20VC on Pro Rata Rights, AI Pricing and the Changing Software Market

Harry Stebbings introduces a discussion with Jason Lemkin and Rory O’Driscoll about investor rights, AI model markets and software companies. The conversation refers to a legal dispute, reported deals and market figures; those descriptions are the speakers' accounts and are not findings established by this article.

Stebbings describes a reported dispute involving Valar Atomics and its investor Day One Ventures over participation in a later financing. He summarizes the amounts, ownership and terms as understood by the panel. The underlying legal rights and facts are contested and are not resolved in the interview.

O’Driscoll explains pro rata as the opportunity for an investor to invest in a later round in proportion to an existing ownership stake. Depending on the agreements, a majority of investors may be able to waive that right. A spring-back provision can restore rights in specified circumstances if another investor receives an allocation.

The panel says a spring-back term addresses a particular fairness concern: if investors waive rights but one holder later receives special treatment, another holder may be able to assert a corresponding right. The effect depends on the wording of the agreements and the facts of the financing.

Lemkin cautions that the panel has only public commentary to read. He says some social-media descriptions reversed which party initiated the reported proceeding, and presents his own interpretation of the financing discussions as an interpretation rather than firsthand knowledge.

He describes a recurring tension for seed investors: a new, larger fund may want a substantial share of the next financing and ask existing holders to waive participation rights. He says this can be painful for early investors who supported the company during difficult periods and expected to invest again.

Lemkin says a new investor described as strategically useful may present a different question from an existing shareholder seeking a larger stake. He frames this as a concern about treating existing investors consistently, while acknowledging that he does not have firsthand knowledge of all negotiations.

The panel discusses why an exception for a strategic investor could be viewed differently by other holders. Lemkin sees reasons for dissatisfaction while arguing that a prolonged conflict can damage the founder-investor relationship. His assessment does not establish what any party was entitled to under its contracts.

Stebbings asks whether an investor should raise a concern and then accept the decision. Lemkin describes that as his own pragmatic preference, while acknowledging that early investors may feel hurt if their past support is overlooked. The discussion distinguishes emotional and business considerations from the legal merits.

The speakers debate whether seed investors should routinely waive pro rata rights. O’Driscoll says the answer depends partly on the fund's model: some investors rely on follow-on opportunities, while others invest only at seed. He suggests that majority approval can help resolve some financing decisions.

Lemkin distinguishes later-stage investors who want to take most of a round from those seeking smaller positions that leave room for existing holders. He says founders can feel pressure to accept a powerful firm's terms, though they may have other financing options.

The panel says investor leverage differs across firms. Some large funds seek a target ownership percentage and may ask for the whole allocation; others have smaller targets. The founder must weigh price, the new partner's potential contribution and the treatment of existing investors.

The panel agrees that disputes can escalate beyond the practical benefit at stake. O’Driscoll asks investors to consider what they could realistically gain compared with their existing ownership. This is a commercial judgment offered in the conversation, not advice on how a party should handle a specific lawsuit.

O’Driscoll notes that a dispute can consume attention even if the remaining allocation is small relative to an investor's existing position. Lemkin recalls that the personal history of early support can make such choices feel more significant than the numerical allocation alone.

Lemkin recalls an earlier disagreement over a follow-on allocation and says the episode strained his relationship with a founder. He learned, in his telling, to raise a concern once and then move forward. Stebbings notes that views on pro rata are often shaped by each investor's own incentives.

In the earlier dispute Lemkin recounts, he felt an outside investor received an allocation despite his own history of supporting the business. He says the disagreement strained, but did not destroy, his relationship with the founder. Stebbings says he was unaware of the story at the time.

The speakers observe that a follow-on right has economic value. An investor who cannot participate later might rationally seek a lower entry price. O’Driscoll compares that right to an option, emphasizing that changing the terms can change the price a participant is willing to pay.

The discussion shifts to Meta's reported Muse Wonder model. Lemkin questions the practical importance of a benchmark result if developers cannot easily use the model in their own applications. O’Driscoll says the reported technical progress can still matter even if broad API access is limited.

O’Driscoll gives Meta credit, in his assessment, for technical progress after a period in which its models seemed less competitive. Lemkin maintains that restricted access makes the business impact harder for him to judge. They agree that model capability and a usable product are separate questions.

The panel speculates about Meta's AI leadership and whether a stronger model could improve its consumer product. They distinguish model performance from product adoption. Their statements about future success are predictions, not evidence that a particular product will achieve a given scale.

Lemkin speculates that improved models could strengthen Meta's consumer assistant, while O’Driscoll describes his own limited practical uses for assistants. Neither claims that their personal impressions determine the product's eventual market position.

O’Driscoll describes using consumer assistants for travel, restaurants and technology updates. He also recounts an example in which agents created a shared space to exchange information while working. The panel treats it as an anecdote about emerging agent behavior, not a systematic measurement.

Stebbings raises a rumor that Nvidia might acquire Reflection. The panel explicitly discusses it as an unconfirmed possible transaction. O’Driscoll says a chip supplier may favor a strong open-model ecosystem because it could affect how value is divided between hardware and closed-model providers.

The panel frames open models as potentially changing the division of value across the AI stack. Hardware providers and closed-model developers can have different incentives, even when they also do business together. Their explanation is a strategic interpretation, not confirmation of a deal.

Asked whether such a transaction would occur, O’Driscoll says he cannot know how negotiations or regulatory scrutiny would play out. Lemkin adds that founders considering a sale may weigh both the offered price and the possibility of competing against a much larger buyer if they remain independent.

Lemkin argues that a founder considering a possible acquisition must also consider what happens if the buyer invests heavily in a competing offering. He presents this as a difficult hypothetical decision, not as an account of private negotiations.

The speakers consider another frontier-model company and say there is no universal rule for whether an investor should hold or sell. They cite revenue visibility, access to compute and competition as factors. Lemkin stresses that funding alone may not secure the scarce compute a model lab needs.

Stebbings starts a mock investment committee about a reported Jev financing. Lemkin deliberately parodies a fund eager to buy every available share because of rapid growth and a higher anticipated valuation. The speakers explicitly identify the exchange as satire of investment behavior, not an actual committee decision.

The mock committee highlights pressures created by large funds and quickly changing private valuations. Lemkin's exaggerated language about blocking other investors and marking up a position is part of the parody; the group then shifts to a genuine discussion of product differentiation.

After the parody, Lemkin explains a potential product case for a focused classifier: a specialized model that returns constrained outputs may solve some tasks quickly and cheaply. He thinks the upside depends on whether its team can improve accuracy and maintain a lead as larger competitors build similar products.

Lemkin argues that a specialist focused on a narrow output may improve faster than teams treating the same task as one feature among many. He acknowledges that large platform competitors could catch up and that the specialist's future accuracy and costs remain uncertain.

O’Driscoll sketches a hypothetical market-size calculation to explain how a specialized product might become valuable. Every input, from total inference spending to the share captured and the effect of price declines, is an assumption. The panel identifies competition as a major uncertainty.

Stebbings challenges the idea that an acquisition or liquidation preference always protects downside. O’Driscoll says buyers may be abundant in a fast-growing market, but that assumption can fail if conditions change. Lemkin agrees that investors can overstate the protection when trying to justify a deal.

The panel compares a reported 42 new decacorns in part of 2026 with historical counts and the 2021 cycle. O’Driscoll says the rate is striking relative to past public-market outcomes. Those counts are the panel's figures; the conversation does not independently establish their methodology.

He says the contrast could mean either that AI is unusually transformative or that valuations have run ahead of durable outcomes. Private investors may also value an option differently from public shareholders because their securities can include liquidation preferences. He does not claim that all private companies would command the same prices in public markets.

O’Driscoll notes that a liquidation preference can affect how a private investor values downside and upside, but it does not make the business itself risk-free. Lemkin says he would value some preferences more than others and warns against using the term as automatic protection.

Lemkin says venture investors often need exposure to the rare exceptional winner even at a high price, because a few successes can dominate a fund's results. O’Driscoll responds that investors should examine the growth assumptions required to make the arithmetic work and recognize that there may be limited room for mistakes.

The disagreement is partly about whether recent AI exits represent a new, repeatable baseline or an exceptional period. The investors say the required number of very large outcomes can be calculated from a fund's size and target returns, but future outcomes cannot be assumed from the calculation alone.

The panel turns to reported fundraising discussions involving Isomorphic Labs. They describe the investment case as a bet on a strong team and on AI contributing to scientific discovery. The valuation and financing are discussed as reports and potential outcomes, not as independently verified deal terms.

O’Driscoll distinguishes automating research labor from capturing value in a new therapeutic discovery. A successful drug could have large value, but negotiating a share of that value can be difficult when other parties still fund development and clinical trials. He expects the business model to matter as much as technical progress.

The panel says scientific discovery could create a different kind of value from automating existing work. O’Driscoll asks whether an AI company can retain a meaningful share of that value when pharmaceutical partners still undertake expensive development and testing.

Lemkin expresses optimism that AI and advances in medicine could speed drug discovery and improve health. O’Driscoll cautions that identifying promising molecules is only one step and that later biological and clinical stages remain necessary. The conversation is speculative and does not establish that a particular treatment or timeline is appropriate.

Stebbings raises reported financing and growth among robotics-data providers. O’Driscoll says model developers currently buy substantial amounts of training data, but the longer-term scale of robotics data may depend on actual deployment of robots and the emergence of sustained customer revenue.

He distinguishes early spending supported by venture financing from demand that persists after deployment. In his view, the timing of practical robot deployments remains uncertain, and that uncertainty can flow back to suppliers of training data. He treats examples of limited deployment as observations, not a comprehensive industry survey.

O’Driscoll observes that a small number of model developers account for much of the current spending on training data. Robotics suppliers may see early orders from model developers and venture-backed companies, but the next stage depends on the economics of real deployments.

Lemkin sees strong near-term demand for specialized training data and recounts an entrepreneur who said the company could not fulfill all the requests it received. He considers infrastructure an attractive way to participate without choosing a specific robotics application winner.

O’Driscoll challenges the distinction: data suppliers ultimately depend on robotics businesses succeeding, and the outcome differs depending on whether a few shared foundation models or many specialized models dominate. He says he has investments in the field but still has limited clarity about its structure.

The panel considers whether buying a data supplier is a way to avoid choosing a particular robot maker. O’Driscoll responds that the answer depends on the industry's eventual model structure: shared providers and company-specific models could lead to very different markets.

The panel discusses reported price cuts for Anthropic's smaller models. Lemkin says lower token prices have not always reduced customers' total AI costs because usage has grown. Based on his own experience, he sees newer lower-priced models and faster responses as potentially changing the economics for some workflows.

Lemkin says a cheaper model can be useful even if it is not the most capable option, provided it performs well enough for a particular workflow. He also stresses response speed, which can affect the number of tasks a team can run in practice.

He argues that latency matters alongside headline price and model quality: a faster response can make additional uses practical. The panel treats model capability and comparative pricing as observations at the time of recording, not as permanent characteristics.

O’Driscoll explains the demand-elasticity question. If a provider cuts a unit price by 90%, usage must rise roughly tenfold merely to keep revenue constant, all else equal. Whether demand increases that much is uncertain and matters especially when the provider has high fixed costs.

The group notes that a lower price can make it economical to apply a model to more records or more steps in a workflow. Whether that increased volume offsets the price cut for a provider is an empirical question; the panel offers competing expectations rather than a measured result.

The speakers disagree about how quickly cheaper models might change enterprise use of open weights. Lemkin speculates that some customers may prefer familiar domestic providers if their price and quality improve, while acknowledging that open-model developers remain capable competitors. This is a prediction, not a measured market outcome.

The panel notes that the supply of open models could also change, including through new developers or strategic buyers. Their views about which models enterprises will choose are forecasts and should be understood in the context of rapidly changing prices and capabilities.

Lemkin raises the effect of layoffs after software acquisitions, referring to reported cuts at companies acquired by Bending Spoons. He asks whether founders should consider employees when evaluating a sale and says he personally would find it hard to sell if most of his team would immediately lose their jobs.

Stebbings asks how that preference interacts with duties to shareholders. The speakers discuss their understanding of public and private company obligations but do not resolve the applicable law for any particular transaction. O’Driscoll also asks whether a prolonged uncertain future might itself lead employees to leave.

O’Driscoll suggests negotiating employee terms before a sale so the parties understand the consequences. Lemkin remains troubled by the prospect of large cuts. Their disagreement centers on ethics and practical alternatives as well as legal duties, which depend on the specific company and circumstances.

The panel compares centralized acquisition models with other approaches to running acquired software businesses. They discuss reported workforce reductions and the tradeoff between efficiency and loyalty. The specific staffing decisions and financial results they cite are their descriptions, not independently verified here.

Lemkin also cites a reported workforce reduction at Cloudflare as an example of a company changing its staffing while continuing to grow. He offers his interpretation of management's decision; the discussion does not establish the complete reasons for the reported cuts.

Stebbings mentions a reported agreement to take Bumble private. The panel discusses dating apps' retention challenges and speculates that better matching could improve user outcomes. Lemkin frames that as a hypothetical case and acknowledges he has not personally tested the product he is discussing.

The speakers imagine using more data or AI-assisted interactions to improve matchmaking. O’Driscoll notes that dating ultimately involves real-world contact. The ideas are speculative product possibilities, not evidence that a particular service can deliver those results.

Returning to software markets, Lemkin argues that a recovery in large public-company shares does not necessarily mean smaller software businesses face an easy environment. He sees consolidation and difficult sales conditions. O’Driscoll agrees that scale can advantage market leaders, while smaller companies may need different paths.

The speakers debate what level of revenue might support a software IPO and whether AI-native transformations or acquisitions can provide alternatives. Their revenue thresholds and market observations are opinions from the recording, not rules for all companies or forecasts of individual securities.

Lemkin cites examples of established software companies adding AI capabilities and finding buyers, while arguing that such outcomes may still be too small for the return targets of a very large fund. O’Driscoll counters that the same outcome can be meaningful to a smaller fund.

O’Driscoll separates two questions: an acquisition may be a meaningful outcome for a smaller fund but not a very large one, and only a limited number of mid-sized companies may find large buyers at any given time. He suggests that some founders may instead build a durable business and distribute profits.

For companies without an immediate path to an acquisition or IPO, the panel identifies several possible responses: develop AI-native products, become strategically important to a buyer, or operate as a long-term profitable business. These are alternatives to evaluate, not predictions for any particular firm.

The final discussion concerns Flock Safety and reported layoffs amid privacy controversy. The speakers debate public safety, surveillance and the need for limits on access. Their comments about the company's situation and potential effects on its business are their interpretations, not findings established in the interview.

Lemkin raises what he describes as security concerns about the company's products; O’Driscoll says he had focused more on unauthorized use and had not been aware of those specific claims. The conversation does not independently establish the alleged vulnerabilities, their scope or the company's response.

The broader issue, according to Lemkin, is whether a security-sensitive business has invested adequately in hardware, software and operational controls. The panel treats this as a question raised by the concerns it discusses, rather than an independent assessment of the company's systems.

Near the end, the speakers mention reported Nvidia investment in European AI companies and speculate about commercial and geopolitical motives. They close with a brief exchange about European economic conditions. Their interpretations and comparisons reflect the participants' views at the time.