20VC on IPO Expectations, Lean AI Startups and Investor Risk

Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss a week of reported IPO plans, AI company growth and fundraising stories. Their conversation took place before the anticipated SpaceX listing; the transaction details and market predictions they discuss should be understood as claims and expectations at that time, not confirmed outcomes.

The panel starts with the reported terms of a SpaceX offering. Stebbings cites a possible $75 billion raise and roughly $1.77 trillion valuation; the guests discuss a reported $135 fixed price, heavy demand and a substantial retail allocation. They treat these figures as information circulating before the proposed listing, not independently verified deal terms.

O’Driscoll says the unusual scale and distribution of the proposed offering would make it difficult to predict the first day. He considers several scenarios, including an early jump or disappointing trading, and cautions that apparent demand is not the same as durable buying once the shares are available.

Lemkin draws a distinction between the excitement around an iconic company and the price investors pay to participate. He thinks the shares could start slowly and appreciate over time; O’Driscoll offers a different view and expects a decline within a year. Both are personal forecasts, made before the anticipated listing, not results.

The guests debate whether setting an offering price instead of maximizing the initial valuation could broaden participation and shape the market after the launch. A large allocation to retail investors might support enthusiasm, but it also adds uncertainty about how much demand reflects long-term conviction rather than an attempt to obtain an immediate gain.

O’Driscoll argues that the size of the proposed sale matters for the rest of the market: an unusually large issue can absorb attention and capital at the same time as other companies are considering offerings. The panel does not claim to know whether this would reduce demand for other stocks or how long any effect might last.

The discussion turns to who owns private-company shares and how IPO liquidity reaches venture funds and their limited partners. A headline company valuation does not itself tell an investor how much a particular fund made: the entry price, ownership, dilution, timing and ability to sell all affect the outcome.

The speakers compare institutional and retail demand and question how much apparent oversubscription will turn into actual orders. They distinguish an expression of interest from cash committed at a particular price, and acknowledge that the distribution of shares can change the dynamics of the market after a listing.

They also discuss the role of early backers and sovereign or pension investors in large private rounds. O’Driscoll corrects a reference to Ontario Teachers rather than an Ohio fund. Their examples are drawn from the conversation and are not a verified accounting of any investor’s holdings or returns.

Lemkin returns to the emotional appeal of a famous company going public. The panel agrees that a compelling story can attract buyers while disagreeing on how much the proposed price already assumes. A strong business and a good investment at a given price are separate questions, they argue.

The guests speculate about how a large listing might affect other private companies weighing public markets. O’Driscoll sees the scale as unprecedented in his experience, while acknowledging that enthusiasm, allocation decisions and the subsequent price are inherently uncertain before the transaction.

Stebbings presses the guests for a near-term price call. Rather than treating one outcome as inevitable, they lay out different possibilities and revisit how institutions, retail buyers and holders seeking liquidity may behave. Their disagreement illustrates how little the announced or reported price alone establishes about future returns.

The panel considers the difficulty of placing a very large amount of stock and the possibility that the company deliberately leaves room for the market to determine the value. Even with a popular brand, the amount of capital required and the make-up of buyers could matter as much as the first-day narrative.

The speakers widen the discussion to the liquidity needs of funds and their backers. Distributions from an exit depend on when shares can be sold and how a fund manages its portfolio; a paper mark is not the same thing as cash returned to investors.

As the SpaceX discussion closes, the panel remains split on both first-day trading and the longer horizon. O’Driscoll emphasizes the danger of paying an ambitious price, while Lemkin gives more weight to the company’s long-term potential. Neither claims to have observed the eventual market outcome.

Stebbings moves to a report that OpenAI had filed toward going public. The panel discusses possible timing and why an AI company might seek public capital, but the captions provide no filing or documentation to establish the report. Their suggestions about motives and timing are speculation, not confirmation of a transaction.

The guests contrast the capital requirements of frontier-model developers with those of conventional software companies. In their view, spending on computing infrastructure can make access to large pools of capital particularly valuable, though public listing decisions also involve disclosure, timing and the preferences of existing shareholders.

The conversation shifts to a reported OpenAI memory update. The panel imagines an assistant that retains context between conversations and develops a more useful understanding of a person over time. They discuss this as an emerging product direction, not as independent verification of the capabilities or availability of a particular release.

The appeal of persistent memory, the guests say, is that users would no longer need to repeat their preferences and history in every interaction. Such context could improve relevance, but it also raises the question of what information is retained and how much control a user has over it.

The panel considers the economics of always-on context: additional processing and storage can make a more personalized service costly to deliver. The speakers do not have a complete cost model, so their comments are hypotheses about the product trade-off rather than measured unit economics.

On a reported plan for Siri to use Google’s Gemini, the guests debate whether Apple should use a partner’s model or build more of the stack itself. One view favors a practical path to a better assistant; another emphasizes the strategic risk of relying on a competitor for a core capability.

Apple’s device distribution and knowledge of the user could give it an advantage in a more personal assistant, the guests suggest. But they also stress that OpenAI and Google are competing for the same user relationship. Their assessments describe potential strengths and risks, not evidence of a settled winner.

The discussion connects memory, trust and privacy to the battle for consumer AI. The most useful assistant might need deep context, but users must be comfortable with how that context is handled. The panel treats the product experience and the trust relationship as important parts of the competition.

Stebbings introduces a reported 23% cut in Uber’s HR organization and a return-to-office change, noting that the company denied AI caused the cuts. Lemkin says recruiting and HR are often among the first functions reduced during a slowdown, although companies sometimes find they need recruiters again later.

Lemkin believes AI could improve some HR tasks, including examining a broader history of work and identifying problems that a biased human might miss. He explicitly says he is not proposing to remove all humans. These are his expectations about possible uses, not proof that automated HR decisions would be fairer.

O’Driscoll guesses that some automation may be involved in workplace efficiency gains, but doubts that a 23% reduction can simply be attributed to AI. He notes the alternative possibility that an organization had too many people. The company’s denial and the guests’ speculation should not be confused with a verified cause.

For O’Driscoll, the more interesting automation story is Uber’s reported experimentation with robotaxi partnerships, including a possible rollout in Madrid. He stresses that deployment has moved more slowly than many expected. Autonomous driving could affect a large workforce, but the scale and timing remain uncertain.

The guests consider two possible futures for Uber: robotaxis could threaten its position, or a range of technology providers could use its app to coordinate rides. O’Driscoll thinks partnerships may let the company participate without owning the entire technology stack, while acknowledging the outcome remains unresolved.

The panel then discusses a reported $115 billion valuation for Revolut. The guests attribute its growth in part to customer dissatisfaction with incumbent banking fees and service, especially cross-border transactions. These are their competitive interpretations, and the valuation and comparisons they cite are not independently established in the conversation.

O’Driscoll adds a caveat to his praise: in his account, Revolut’s business is not primarily long-term lending. He asks what happens if a highly valued banking brand does less of the lending that traditionally connects savings with investment. The panel treats that as a question about business models, not a finding about the entire banking system.

Stebbings brings up founders’ accounts of uncomfortable venture fundraising experiences. Lemkin recalls carrying grudges from his own time as a founder, including feeling that an investor used meetings to gather information about a competitor. He now advises founders to recognize when a meeting is useful to them as well.

Lemkin argues that fundraising is a form of selling and that founders need resilience when a prospective investor does not commit. He distinguishes an unpleasant rejection from more consequential events such as being removed from a company. His advice is to learn from the process rather than assume every failed conversation merits a lasting feud.

O’Driscoll agrees that resilience matters but argues that rejection can feel especially personal because founders are selling their own vision and identity. He remembers limited partners who rejected his funds. At the same time, he says venture investors necessarily decline most opportunities, making a disappointing answer the default experience.

The guests discuss a disputed account of a historical fundraising meeting and how investors should raise concerns about a founding team. One participant denies the account, as O’Driscoll notes. The panel cannot resolve what happened; it uses the disagreement to consider the difficulty of delivering candid feedback without damaging trust.

O’Driscoll says even experienced investors sometimes handle a rejection badly and should apologize when they recognize it. Stebbings argues that moving on can be more useful than remembering every slight. The guests’ personal stories point to different ways of processing rejection over long fundraising careers.

Stebbings cites reported milestones of $500 million in annual recurring revenue for Lovable with 146 employees and $4 billion for Cursor, with a $6 billion year-end target. These figures and target are attributed to the discussion; they are not audited results, and the target is not a realized outcome.

Lemkin says the important question is whether fast-growing AI companies can remain lean as they scale or will eventually add the organizational layers associated with earlier software businesses. He sees some teams shipping complex features with few people, and thinks that could reshape both employee opportunities and investment expectations.

He argues that a small, highly productive workforce could command higher compensation while offering fewer roles overall. That is a forecast, not a demonstrated economy-wide outcome. He also challenges leaders who request large hiring increases without explaining why the work cannot be accomplished differently.

O’Driscoll cautions against comparing revenue per employee without accounting for different costs. If an AI application spends a substantial share of revenue on outside model usage, it cannot also spend that same share on staff. His 50–70% example is illustrative, not a measured cost ratio for the named companies.

The guests discuss whether AI can give a small team enough leverage to produce and sell more. O’Driscoll thinks an employee in such a team could benefit if the company’s economics allow it to share the gains. But higher revenue per head alone does not establish profit, pay or the quality of work.

The panel disagrees over enterprise sales. O’Driscoll expects selling to large organizations to require substantial sales teams, as in earlier software cycles. Lemkin points to different go-to-market choices among AI firms and believes some founders will accept less marginal revenue in exchange for remaining lean.

Lemkin suggests that enterprise AI companies may still need more people than product-led companies, while achieving several times the efficiency of older peers. O’Driscoll agrees that the business models differ. Their multipliers are scenarios, not established measurements or guarantees for any particular company.

When Stebbings compares revenue per employee at an AI startup with Salesforce, O’Driscoll again warns that the comparison mixes enterprise sales, research and model costs. He uses rough token-spending estimates to illustrate why revenue divided by headcount does not by itself measure comparable profitability or productivity.

Lemkin predicts that, over the coming years, some startups could operate at roughly half the headcount previously needed for a given level of revenue. He considers that more consequential than any individual layoff announcement. O’Driscoll links the thesis to AI augmenting labor, though neither establishes that the predicted employment effect will occur.

Stebbings raises an alleged Cursor acquisition associated with Elon Musk and asks whether it would prove prescient if a revenue target were reached. O’Driscoll praises the broader strategy he describes, but the captions do not substantiate the transaction, terms, revenue target or the infrastructure figures cited. His assessment is conditional and speculative.

O’Driscoll argues that committing capital to AI infrastructure ahead of demand could become valuable if other model developers need capacity. He also distinguishes selling computing capacity from building a successful frontier model. The large customer-payment and annualized revenue figures he mentions are estimates in the conversation, not verified contracts or results.

Stebbings cites a reported Ramp round of $750 million at a $44 billion valuation and mentions annual recurring revenue and free cash flow. O’Driscoll frames the price as a bet on sustained growth: in his view a fast-growing financial business may command a different multiple from a mature one, but slowing growth could change that calculation.

The same question applies to Revolut, O’Driscoll suggests: investors must decide how much continued growth is implied by a private valuation. The panel’s cited revenue and profit numbers are presented as their own reported figures. They do not establish what either company will earn or what its securities are worth.

Stebbings also mentions a reported financing for AI music company Suno. Lemkin says he pays for the product and enjoys it, but uses it lightly enough that he might cancel if he were more cost-sensitive. He is uncertain whether the rapid increase in the reported valuation can be supported over time.

Asked where the money for large rounds comes from, O’Driscoll says risk appetite can matter more than a literal increase in available cash: when investors are confident they put money to work, and when frightened they retreat. He cannot say when sentiment will turn, and treats the timing as unknowable.

Lemkin notes that unusually fast startup growth can make investors behave as though all previous limits have disappeared. O’Driscoll agrees that new technology expands possibilities but argues that participants can become progressively more aggressive until someone overreaches. Both are describing a risk, not predicting the date of a downturn.

The guests discuss a recent market decline and offer different accounts of the immediate trigger. O’Driscoll’s broader point is that expectations can leave highly valued stocks vulnerable to even a modest disappointment. The episode’s cited market figures and causal explanations are the speakers’ claims, not a conclusive market analysis.

The panel turns to Bending Spoons and a reported plan to go public. Stebbings describes its acquisition-led portfolio and cites revenue and valuation figures. Lemkin admires what he sees as the turnaround of older products; O’Driscoll asks how much of the reported growth comes from acquisitions, price increases and underlying customer growth.

O’Driscoll describes a playbook of buying established products, reducing costs and raising prices for customers who remain. He says it can produce cash but makes organic unit growth difficult to assess. His examples and approximate price changes are his interpretation of the business rather than a verified breakdown of each acquisition.

The speakers distinguish a good operating business from the valuation at which it might be attractive. O’Driscoll questions whether an acquisition-led company deserves the same growth multiple as one expanding organically. Lemkin is more open to the model if management can continue finding suitable targets and executing well.

Stebbings asks whether such a listing would challenge their assumptions about how large a company must be to go public. The guests differ on whether public investors care about the source of growth, but agree that acquisition prices, continued access to targets and sustainable profits matter to the strategy’s future.

The panel then discusses a reported new Databricks private round and valuation. O’Driscoll speculates that a software company with more manageable capital needs may be able to stay private longer than a frontier-model developer. He also notes that public-market liquidity and acquisition currency can be reasons to list, but does not claim to know the company’s decision.

O’Driscoll believes public capital can eventually be cheaper, while recognizing that private investors may currently offer attractive terms to a rapidly growing company. He also suggests that a crowded calendar could influence timing. These are his hypotheses about the choice, not an announced schedule or an investment recommendation.

In the closing round, Lemkin highlights Microsoft’s reported new models and says that, in the version he had encountered, they could not search the web. He views that limitation as a reminder of how difficult it is to predict whether a new entrant can match the pace of leading model providers.

O’Driscoll sees strategic value in Microsoft developing capabilities it controls, while conceding that he had not used the models himself. He asks whether the company can reach a useful level for its customers over several years, comparing the challenge with past successes and failures in other markets. He does not claim to know the answer.

The guests also debate whether open models can remain close enough to the leading proprietary systems to create competitive pressure. They discuss models from Chinese developers and the importance of a broader ecosystem, while recognizing uncertainty over future releases, licensing and performance.

Stebbings relays a guest’s view that model consolidation could threaten infrastructure providers, and another guest’s argument that export controls have encouraged alternative engineering approaches. These are attributed opinions, not established effects of policy. The panel closes on the unresolved question of whether foundation models will be concentrated among a few providers or support a wider field of competitors.

The conversation ends with a lighter exchange about a quotation from a previous episode. O’Driscoll says a version may have appeared in an older investing book and does not claim to be its original author. Stebbings notes how strongly the line resonated with listeners.