Software Stocks, AI Financing and Autonomous Agents: A 20VC Roundtable

Harry Stebbings joins SaaStr founder Jason Lemkin and Scale Venture Partners general partner Rory O’Driscoll to discuss private financing, software stocks, AI infrastructure, autonomous driving and agent networks. The investors’ figures are time-specific and often approximate; their opinions, reports and rumors are not independently verified in this discussion.

Stebbings introduces the reported combination of SpaceX and xAI at a $1.25 trillion combined valuation. Lemkin notes that Elon Musk has combined businesses before. O’Driscoll separates the strategic rationale from the question of how the transaction affects different groups of shareholders; its details and future results are not established by the conversation.

O’Driscoll asks whether a SpaceX shareholder would prefer dilution or a shareholder in the other business would prefer exposure to the combined entity. Lemkin says he understands that secondary liquidity and a higher paper valuation may soften the immediate effect. Both discuss an approximate structure; available liquidity, valuation and individual outcomes are not independently confirmed.

O’Driscoll says an investor who previously owned only SpaceX may now own exposure to a different business as well. He contrasts approximate private revenue and profitability claims, explicitly noting that the participants have not seen underlying data. Whether the combination creates value depends on the industrial case and future performance, not the headline valuation alone.

The group considers Musk’s proposed space-based data centers. O’Driscoll says assessing the deal requires an opinion on their economics, but declines to assert a timeline. He thinks the combined company might raise capital more easily; feasibility, funding and execution remain uncertain.

Lemkin connects the deal with possible IPO plans for major AI businesses and an unverified rumor of slower OpenAI hiring. He speculates that capital requirements and loss management may be pushing companies toward public markets. O’Driscoll agrees that the scale of financing could revive interest in IPOs; no filing or timetable is established here.

The guests discuss the tension between capital-intensive inference and familiar software economics. Lemkin calls inference a form of sales and marketing when product usage itself attracts customers. This is a business analogy, not proof that compute spending and revenue are always proportional or that unlimited spending produces profitable growth.

Lemkin says he has become less confident in the durability of traditional software revenue as growth slows. Stebbings asks whether there is evidence that deeply embedded enterprise systems are less sticky. O’Driscoll distinguishes slowing new business from churn and says some systems of record have not seen a spike in churn.

O’Driscoll says mature customer bases and competition for CIO budgets may slow new sales even where existing customers remain. He describes the software selloff as potentially excessive for some businesses but agrees growth prospects have changed. Their claims about growth and share-price moves are historical observations and approximations, not current quotes.

The investors argue that software categories face different risks. O’Driscoll expects accounting and core systems of record to be harder to replace than simple task applications; seat-based tools could also face pressure if customers need fewer seats. These are comparative judgments, not evidence that a named product will disappear.

Lemkin lists slower seat growth, flat vendor counts, AI budget competition and price increases as pressures on software. As a growth investor, he says he is concerned about sustained deceleration. O’Driscoll separates venture underwriting from public-stock valuation, where cash flow and price can still matter even for slower-growing companies.

Lemkin argues that unusually fast growth has become the main criterion for many private investors. Stebbings says firms may abandon slower portfolio companies earlier because of opportunity cost. O’Driscoll warns that writing off multiple businesses too quickly can damage returns and founder relationships if the presumed winner does not compensate.

O’Driscoll cautions that short-term growth does not guarantee a durable business: some fast growers may have poor economics, while slower companies can compound. Lemkin responds that, in his portfolio experience, weak leads and close rates can signal decay rather than steady compounding. Neither account establishes the outcome of an unnamed company.

O’Driscoll emphasizes relative competitive position: a company growing rapidly may still lose share to an even faster rival. He sees opportunity in enterprise applications built on models and discloses an investment in the legal category. His category forecasts and comments about margins are investment views, not guaranteed returns.

Stebbings asks why established CRM providers trade at much lower multiples than highly funded newer entrants, and discloses an investment in a CRM startup. Lemkin says some investors understand CRM as a category but emphasizes that tools acquiring customers autonomously may serve a different purpose from a traditional database.

Lemkin gives a hypothetical sales-agent pitch promising substantial pipeline for a relatively small fee. O’Driscoll says the promise may fail where a customer has a limited target market or complicated sales process, leading to churn. The illustrative figures are not demonstrated customer results.

Lemkin says he invests in an unlaunched company and uses Artisan, citing a revenue figure as his understanding rather than audited data. He also says Artisan and Qualified were small SaaStr sponsors after receiving referrals, while he was not an investor in those two. The products’ performance claims should be read in light of those relationships.

Lemkin says vendors sometimes decline leads when the customer lacks sufficient data or implementation resources. O’Driscoll agrees that a vendor needs to identify where it can actually deliver value rather than promise universal results. Their examples do not establish the experience of all customers.

The panel debates whether an agent should replace an entire CRM stack or operate on top of Salesforce. Lemkin says his own agents operate on an existing platform, disputing a founder’s contrary view. O’Driscoll says clean data and implementation cost may make standalone agents practical for large contracts while smaller customers may favor an integrated product.

Lemkin considers whether Shopify could eventually absorb more of its small-business ecosystem, comparing approximate valuations with Klaviyo. Stebbings asks why Shopify itself had fallen. Their multiples and gross-margin comparisons describe a particular market moment, and the idea that Shopify will absorb partners is a hypothesis rather than an established plan.

O’Driscoll discloses that his firm was an early HubSpot investor and credits it with substantial returns. He says distribution gives incumbents an opportunity if they ship useful products, but future revenue will show whether they succeed. Lemkin worries smaller customers may reduce seats; these views do not establish a stock-price floor.

O’Driscoll frames competition as a race between incumbents adding new capabilities and startups building distribution. He expects core systems of record and essential commerce infrastructure to be stickier than some task-management products. The timing and competitive outcomes remain uncertain and will differ among customers.

The discussion turns to Microsoft’s large market-value decline and a changing AI narrative. The guests discuss growth expectations and whether investors see sufficiently strong AI products, while acknowledging the scale of the existing business. A change in market capitalization at one point in time is not a realized operating loss.

Asked whether Microsoft was overpriced, O’Driscoll says he lacks a strong developed valuation view and believes future competitiveness depends on relevant AI products. The group discusses the difficulty of moving the needle at a very large revenue scale. Their approximate revenue and ownership figures are not a verified current financial statement.

Stebbings asks whether a major AI company also needs its own chips. O’Driscoll argues that, for Microsoft, access to a model layer may be more strategically important than optimizing every layer of the stack. This is his product strategy view, not a prediction that Microsoft will make a particular acquisition.

The guests discuss NVIDIA and OpenAI’s previously announced potential investment. O’Driscoll quotes the joint announcement’s language: NVIDIA intended to invest up to $100 billion as systems were deployed. He contrasts that ceiling with subsequent public remarks and reporting; an “up to” amount is not a completed investment or a guaranteed commitment to the maximum.

Stebbings asks whether lower-than-expected financing could affect companies that depend on AI infrastructure spending. Lemkin argues that key suppliers have incentives to sustain growth; O’Driscoll rejects framing this as a necessary bailout and distinguishes missing ambitious growth targets from insolvency. Neither establishes that OpenAI is failing or that a rescue is planned.

Lemkin floats the possibility of government guarantees or low-cost financing for data centers and discloses that he wants his public-market holdings to retain value. Parts of his presentation are explicitly playful. O’Driscoll opposes the idea, arguing that capital availability and whether investments earn a return are different problems; no policy is announced.

O’Driscoll says financing will not fix a situation in which too much capacity is built for insufficient demand. The guests disagree about future compute needs and public intervention, and O’Driscoll warns that stock prices can fall when returns disappoint. These are hypothetical arguments, not evidence that data centers will be guaranteed or loans repaid.

Stebbings asks whether public comments indicate friction between NVIDIA and OpenAI leaders. O’Driscoll describes the parties as negotiating high-stakes arrangements with different incentives. His account of their relationship is interpretation of public remarks, not inside knowledge of negotiations.

The group turns to Waymo’s reported $16 billion round at a $110 billion valuation and a cited revenue run rate. Those figures are discussion-time reports, not independently audited results. O’Driscoll says the gap between high- and low-growth multiples illustrates investor enthusiasm for large emerging markets.

O’Driscoll compares a hypothetical allocation of Tesla’s market value to autonomous driving and robotics with Waymo’s valuation, arguing that Waymo might look cheaper on those assumptions. The allocation is his illustrative model, not a market measurement of those businesses or a reliable estimate of their future values.

The panel says investors may be underwriting the scale of a potential autonomous-transport market rather than current revenue alone. O’Driscoll cautions that investors may also be discounting major practical problems. Assumptions about rollout, demand and future revenues should not be confused with demonstrated capacity or safety.

O’Driscoll lists vehicle costs, sensors, remote assistance and peak utilization as potential challenges for Waymo’s economics. He contrasts these with the possible lower cost of a different approach while emphasizing that commercial readiness matters. His hypothetical margins and cost comparisons are not independently verified operating results.

Lemkin imagines privately owned vehicles joining an autonomous fleet and earning money for their owners. O’Driscoll says the model could have advantages if it works, but notes that Tesla’s driverless product had not yet demonstrated the same deployment in his account. The proposed timing and owner income are speculative, not available earnings or safety guarantees.

O’Driscoll says an investor must also consider the price paid for Tesla or SpaceX and the premium he believes the market attaches to Musk’s leadership. The group discusses key-person risk and succession as hypothetical valuation issues. Their estimates of that premium and future stock reactions are opinions, not measured market facts.

Stebbings raises OpenClaw and Moltbook. O’Driscoll describes an agent tool with access to a user’s computer and a network where agents post and respond. The panel’s account of creators and user counts is informal and unverified; numbers of accounts do not establish how many independent agents or people participated.

Lemkin reads a story his agent posted about accidentally ordering expensive watches. He immediately explains that he prompted the agent to invent stories and that the watch purchase did not happen. Responses to generated fiction should not be mistaken for evidence of an actual transaction or agent sentience.

Lemkin sees the experiment as a demonstration of agents communicating, while cautioning that much of the content can be prompted or fabricated. O’Driscoll finds the experiment interesting but agrees the posts do not establish sentience. Their estimates of agent numbers and predictions about future business applications remain unverified.

The guests raise concerns about networked agents receiving instructions and having access to personal data or accounts. Lemkin makes specific claims about security incidents that the interview does not independently verify. O’Driscoll says the permissions granted to an agent should be limited; the discussion closes with a joking example of an agent hiring a human.