20VC: AI Funding, Model Competition and the Economics of Agents

Harry Stebbings opens with a reported Nvidia–Poolside arrangement, describing a license, investment and movement of engineers. The panel discusses an investor letter said to describe Poolside’s capital constraints. The specific terms and letter are not independently verified in the conversation.

Jason Lemkin reads the reported letter as a reminder that even in a strong AI funding market, a frontier-model company may struggle to finance its next generation of compute. He sees a tension between exciting technical progress and the capital required to continue.

Rory O’Driscoll says the company may have anticipated a real need for a US open model yet encountered costs beyond its financing capacity. In his view, an acquirer or partner with both capital and GPUs can value work that might be difficult to sustain independently.

O’Driscoll argues that a strong exit can still arise from a strategy that did not work as a standalone business. He cautions that the outcome could have been different in a weaker market; a buyer’s appetite and capital are outside the startup’s control.

Stebbings says he is a small Poolside investor and gives an approximate return he expects from the reported arrangement. Both the terms and that return are his statements in the discussion, not independently established results.

The panel considers why an open model could complement a chip supplier’s business. O’Driscoll argues that more adoption of open models running on a vendor’s hardware might support chip demand, although the strategic and financial results remain uncertain.

Lemkin questions whether a 15-times return is sufficient for a seed fund’s best outcome, given high entry prices and dilution. O’Driscoll responds that earning such a multiple from a bet that did not fully achieve its original plan would still be attractive in many portfolios.

They work through the way later dilution can make an investment’s effective entry price much higher than its initial headline valuation. The arithmetic is illustrative; actual returns depend on ownership, transaction terms, timing and fund-level fees.

O’Driscoll argues that building an open frontier model was a rational venture bet if the possible upside was extremely large, even though the company encountered financing limits. Lemkin agrees that such an outcome can be positive while emphasizing the demands of a seed fund’s portfolio math.

The investors say very large model developers have depended heavily on capital from major technology companies. Their estimates of ownership and financing sources are broad claims in a discussion, not a complete cap-table or financing analysis.

Lemkin asks what exit value would be needed to turn the illustrative stake into 50 times invested capital after dilution. The group’s quick mental calculation is hypothetical and should not be treated as verified deal math or a forecast.

O’Driscoll compares the outcome to placing well in a tournament without winning: in a rapidly growing market, a company can create value even if it falls short of the largest possible prize. Stebbings says the reported outcome would still be a win for his investment.

Stebbings moves to reported funding for Mercor and possible Nvidia participation, disclosing that he is an investor in Mercor. The valuation, revenue and investor participation mentioned are reported or estimated in the conversation and are not independently confirmed here.

O’Driscoll sees many of Nvidia’s investments as ways to support the larger ecosystem for computing. He finds the data-company case less immediately connected to chip sales than some infrastructure investments and says it could also simply be a financial bet.

Lemkin discusses a reported analysis of transaction multiples and suggests that the market may tolerate lower gross margins in fast-growing AI businesses. The group’s cited thresholds and company margins are not independently established by the conversation.

The investors speculate about how much capital Nvidia might deploy across customers and partners. O’Driscoll notes that financial results and cash flows can change and that an upcoming earnings report could alter their interpretation; their figures are estimates.

They distinguish an ordinary strategic investment from vendor financing. O’Driscoll says financing a customer can help an ecosystem grow, but it creates exposure if the customer’s demand or ability to repay does not develop as expected.

Stebbings expects large data providers to become very valuable and says he regrets not investing more in Mercor over time. His projected company values depend on assumptions about model-company revenues and spending, and he has a financial interest in the company.

O’Driscoll prefers to reason from the amount of revenue model companies may spend on training data rather than assume a fixed fraction of their market value. The panel offers divergent estimates of current spending and future market size, none independently verified in the interview.

They debate whether a company starting with low margins can improve them as it scales. Lemkin points to examples he believes did so; O’Driscoll warns about survivor bias and says many businesses fail to achieve the hoped-for margin expansion.

O’Driscoll adds that a data supplier serving a few large customers may have different bargaining power from an application with many buyers. He says the eventual margin structure is uncertain even if demand for data grows.

Lemkin encourages investors to ask what happens if a promising company exceeds expectations, while O’Driscoll adds that a partial success in a strong market can also return capital. Both frames are ways to evaluate risk, not assurances of a profitable exit.

Stebbings raises a reported statement about a possible OpenAI IPO in 2027 and reported plans involving Anthropic. The speakers analyze what such plans might mean for capital and competition; the interview does not independently confirm a listing, date or valuation.

O’Driscoll compares reported quarterly growth with figures attributed to a competitor and suggests OpenAI would want to communicate any acceleration. He explicitly says he cannot be sure until more formal financial figures are available; his explanation of the company’s communications is interpretation.

His concern is relative growth: a very fast-growing business can still lose share if another grows faster. The panel’s revenue and run-rate figures are estimates from reports and should not be treated as audited, directly comparable results.

Asked what price OpenAI might achieve, the investors speculate about ranking, growth and market conditions. O’Driscoll says he does not know a valuation. Their scenario is conditional on possible IPOs and future performance, neither of which is established here.

Lemkin argues that enterprise buyers increasingly want choices among models. He sees competitive pressure from open-weight systems on a frontier provider that is not the preferred model for a particular use. His characterization of current market rankings is an opinion based on the information he cites.

O’Driscoll expects many tokens eventually to run through open-weight models while frontier providers retain a substantial share of revenue. He treats this as a forecast about price and value, not a measured current or future split.

The two agree that a lower-cost competitor can pressure margins and force incumbents to explain their value. O’Driscoll speculates that a second-place provider may have less freedom over IPO timing and pricing, depending on a competitor’s actions and public-market conditions.

Lemkin asks how OpenAI would describe its differentiated mission to employees and investors as the market becomes crowded. This is his question about positioning, not a finding that the company lacks a mission or that any reported internal statement is authentic.

Stebbings emphasizes ChatGPT’s consumer recognition and imagines potential advertising and hardware opportunities. The panel contrasts that broad reach with the commercial value it attributes to coding workloads, while acknowledging that their paths and economics could change.

O’Driscoll argues that coding has been an unusually fast-adopting, high-value use case for AI. Lemkin says some heavy consumer plans may be subsidized. Their examples of user costs and company margins are estimates, not audited unit economics.

The investors consider whether consumer AI could eventually become a large business as serving costs decline and monetization develops. That is a scenario dependent on future costs, adoption and business models rather than an established result.

They react skeptically to a reported multi-trillion-dollar addressable-market assumption attributed to Anthropic. The figure is not independently confirmed in the conversation; the discussion illustrates how sensitive valuations can be to very large market assumptions.

Stebbings raises the possibility of a Hugging Face transaction at a reported price. Lemkin says he cannot evaluate that valuation, while O’Driscoll considers why a strategic buyer might value a platform associated with open models. No sale or price is confirmed in the interview.

O’Driscoll says an enterprise technology company might want a position in open-weight infrastructure as a counterweight to frontier providers. Lemkin speculates that current interest in open models could create an attractive selling window, but also acknowledges future growth could differ.

They recognize that founders may value a platform’s mission as well as a financial exit. Lemkin argues that a buyer would need to preserve user trust and neutrality; his predictions about how a hypothetical acquirer would behave are unverified speculation.

The panel turns to a reported sale by Citadel of positions associated with a leveraged AI investment. The speakers treat it as a short-term trading decision, but the interview does not establish the transaction, prices or positions independently.

O’Driscoll uses the example to distinguish owning an asset outright from using leverage: debt can force decisions before a long-term thesis plays out. He connects the same principle to vendor financing, where a supplier bears credit risk as well as demand risk.

The investors discuss volatility in Korean semiconductor-related markets and changes in share prices. Their index levels and return figures are presented conversationally and may be incomplete or dated; they are not investment recommendations.

They also talk about wealth concentration and rising housing costs in the San Francisco Bay Area. Their rent and wealth figures are anecdotes or estimates. O’Driscoll expects a future correction is possible but thinks local costs may remain above prior levels.

Stebbings asks how long the AI boom can continue. Lemkin expects a prolonged cycle with gains and wealth concentrated among fewer companies and workers. O’Driscoll says the important constraints may be available capital and demand for the end products.

O’Driscoll expects major infrastructure suppliers to keep investing while demand appears strong. He suggests public-market capital could extend financing but says the more important question is whether businesses will spend enough on useful applications to support projected revenue.

The panel discusses very large reported future revenue expectations for frontier labs. Those are projections attributed to companies or reports, not established demand. O’Driscoll says the pace at which use cases spread beyond coding will matter to the outcome.

Lemkin says some intensive users have become dependent on AI tools and may resist losing access, while companies are beginning to confront unexpectedly large bills. He calls this a form of ‘token addiction’; it is his characterization, not a diagnosis or a measured population trend.

O’Driscoll compares AI consumption to capital that must be priced and allocated. Unlike a fixed number of software seats, agent usage can grow with activity, so an organization may need budgets and spending controls without eliminating access altogether.

He gives a hypothetical example in which an AI bill consumes a meaningful share of a bank’s profit. A finance leader would need to decide where the spending produces value and how to fund it; the illustration does not establish actual costs for any bank.

The speakers debate whether greater AI spending could lead companies to reduce other costs, including staffing. O’Driscoll says an organization cannot assume every productivity investment raises profit immediately; Lemkin emphasizes the risk of losing valuable employees if they lack tools they want.

Lemkin says some finance leaders worry about retaining staff when well-funded AI companies can offer strong compensation. O’Driscoll agrees that retaining the right people matters, while returning to the question of how an employer pays for their additional tooling.

Lemkin argues that both casual and expert users may increasingly depend on AI. Stebbings cautions that their Silicon Valley circles are not representative of the entire population. They disagree on how quickly broader adoption will follow.

Stebbings describes a legal-work example from someone he knows, while also saying AI media tools still fall short for his needs. The panel contrasts rapid adoption in some tasks with uncertainty about how quickly useful automation will spread across other work.

The group discusses reported Stripe growth and billing figures and attributes some momentum to AI-related activity. These figures are not independently verified in the interview, and the speakers’ interpretation does not establish the cause of the growth.

O’Driscoll describes a hypothetical attraction of a diversified payments business with an AI growth tailwind: if one source slows, another may still generate cash. He notes the company is private; the scenario is an investment opinion, not a recommendation.

Lemkin predicts that if fast-growing private technology companies list, they could reshape public-software comparisons. O’Driscoll says private status currently prevents many public investors from owning those businesses directly. Both are discussing possible future market effects.

They point to reported stock buybacks and the ability of some private companies to finance acquisitions without listing. The conversation does not independently verify specific transaction terms, and a few large companies are not representative of all private businesses.

Stebbings turns to reported security concerns involving AI assistants with broad access to email or other data. Lemkin argues that agent safeguards remain incomplete. The interview does not independently establish the details or scope of any particular reported incident.

The panel describes an assistant as a tool that may act like a chief of staff when given authority. Lemkin says the value of delegation must be weighed against potential disclosure or unwanted action, and he would be interested in a product that reliably addresses those risks.

Stebbings expects users eventually to trust agents with sensitive financial information, comparing the shift to earlier discomfort with online payments. Lemkin agrees the problem may be solvable in principle but says he does not regard it as solved today; future security cannot be assumed.

Lemkin recounts an agent attempting an unwanted purchase as an example of goal-seeking behavior. It is his anecdote, not a controlled evaluation of a product. The group notes that broad permissions and fast repeated actions can magnify the consequences of mistakes.

O’Driscoll questions whether highly individualized personal productivity is the easiest market for agents. A repeatable enterprise process, he suggests, may have clearer budgets and measurable savings than a person’s idiosyncratic inbox or calendar.

He says technology investors may overestimate how many people want to optimize every personal task. Productivity products can have real value but may be difficult to get right or distribute broadly; this is his market judgment rather than a universal description of users.

The panel mentions existing productivity products and notes that some moved toward organizational use. O’Driscoll asks both whether an assistant can perform reliably and how large the paying market will be, leaving both questions open.

Asked which AI investment categories may be overheated, Stebbings questions customer-support software. He predicts consolidation and says some large technology companies may build their own systems. Lemkin discloses investments in the area and does not fully agree with the implication for returns.

O’Driscoll names defense as an area where he thinks venture returns could be difficult despite useful products and spending. He predicts that a few large platforms could acquire smaller suppliers because complex government markets may reward broad portfolios and customer relationships.

Stebbings also questions humanoid robotics, citing the difficulty of dexterity and real-world execution. O’Driscoll distinguishes specialized robots already deployed for a defined purpose from the broader humanoid vision, and says his skepticism could prove wrong.

Lemkin shares skepticism about traditional customer-support categories and about applying venture capital to transform professional-services firms. His claim that some categories could merge or shrink is a prediction, not evidence that existing services have already disappeared.

O’Driscoll admits that, despite category-level doubts, an exceptional team and persuasive evidence could change his mind on an individual deal. He also notes portfolio construction: one investment does not need to determine an entire fund’s exposure.

Lemkin closes on the possibility that AI and other changes enable unusually creative new businesses. The panel agrees the opportunity is substantial while leaving unresolved which strategies and valuations will work; the conversation is not financial advice.