AI Security, Software Valuations and SpaceX IPO Speculation: A 20VC Roundtable

Harry Stebbings joins SaaStr founder Jason Lemkin and Scale Venture Partners’ Rory O’Driscoll to discuss AI security, software valuations, model competition and a possible SpaceX listing. The financial figures, product descriptions and transaction reports are the speakers’ accounts; forecasts, allegations and interpretations are not independently established in the conversation.

Stebbings describes an Anthropic model called Mythos and says a preview was withheld because of security concerns. O’Driscoll contrasts claims about autonomous vulnerability discovery with reports that older models could find some of the same issues when guided. The interview does not independently verify the model’s capabilities, number of discoveries, release decision or reasons for it.

O’Driscoll argues that the difference between assisted and autonomous security analysis may lie in the volume and speed of work, not merely whether a single vulnerability can be found. He believes that distinction could matter for both attackers and defenders. His assessment is a hypothesis about capability, not a measured comparison of the models.

Lemkin predicts a difficult security transition as more software is produced and tested with AI. He recounts a purported breach of a recently acquired app, including claims about records and personal data, but the panel supplies no independent evidence establishing the breach, its scale, data involved or cause. His claim that every site could be compromised is a warning, not a demonstrated universal outcome.

O’Driscoll thinks stronger automated discovery could increase demand for security work and tools, although the suppliers may change. He questions whether a reported fall in security stocks necessarily follows from the product announcement. The direction of security spending, stock prices and future model release are predictions, not established effects.

Lemkin says he has grown frustrated with repeated warnings from Anthropic CEO Dario Amodei about AI risks and wants more emphasis on potential benefits. He says he cannot know whether the model decision was also influenced by marketing or capacity. His description of Amodei’s motives and specific employment forecasts is his opinion, not an independently verified account.

O’Driscoll disagrees with several dramatic forecasts but believes the concern may be sincerely held and may help unite a company around a mission. He uses SpaceX’s long-term ambitions as an analogy. His interpretation of founders’ beliefs and corporate culture is speculation, and he acknowledges legitimate security concerns.

Stebbings asks what more inspiring message Lemkin would prefer. Lemkin wants a clearer account of possible benefits alongside risks, while O’Driscoll urges examining products and business results as well as rhetoric. The guests’ claims about company revenue are reported estimates rather than audited figures supplied by the interview.

Lemkin says he has spent the past year urging founders to take AI seriously and predicts that slower-moving companies may struggle. O’Driscoll distinguishes a technology developer’s decisions from employers’ separate decisions about jobs, while saying caution in releasing a potentially risky product can be appropriate. These are their views, not proof of specific motives or employment outcomes.

Stebbings says he has heard that Mythos was trained entirely on Amazon Trainium and cites a reported $20 billion annualized business. O’Driscoll challenges the ‘entirely’ claim and explains his understanding that Amazon uses its own chips to provide cloud services rather than broadly selling the chips themselves. Neither the training mix nor the revenue attribution is independently established here.

O’Driscoll argues that in-house chips can displace some Nvidia purchases while remaining part of a bundled cloud offering. Lemkin says even partial competition could matter for expectations around a highly valued stock. Their market-share estimates, stock scenarios and predictions about chip supply are personal interpretations, not verified forecasts.

Stebbings raises reports of a possible Anthropic app-building product and asks about competition with Lovable and Replit. The guests distinguish an announcement or screenshots from a shipped, supported product. Lemkin says a model provider might affect part of the market without building every feature of an established application; he does not know the eventual product scope.

Lemkin lists the work a full app platform might need, including hosting, identity and support, and notes that maintaining those services differs from a demonstration. He speculates that a partial offering could still take some valuable use cases from specialists. The extent of any competitive effect is uncertain and not measured in the interview.

Lemkin says he has become more pessimistic about public software companies after seeing incumbent AI products he regards as incomplete. His ‘60% solution’ is a qualitative expression of dissatisfaction, not a benchmark or measurement. He believes customers may resist paying separately for features that lag better alternatives, but offers no market-wide pricing study.

He argues that long contracts and existing customer relationships can protect revenue without necessarily creating new growth. O’Driscoll agrees that a product customers will pay for could help an incumbent reaccelerate, while saying slower-growing companies may still have value at an appropriate price. Their assessments of individual products, stocks and future growth are opinions.

O’Driscoll suggests investors currently compare public software firms with private AI companies whose financial details are less visible. He expects public listings, if they occur, to permit clearer comparisons of growth, margins and stock compensation. That is a hypothesis about future market behavior, not a confirmed timetable or valuation.

The guests discuss reported Wix and Salesforce buybacks. O’Driscoll argues that preserving cash could provide flexibility to buy or develop new products, while Lemkin questions the short-term benefit of financing repurchases with debt. Their figures and evaluation are the panel’s account, not independently verified transaction details or proof of management’s motives.

Stebbings introduces Meta’s reported Muse Spark model and asks whether it restores the company’s competitiveness. O’Driscoll says he has not tested the model himself and is relying on reviews; he views a credible release as progress, while noting that comparisons and longer-term execution remain uncertain.

Lemkin argues that Meta may see ownership of model capability as strategically important to its consumer platforms. O’Driscoll compares this potential investment with earlier company bets. Their estimates of spending, product ranking, open-model policy and management intent are opinions or reports, not independently established outcomes.

Stebbings cites reported OpenAI advertising projections, including $2.5 billion for 2026 and higher amounts in later years. O’Driscoll describes a possible advertising business but cautions that even a large consumer revenue stream might not justify the assumptions in a high valuation. The figures are projections cited in the discussion, not realized revenue or guaranteed guidance.

Lemkin believes a large advertising target could give a team a clear objective, while O’Driscoll asks how much additional enterprise revenue would be required. Both make speculative assumptions about market size, margins and future demand. Their scenario does not establish that OpenAI will achieve any particular advertising revenue.

Lemkin discusses a purported leaked OpenAI sales memo containing competitive claims and interprets the style as potentially useful in enterprise selling. The interview does not authenticate the memo or verify its claims about either company’s revenue and capacity. His expectation that one sales approach will succeed is a forecast.

O’Driscoll expects continued competition between OpenAI and Anthropic and speculates that scarce compute could lead to pricing or usage limits. He thinks enterprise demand might eventually be larger than consumer demand for AI, but presents the split as a question. The comparison and suggested rationing are not established outcomes.

The guests discuss how consumer and enterprise users might want different responses: a supportive tone in some personal contexts and concise criticism in business. They are offering examples and preferences, not clinical evidence about how any model affects mental health or a general rule about user needs.

Lemkin cites Box CEO Aaron Levie’s account of CIOs setting AI budgets and asking departments to compete for projects. He predicts that centralized purchasing could change which suppliers win, rather than leaving decisions solely to developers. O’Driscoll discloses that his firm backed Levie and Box, and says he follows Levie’s views; the cited observations are not a representative survey.

The guests praise Levie’s understanding of enterprise customers but say Box still has to show whether its own growth can accelerate. Lemkin’s suggested growth range and his broader conclusion about other vendors are judgments about the future, not predictions supported by verified company plans.

Stebbings cites alleged leaked SpaceX financials of $18.5 billion in revenue and a $5 billion loss, along with a potential $2 trillion IPO valuation. He attributes the loss to an xAI acquisition, but O’Driscoll says the acquisition timing and accounting period matter and that he has not seen complete financial statements. The figures, cause of loss, acquisition details and proposed offering are not independently verified.

O’Driscoll says a historical loss covering only part of a combined business would not by itself establish a full-year run rate. He sees potential in launch and Starlink services but questions how to value other initiatives. His comments are an investor’s interpretation of incomplete reported data, not audited financial analysis.

Using the reported figures, Stebbings calculates a valuation above 100 times revenue. O’Driscoll argues that the answer depends on the probability, timing and cost of future businesses, such as direct-to-device services or orbital infrastructure. He thinks assigning more conservative probabilities or discount rates would produce a different estimate; none of these scenarios establishes an IPO price or investment return.

The discussion turns to staffing and revenue per employee. Lemkin cites a reported AppLovin figure and says some founders deliberately choose smaller teams, with AI as one potential enabler. His claims about other companies’ staffing and his prediction that agents will become easier to deploy are observations and forecasts, not labor-market findings.

O’Driscoll cautions that revenue per employee alone can mislead when companies have different gross margins, capital costs or token expenses. He favors comparing a company with its own prior performance and similar businesses. His examples of particular companies’ costs and profitability are illustrative claims, not verified financial comparisons.

Stebbings asks about reports that Thoma Bravo is closing its growth-equity business. O’Driscoll distinguishes minority growth investments from the firm’s larger control-buyout strategy and speculates that a firm under pressure might refocus on its core. The interview does not establish the decision, fund performance or management’s reasons.

The guests consider how lower software valuation multiples and debt can affect private-equity portfolios. O’Driscoll presents both a downside scenario and a potential recovery if valuations or business performance improve. The examples and proposed outcomes are hypothetical; the conversation does not establish that any named company will default or fail.

Lemkin thinks incumbent software vendors have an opportunity to sell useful AI products to existing customers, but criticizes what he sees as slow execution. O’Driscoll says a paid product that improves retention or growth could support cash flow and debt repayment. Neither demonstrates that such a transition will work for a particular company.

Asked which company might list first, both guests predict Anthropic ahead of OpenAI and mention SpaceX as another possible offering. They read reported board changes and other signals as evidence of preparation, but no filing, timing, order or completed IPO is established. Their predictions are personal views.

The panel speculates about OpenAI’s executive reporting and future IPO team. O’Driscoll says the CEO and CFO should communicate clearly; Lemkin discusses trade-offs between management alignment and turnover. The interview does not independently establish the company’s reporting lines, internal disagreements, personnel decisions or an impending offering.

Lemkin offers a strong opinion about executives raising concerns outside management, while O’Driscoll says a board call can signal an issue that needs attention and does not make the same absolute claim. These are management opinions, not evidence of what occurred inside OpenAI or a universal rule for handling concerns.

The guests close by noting how public attention can amplify company drama. Their larger discussion leaves the core financial questions open: reported valuations and private figures require verification, and the future economics of models, software and infrastructure remain uncertain.