AI Financing, Venture Scale and the Risks Behind Growth Valuations

Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss reported AI financing, competition between model providers, the economics of venture funds and a proposed California wealth tax. Their valuation calculations, market-share judgments and policy predictions are opinions made at the time of recording, not independently verified facts or investment or tax advice.

Stebbings begins with a reported $10 billion Anthropic financing at a $350 billion valuation and asks whether it could be the last round before an IPO. O’Driscoll thinks an IPO is plausible because of the company’s stated ambition, but does not present a confirmed listing or financing outcome.

O’Driscoll contrasts the reported current valuation with an earlier round and rapid run-rate growth. He cites roughly $100 million at the end of 2023, $1 billion at the end of 2024 and an alleged $9 billion–$10 billion at the end of 2025; the last figure is explicitly an assumption, not audited revenue.

In a hypothetical threefold rise from a $10 billion to a $30 billion annualized run rate, he averages the opening and closing figures to estimate $20 billion in annual revenue. He says the implied multiple might look attractive if exceptional growth continues; an average of endpoints is a simplifying assumption, not a revenue forecast with evidence of realized results.

Lemkin sees the reported fundraising size as potentially consistent with strong unit economics and argues that model-based coding and application creation are still early markets. This is his subjective reading of company economics and adoption, not proof of margins or ownership of the enterprise market.

He says Claude appears strong at the enterprise API level and points to products built using models. O’Driscoll separates the API business, a first-party coding application such as Claude Code, and a possible AI workspace for other knowledge workers; his market-share comparisons are estimates rather than measured results.

O’Driscoll argues a model provider that sells an application can capture more of the customer relationship than one supplying only an API. He also notes competition from Cursor and GitHub, and distinguishes coding tools from a newer workspace concept he had not yet tried.

He imagines a workspace bringing documents, spreadsheets and presentation tools into an AI environment, rather than adding AI separately to each application. He calls the opportunity significant but says early feedback suggests the product is not yet polished. O’Driscoll mentions an investment in an adjacent company, ObviousAI.

Stebbings says product leaders he has spoken to increasingly mention Claude Code instead of Cursor and asks whether investors should worry. These conversations are anecdotal, not a representative measurement of usage. O’Driscoll says the answer depends partly on entry price and the intensity of competition.

O’Driscoll says Cursor faces Microsoft and a model supplier that also competes with it. He would be more nervous at a reported $27 billion pre-money valuation than at an early entry price, while praising the product. The discussion does not independently verify Cursor’s financials or future valuation.

Lemkin says rapid product change makes disruption a routine risk for AI investors. He speculates that Anthropic might someday limit competing applications’ access to its strongest models or build similar interfaces, while acknowledging that supplying those businesses can itself be valuable; no such future restriction is established.

The group discusses Google’s Gemini role in Siri and the strategic value of distribution on Apple devices. O’Driscoll is unsure how money flows in the AI arrangement and distinguishes it from advertising-funded search placement. His discussion of possible payment directions is speculation, not a disclosed contract.

Lemkin argues privacy and deployment requirements can matter deeply to large enterprises and cites Apple’s privacy rationale. He observes that security expectations may be higher than startup suppliers assume. The discussion does not independently assess the security of the vendors mentioned.

Stebbings asks whether OpenAI is vulnerable to competition from Anthropic and Google. The guests distinguish OpenAI’s commercial enterprise from its nonprofit shareholder; they disagree on how precarious the business is and discuss the effects a loss of value could have on its stakeholders.

O’Driscoll argues that OpenAI has a substantial consumer audience, subscriptions and product appeal, so he considers a zero valuation implausible. He uses relative private valuations as an informal comparison of competitive progress; those valuations and the proposed measure of management success are not audited operating metrics.

Lemkin presents a downside scenario: rapid model improvement coincides with a capital-market shock, leaving OpenAI unable to fund the next generation while rivals continue. His claim about model shelf life and future capital needs is an estimate and scenario, not evidence that such a shortfall will occur.

O’Driscoll suggests maintaining a substantial cash runway and adjusting ambitions if financing tightens. Lemkin responds that slowing development may itself damage competitiveness. They disagree over consumer stickiness and the severity of a funding shock.

The debate turns to whether consumers would abandon an older model when newer alternatives improve. O’Driscoll argues the worst case requires several adverse conditions at once, including weak macro conditions and continuing rapid model advances. Lemkin thinks venture markets may be underestimating the possibility of a downturn.

Stebbings cites a claimed decline in ChatGPT usage after new Gemini releases, and the guests mention personal choices by family members. Neither the figure’s source nor its measurement is established in the discussion; the anecdotes cannot determine broader retention.

O’Driscoll summarizes the disagreement as a question of combined probabilities: a funding shock, low cash and a fluid market could cause trouble, but he considers the combination unlikely. He believes a more probable response would be to narrow ambitions and build on the consumer business.

The conversation shifts to a16z’s reported $15 billion fundraising and whether venture is moving toward giant platforms and small specialists. Stebbings cites a share of venture fundraising, while Lemkin praises the firm’s founder brand and reported returns; the show does not independently verify those estimates.

O’Driscoll separates two questions: can the industry’s total invested capital earn attractive returns, and can a single firm deploy a large share successfully? He treats the industry as roughly in equilibrium, a judgment dependent on assumptions about fundraising, exit values and ownership.

He illustrates his argument with an implied $75 billion raised industrywide if $15 billion represented 20%, and roughly $300 billion in exit value. He explicitly notes that not all exit value belongs to venture investors, so the arithmetic is a rough illustration rather than proof of a 3x fund return.

If the firm raised on a two-year cycle, O’Driscoll estimates a sustained share closer to 10% and argues it would need meaningful participation in successful early and later rounds. He cites a study of large Series A outcomes as context; neither the study nor the projection guarantees a future share.

He identifies operating risks: a larger team must keep investment selection quality high, and a larger fund has to participate in the relatively few companies that account for much of private-market value. Missing several exceptional outcomes could make the return math harder.

Lemkin argues a strong brand and broad sourcing team could see most important deals and asks whether a platform could target a much larger share. He proposes managing conflicts between investments with separate teams; his claim that conflicts can be solved is an opinion, not proof that founders face no risks.

O’Driscoll cautions that adding more investment decision-makers can lower the hit rate, even if a larger firm participates in more winners. He compares scalable index investing with stock picking and argues selection can become harder as the number of partners and deals rises.

Stebbings asks whether smaller early-stage firms need a large growth fund to compete. O’Driscoll says firms can succeed with focused early funds, while a late-stage vehicle can offer follow-on capital, build profile and support a broader platform. He views this as one viable strategy rather than the only one.

Lemkin doubts most Series A founders choose investors mainly for the promise of later financing; he thinks immediate help and terms matter more. O’Driscoll replies that a history of supporting companies over time can help on the margin, even if it is not decisive.

O’Driscoll outlines a different advantage of a growth fund: a firm can make several early bets, accept some losses, and later concentrate much larger sums in winners. His numerical example is illustrative. Such follow-on investing also creates its own price and execution risks.

Stebbings says access to later capital may make a firm more willing to pay a higher early valuation, and repeats the thesis that middle-sized venture firms could be squeezed. O’Driscoll agrees that giant funds increase competitive pressure but challenges the claim that every focused middle-sized firm must disappear.

O’Driscoll says his firm concentrates on enterprise software and typically makes 20–30 investments per fund. At its stage, he argues, check sizes and reserves require a meaningful fund, while a specialist must know its market and identify opportunities before a large platform reaches consensus.

He points out that a16z has organized capital into sector-specific funds, which he likens to boutiques operating under a shared brand and backed by late-stage resources. He argues independent specialists can still compete if they bring differentiated insight, though he acknowledges the platform’s advantages.

Lemkin asks whether small investors can still find exceptional companies outside increasingly efficient accelerator and venture networks. He notes a reported early rejection history for Anthropic but treats the broader question—whether overlooked opportunities remain—as unresolved.

He describes a possible niche in companies whose growth falters and later accelerates, calling this a ‘glitch in the matrix.’ He cites an investment in Owner as an example and notes his own investment interest. O’Driscoll agrees that recognizing such reversals before others is difficult.

O’Driscoll offers estimates of accelerator participation among unicorns and leading investors’ presence by financing round. The figures are his observations without a stated complete sample or method. He argues that historically prominent companies have sometimes struggled for early venture acceptance, so broad deal coverage alone does not ensure good picking.

The group debates whether AI startup growth has become unusually obvious and linear. Stebbings cites several European companies as examples; O’Driscoll warns that confidence in uninterrupted growth can itself be dangerous. Their observations are selective and not a tested industry-wide trend.

O’Driscoll identifies the unresolved pricing risk in strategies that deploy substantial money into late rounds. When an opportunity looks obvious, competitive investors may pay high multiples. If growth slows and public-market multiples contract, strong execution may not protect a late investor’s entry price.

He uses Databricks as a hypothetical illustration, citing an approximately $100 billion valuation and comparing a high-growth multiple with a lower multiple for slower growth. These are rough assumptions, not a valuation of the company or a prediction that its growth will decline.

He contrasts early-stage business risk, which can differ company by company, with late-stage valuation risk that may move across many investments at once. His concern is that even a modest change in growth or multiples could produce correlated losses if investors have leaned on similar assumptions.

Lemkin describes building a founder simulation and integrating ElevenLabs voice into it. He praises the product but says a small number of players consumed about $30 in credits over a short period. This personal experiment illustrates a potential cost sensitivity, not a representative customer study.

Asked whether he would invest at an $11 billion valuation, Lemkin praises the founder and sees a possible large voice market but says he would need more work at that price. He already considered lower-cost alternatives for his own game; a single user’s experience cannot establish the company’s retention or economics.

He adds that spending on several AI services may prompt customers to substitute bundled or cheaper tools. He compares a paid authentication product with a feature included in a development platform. His expectation that substitution will become more important is a forecast, not evidence that these vendors have lost a particular customer.

O’Driscoll recalls evaluating a speech company whose large customers said they would replace it, only to remain years later. The anecdote cuts against assuming stated switching intentions become actual departures. He suggests larger customer spending may nevertheless change the incentive to substitute.

Thinking aloud about ElevenLabs, O’Driscoll illustrates how a $30 billion outcome might require substantial annual revenue at an assumed public-market multiple. He suggests a diversified base of many smaller customers could face different pricing pressure from reliance on a few very large buyers. The numbers are illustrative and depend on unknown future outcomes.

Lemkin reiterates that he found the ElevenLabs API easy to use and estimates a higher monthly bill if his game scales. O’Driscoll notes that ease of adoption may also help retain users if alternatives do not match it. Neither statement establishes comparative product quality or future margins.

Stebbings asks how a proposed California wealth tax might affect the startup industry. O’Driscoll says he has not read the core text and offers views about mobility and assessment of voting control. His description is not a verified account of the proposal, enacted law or tax liability.

O’Driscoll predicts wealthy residents could relocate and questions whether the measure would raise its projected revenue. He acknowledges why the public might be unsympathetic to wealthy taxpayers. His claims about other countries and expected fiscal effects are opinions offered without a detailed comparison.

Lemkin argues that proponents might pursue an annual tax and lower thresholds later, and predicts founders could leave before later funding rounds. These are claims about political intent and future proposals, not established features of the measure or proof that such changes will pass.

O’Driscoll expects the proposal may fail at the ballot but believes even discussion can affect relocation. He leaves before the end of the show. Both investors’ economic and political predictions are uncertain, and their personal financial interests are relevant to their views.

Lemkin says a California ballot measure could pass with a majority and reiterates his concern about a future annual levy on illiquid startup holdings. He presents a scenario in which founders build a team in San Francisco and later relocate; the interview does not establish that this would happen or verify the tax calculation.

Asked where people might go, he suggests cities such as Miami or Austin. He says he has considered leaving California and would consider it more seriously under a recurring tax at lower thresholds. This is a personal hypothetical, not a report that he has moved.

Lemkin also says technology can adapt through distributed teams, though he sees costs to leaving an AI cluster. Stebbings and Lemkin then raise a broader concern that AI gains and wealth could become concentrated while some workers experience displacement.

Lemkin speculates that some software businesses may eventually generate $1 million–$2 million in revenue per employee and require fewer staff. He cites company examples but offers no systematic employment analysis. Productivity, future staffing and job displacement remain uncertain.

The pair discuss claimed Nvidia employee wealth and the possible effects of concentrated gains on housing and social attitudes. The employee-wealth numbers and housing assertions are not independently sourced in the interview; they should not be treated as verified statistics.

Lemkin worries that layoffs and long job searches could increase resentment if the benefits of AI are unevenly shared. Stebbings shares concern about disparities. They close by acknowledging the risk of social strain, which remains a prediction rather than a demonstrated outcome of a particular tax or product.