would pay for an agent that performs work rather than merely tracking it. Higher willingness to pay could support a larger market.

The panel notes that service management has grown beyond a narrow help desk into a broad system for corporate work. Replacing an incumbent could be valuable, though building that position may take years.

They contrast a long compounding opportunity in enterprise applications with the faster spending velocity they see in coding agents. Lemkin warns that many vertical AI startups may face more competitors without expanding customers’ overall budgets.

Inference infrastructure can generate revenue quickly, but the investors also see intense competition and a risk of commoditization. Stebbings names other inference providers he considers strong.

O’Driscoll sums up the choice: horizontal inference with competition, an enterprise coding agent at a high price, or an application with a more visible incumbent market. He says the comparison makes him reconsider the tradeoffs.

The conversation turns to the rapid collapse of a large law firm and an AI-oriented legal company’s interest in its assets. The panel treats the law firm’s partner structure and AI legal services as related but distinct stories.

One investor questions whether an AI-assisted law firm can retain profitable lawyers if traditional firms offer higher compensation or partnership economics. Another distinguishes bespoke advice from more standardized transactional work.

For routine documents, the panel debates whether automation could reduce errors or whether inadequate legal work can still be costly. They do not resolve whether the proposed model can deliver consistent quality.

O’Driscoll argues that some lawyers value predictable hours and a good salary over becoming a highly paid rainmaker. He still questions how much corporate value can be built after compensating the lawyers who do the work.

The group agrees that an AI-assisted legal business may be viable while disagreeing about its potential scale. Venture returns depend on which work can be standardized and how much value the company can retain.

Stebbings raises Physical Intelligence’s reported fundraising discussions and compares a high robotics valuation with an established software company’s more modest valuation. The panel says the two businesses are priced on very different expectations about the future.

O’Driscoll is enthusiastic about recent robotics model demonstrations, particularly dexterity. He cautions that deployment in factories and other physical settings usually takes longer than deploying a language model.

He questions the revenue required to support a very high robotics valuation when the installed base of robots is still limited. The technology may be impressive while the pricing and deployment model remains unsettled.

The panel suggests that major model providers might build or acquire robotics capabilities to address physical work as well as knowledge work. That is an investment thesis about possible future demand, not a forecast of any particular deal.

O’Driscoll acknowledges a tension: a detailed deployment analysis can make the opportunity look slow, while investors may be drawn to a major technical advance and a capable team before the commercial path is clear.

Lemkin says venture investors need acquisitions as well as public listings to return capital. He mentions reported scrutiny of a large AI transaction and says he is watching whether merger review slows exits.

He points to Salesforce’s recent acquisition activity as evidence of demand from strategic buyers. Salesforce had announced an agreement to acquire Cresta; the panel discusses the value of recurring exits for venture portfolios.

The investors describe acquisitions as a way for a large company to add capabilities and move into new markets. Lemkin suggests different business-unit leaders may identify targets, with some acquisitions expected to work better than others.

They compare buying a more mature company with existing enterprise customers against buying a newer AI-native startup. The newer business may grow faster but command a much higher multiple of current revenue.

Using the examples discussed on the show, the panel argues that the buyer must decide whether established scale or an early technical position is worth the price. Reported revenue and transaction figures are estimates used in their discussion.

Lemkin says large buyers often look for around $100 million of revenue before committing a major distribution organization to an acquisition. Smaller technology purchases can happen, but integrating them may be harder to justify at scale.

The panel agrees that a revenue threshold can help an acquirer see how a target could become material within a very large business. Stebbings closes the discussion and thanks the guests.