AI Financing, Model Routing and Security Risks: A 20VC Roundtable

Harry Stebbings joins SaaStr founder Jason Lemkin and Scale Venture Partners’ Rory O’Driscoll to discuss reported model-company finances, management changes, a media acquisition, AI infrastructure and security. Company figures and transactions are the panel’s accounts or reports; allegations, legal interpretations and forecasts are not independently established in the conversation.

Stebbings cites a reported $30 billion Anthropic revenue figure and says it has surpassed OpenAI. Lemkin describes rapid growth from a lower figure at the start of the year. The interview does not establish whether the figures are actual recognized revenue, annualized run rate or comparable between companies, and it does not independently verify either company’s financials.

O’Driscoll believes capacity constraints may lead model providers to limit fixed-price use or allocate compute by price. He points to reported changes involving agent use and video, but the actual terms, capacity and business motives are not established here. His expectation about future pricing is a scenario, not a guarantee.

Lemkin cites a news report alleging Anthropic’s training costs are a fraction of OpenAI’s and infers a competitive advantage. The panel does not examine the underlying figures, accounting definitions, periods or the companies’ complete costs. Comparable revenue, profitability and any conclusion about which investors received a better deal therefore remain uncertain.

The guests discuss reported financing at OpenAI, distinguishing cash available now from conditional or staged commitments. They disagree over how to interpret the round’s strength. Their descriptions of investors, instruments, timing and valuations are not independently verified financing documents, and their hypothetical long–short trade is not investment advice.

Asked what employees should do with private shares, O’Driscoll says he generally takes liquidity windows seriously because they may be infrequent. That is his general view, not advice suited to an individual’s tax, financial or employment circumstances. The discussion does not establish that a particular tender will occur or at what price.

Stebbings lists reported OpenAI leadership changes, including health-related absences. O’Driscoll notes that outsiders do not know people’s circumstances and wishes affected individuals well. Lemkin views changes as a response to competitive pressure; the speakers’ descriptions of roles, motives and internal dynamics are not independently verified.

Lemkin says bringing an experienced executive into a broad role during a turbulent period can be difficult, citing his own investing experience rather than a measured probability. O’Driscoll asks what new capabilities would be added. Their comments are general management opinions, not findings about any individual’s performance or health.

The panel discusses a reported OpenAI acquisition of TBPN. O’Driscoll questions whether acquiring a media asset fits a stated focus on core products; Lemkin argues that some acquisitions can run with limited management attention and may provide marketing reach. The conversation does not independently establish the transaction’s terms, editorial arrangements or management’s reasons.

Stebbings asks for a fair case on both sides. Lemkin argues that acquiring a media asset can convert balance-sheet resources into marketing capacity, while O’Driscoll responds that a widely known company may not need more attention. Their debate concerns a proposed strategic rationale; it is not evidence that the buyer sought to influence coverage.

Stebbings says he understands the media asset to be editorially independent, then speculates about communications benefits. The guests do not produce the agreement or verify the actual editorial protections. O’Driscoll says owners should consider the time and financial demands of media assets, and suggests that stronger communications could address some concerns directly.

Lemkin speculates that talks may have begun earlier, when company priorities differed, and says a change in management can derail acquisitions. O’Driscoll counters that leaders can still stop a deal if circumstances change. Their claims about timing, internal approvals and what would happen today are conjecture, not a reconstruction of negotiations.

Asked to compare hypothetical inverted valuations for Anthropic and OpenAI, O’Driscoll says he might prefer OpenAI if its leadership focused on a few priorities. Lemkin says management turnover and technical leadership are risks he weighs differently. These are hypothetical, time-specific investment preferences; neither reported private valuation is independently confirmed here.

Stebbings cites reports that SpaceX confidentially filed for an IPO at a possible $2 trillion valuation, with a possible $75 billion raise and estimates of historical revenue and earnings. The interview does not verify a filing, financial statements, merger details, valuation or offering terms. A prospective or confidential filing does not establish a completed listing.

Lemkin predicts that potential offerings by SpaceX, OpenAI and Anthropic could collectively dwarf prior IPOs. O’Driscoll discusses concentration and the psychological effects of comparing other ventures with a few very large outcomes. Their aggregate comparison and assumed offerings are speculative and not calculated from a disclosed dataset.

O’Driscoll discloses a substantial QQQ holding and speculates that a future SpaceX listing might enter an index quickly. Neither the listing nor index eligibility and timing are established in the interview. He distinguishes an estimated value for existing assets from a premium investors might place on a founder’s future plans.

Lemkin expects an influential founder and retail demand could support a proposed offering price briefly, but says he does not know where shares would trade later. O’Driscoll discusses reported earlier private valuations and a possible business combination; those transactions and figures remain unverified. Their comments do not establish any market price or future return.

Stebbings reports that Doug Leone is returning to an investing role at Sequoia while other leaders retain management responsibilities. Lemkin guesses that continuity might reassure limited partners; Stebbings instead emphasizes competition for deals. O’Driscoll sees possible value from experience. These interpretations of the person’s and firm’s motives are opinions, not verified internal decisions.

The panel turns to allegations concerning compliance startup Delve and a reported removal from Y Combinator’s community. The speakers make serious claims about audits, code and conduct, but provide no independent investigation, findings, company response or verified grounds for any action. The interview therefore cannot establish guilt, fraud or the reason for a reported removal.

Lemkin speculates that the alleged handling of another company’s software may have influenced the reported decision. O’Driscoll discusses how an accelerator could protect community trust, but his numerical analogy about criminality is not evidence about Delve, its founders or Y Combinator companies. Their descriptions remain allegations and conjecture.

Stebbings describes OpenRouter as a model marketplace and cites a reported $1.3 billion valuation and $50 million in annual recurring revenue. Lemkin says he uses the product and values its ability to access or route among models. The company’s valuation, revenue, fees, market position and product capabilities are the panel’s claims, not independently verified metrics.

O’Driscoll explains the general routing idea: an application can access several models through an interface rather than building each connection separately. He cites an approximate fee as an example. Actual model availability, routing, fees and contractual terms can differ; the conversation does not establish the company’s complete pricing or accounting.

The guests suggest routing may help organizations compare premium models with cheaper alternatives for appropriate tasks. They do not claim that a cheaper model is always adequate, and their comments about model popularity and national origin reflect the data they say they have seen, without an independently examined dataset.

Lemkin worries that a low take rate requires very large transaction volume to generate substantial revenue. O’Driscoll offers a hypothetical future API market and routing share, then acknowledges that the illustrative calculation would require an unrealistically complete share of its assumed market. Neither the market forecast nor the outcome is established.

Lemkin discloses that he and Stebbings are investors in RevenueCat and says he was its first investor. He uses the company as an example of how a small percentage of a large market can still make revenue scaling difficult. His claimed market share, growth and future path are his account, not independently verified company data.

O’Driscoll compares low-take-rate businesses with payment networks and argues that large transaction volume can make small fees valuable. Lemkin replies that a theoretically large market may prove narrower in practice. The comparison illustrates different investment theses, not a prediction of OpenRouter’s or RevenueCat’s return.

The guests agree that a young infrastructure company may need to add products and deepen customer relationships while demand is strong. O’Driscoll warns that competition and cost optimization may compress some markets later; Lemkin says expanding into multiple products requires capable leadership. Their timelines and outcome scenarios are forecasts.

Stebbings asks about a reported $10 billion Supabase valuation. Lemkin likes its positioning around easier database deployment and describes integrations with app-building platforms. He says more databases may be created by agents than people, but does not supply the exact numbers or methodology. The valuation, usage and product claims are not independently verified.

O’Driscoll thinks Supabase has benefited from an existing product finding a new wave of demand. He argues that partnerships and developer adoption may matter if platforms later choose to build their own alternatives. His competitive and valuation predictions are hypothetical, not proof of lasting market leadership.

The guests compare progressively easier database deployment across generations of software. Lemkin says some users can create applications without realizing a database has been provisioned. Their examples are personal product experiences; ease of setup does not by itself establish security, reliability or suitability for every deployment.

Lemkin criticizes investors who, in his view, reassure founders without adequately confronting fast-moving competition. O’Driscoll distinguishes constructive support from avoiding difficult facts and says board members should understand competing products and recent wins and losses. They offer governance opinions based on their experiences, not evidence about any named board.

Stebbings addresses a reported Mercor security incident and says a major customer may have paused or reduced business. Lemkin discusses an unnamed executive’s strict expectations for vendor security. The interview does not independently establish the incident’s scope, affected data, actor, cause, customer decision or effect on Mercor’s business.

The guests discuss possible consequences for a vendor after a security event, including customer trust, remediation and the availability of alternatives. O’Driscoll says he lacks forensic detail and offers a conditional recovery scenario; Stebbings attributes responsibility to a particular group without presenting evidence. No attribution, ransom payment or outcome is established by the conversation.

Lemkin predicts that AI could make security incidents more frequent and says smaller companies may lack dedicated teams. O’Driscoll distinguishes security concerns in AI applications from the possible use of AI to automate attacks. Their predictions are warnings, not proof that every company will be compromised or that a specific defense will work.

O’Driscoll argues that organizations should consider security as more important services move online, and thinks the market may need different tools. Lemkin mentions configurations he has seen users alter; the panel does not test any named vendor or establish an incident from those observations. Their discussion remains a high-level risk assessment.

Stebbings introduces a report about a very small business said to have generated $1.8 billion in GLP-1-related revenue. Lemkin questions the headline and makes allegations about marketing practices; O’Driscoll adds allegations about regulatory compliance. The interview supplies no independent financial statements, regulatory findings or company response, so staffing, revenue, conduct and legality are not established here.

Lemkin separates his concerns about the reported practices from his prediction that AI may enable more personalized marketing at scale. O’Driscoll agrees that techniques can spread beyond early adopters, but the effects on customers, privacy and future advertising performance remain uncertain. The discussion is not medical advice or endorsement of the alleged marketing.

The guests compare possible changes in marketing with earlier waves of affiliate marketing and search optimization. Lemkin cites an estimated payment per customer acquisition as an illustrative claim; the panel provides no audited economics or evidence that the named company achieved it. Their conclusion about future marketing is a forecast.

The panel closes by returning to the challenge for investors: rapid growth can make new infrastructure, software and marketing models attractive, but the economics, governance and risks require careful evaluation. The reported transactions and company-specific claims discussed in the episode remain subject to independent verification.