AI Compute, Private Equity and the Narrowing Exit: A 20VC Roundtable

Harry Stebbings joins SaaStr founder Jason Lemkin and Scale Venture Partners’ Rory O’Driscoll to discuss reported AI company results and financing, competition among models, a disputed cross-border transaction and private-equity exits. The financial figures and transaction descriptions are the panel’s accounts or reports; allegations, forecasts and investment views are not independently established by the conversation.

Stebbings cites reports that OpenAI missed user-growth and revenue targets and says CoreWeave and Oracle shares fell. O’Driscoll interprets the news as reflecting earlier model performance relative to Anthropic, while saying newer coding models may have improved. The figures, stock movements and causal explanation are not independently verified in the discussion.

Lemkin agrees that the reported figures look backward and says he sees a new question: as agents take on workflows, they may help select models and tools. He describes his own agents as working well with OpenAI and predicts a more competitive market. His experience is anecdotal, and his forecast about how most workflows will be managed is not established fact.

O’Driscoll asks whether agents choosing tools would reduce human brand loyalty and whether the major model providers would still dominate. Lemkin describes an experiment in which models graded APIs, and says the results favored prominent, innovative providers. He presents this as his own test; it does not prove objective product quality or predict how every agent will choose.

Lemkin reports that his model-based grader favored Stripe and ranked several established marketing tools poorly for agent use. The grades reflect the prompts, models and criteria in his experiment, not an independent evaluation of the products. He predicts that vendors will need to serve agents as well as human users.

The guests consider whether the company controlling an agent harness could favor its own model and create lock-in. Lemkin thinks ownership of the agent layer could matter as much as the underlying model, while allowing that agents might eventually select independent alternatives. Both the degree of lock-in and future market structure remain uncertain.

Stebbings asks whether long enterprise contracts protect incumbents. Lemkin argues they can defer customer churn rather than eliminate it; O’Driscoll says systems of record may persist even if growth slows. Their examples and statements about contract terms are views on potential risk, not proof that named companies will lose their customers.

O’Driscoll interprets analysts’ questions about ServiceNow’s agent revenue as evidence that investors want to see how an incumbent benefits from agent activity, not just legacy subscriptions. His account of the call and the financial measures is interpretive. The discussion does not determine how much future growth any platform will achieve.

Lemkin praises Canva’s newer AI features but questions whether autonomous agents will need the same interface as individual users. He nevertheless predicts a successful possible IPO and says the timing of any change in demand is unclear. Stebbings notes his friendship with Canva’s Cliff Obrecht; there is no confirmed offering date or guaranteed outcome here.

Lemkin suggests businesses may automate multi-person workflows sooner than individual consumers replace their own tools. He admits he does not know when the crossover would occur. O’Driscoll considers a spectrum from shrinking legacy products, through persistent systems of record, to platforms that gain agent-driven activity; it is a framework rather than a classification proven for individual companies.

The guests discuss Canva and Rippling as substantial businesses that might be priced on growth and cash flow rather than the highest speculative software multiples. O’Driscoll argues venture returns can be especially large when markets attach a narrative premium, but his proposed valuation outcomes and IPO expectations are predictions, not known transaction results.

Stebbings cites reports of Google committing up to $40 billion to Anthropic, including an immediate cash component and milestone-dependent amounts, and a further $5 billion from Amazon. He asks whether in-house chips could affect Nvidia. The panel does not independently verify the financing, valuation, milestones, chip commitments or amounts actually funded.

O’Driscoll divides the model-provider challenge into building competitive models and securing enough compute to meet demand. He believes OpenAI and Anthropic faced different constraints, but the relative product performance, capacity and management decisions are his interpretations. Both activities require significant resources, and future demand is difficult to predict.

To illustrate capacity risk, O’Driscoll imagines a provider growing from a $10 billion run rate toward $100 billion in two years and assumes several dollars of capital spending per dollar of additional capacity. He estimates hundreds of billions in investment shared with partners. These are hypothetical inputs and rough arithmetic, not a disclosed Anthropic budget or forecast.

His example shows the consequences of being wrong in either direction: too little compute can constrain service, while excessive contracted capacity may become expensive if demand disappoints. He attributes different levels of risk tolerance to the companies’ leaders, but those personal motivations and the suggested cost ratios are not independently established.

Lemkin says a reported OpenAI shortfall could be an early sign that compute and revenue do not always move one for one, while stressing that the definition of the reported miss matters and he cannot be certain. O’Driscoll agrees that capacity alone does not generate demand without products customers want. Both claims are economic interpretations, not proof about a specific company’s results.

The guests expect agents to consume more compute in some workflows. Lemkin reports that SaaStr’s Salesforce bill rose as its seat count fell, and says he does not know the precise token figure. O’Driscoll cites a rough comparison of agent and human usage; neither example establishes a universal consumption ratio.

Stebbings asks whether it is easier to sell excess capacity or acquire more when demand exceeds supply. The guests speculate that cloud providers might redirect or sublet contracted capacity to other customers, subject to the actual agreements. Long lead times make both over- and under-provisioning consequential; no particular resale transaction is established.

Lemkin argues that Google may benefit from its own models, cloud customers and infrastructure investment. O’Driscoll adds that Google and Amazon can combine their chips, cloud capacity and financing when competing for customers. The scale, utilization and economics he describes are estimates, not proof of the terms of any specific deal.

O’Driscoll uses a hypothetical data-center budget and Nvidia gross margin to illustrate why cloud providers might seek in-house chips. He acknowledges that evaluating performance and support requires technical detail. His calculations should not be treated as a verified cost breakdown or as proof that a particular chip is superior for every workload.

Asked to choose between Google and Nvidia, O’Driscoll says he would reluctantly favor Google on a risk-adjusted basis because it has several businesses; Lemkin prefers Nvidia as a more direct AI infrastructure exposure. They joke about concentrated portfolios, and O’Driscoll notes diversification and survivor bias. These are time-specific personal opinions, not investment advice.

Stebbings describes reports of Chinese action affecting Meta’s roughly $2 billion Manus acquisition and says investors had received distributions. The panel does not establish the regulator’s actual decision, jurisdiction, status of the deal, funds paid or any employee’s circumstances. Its ensuing discussion is a set of scenarios, not confirmation that a transaction was unwound.

O’Driscoll speculates about how regulatory leverage could differ for a company with significant local operations and raises concern about people who may remain in China. He explicitly does not know the scale of Meta’s local business. The proposed outcomes, travel restrictions and attributed government motives are conjectures, not findings in the interview.

The guests frame the reported dispute within broader US–China technology tensions. O’Driscoll attempts to imagine both sides’ perspectives, while Lemkin predicts political friction and social disruption as AI changes the economy. These are opinions about policy and future events, not evidence of a particular government’s intent or a guaranteed political outcome.

Stebbings cites reports that Thoma Bravo handed Medallia to creditors and references a $5.1 billion equity loss and about $3 billion in debt. The guests are uncertain about the exact figures and compare the case with Pluralsight. Their account does not independently establish the restructuring, the precise recoveries or the ultimate loss to any investor.

O’Driscoll recalls an acquisition financed with a substantial equity contribution and says a high purchase price can still produce losses even when initial debt is a minority of the deal value. He is unsure whether proceeds were previously returned to investors. His estimates of EBITDA, debt service and recovery are rough and should not be used as a factual reconstruction.

The guests distinguish leverage at acquisition from debt relative to a company’s current revenue and cash flow. O’Driscoll argues that a slow-growing business with significant obligations may have limited room to invest in a transition. Lemkin adds his view that customers may consolidate software vendors; neither establishes the exact reasons for Medallia’s reported outcome.

O’Driscoll discloses that his firm has an investment in Raptor, a customer-sentiment analysis company, as he discusses potential competition for older survey products. Lemkin mentions other software businesses he thinks may face debt pressure. O’Driscoll explicitly cautions against saying named firms are going under; no default or inability to repay is established for them here.

The panel describes possible warning signs in leveraged software, such as debt trading below face value, payment-in-kind features and refinancing deadlines. O’Driscoll says actual losses could affect limited partners’ risk appetite, particularly if investments they regarded as relatively safe underperform. He presents this as a conditional scenario, not a claim that the named companies will fail.

The guests consider a second consequence: private-equity buyers could become more selective, narrowing an exit route for venture-backed businesses that are too small to list or do not fit a strategic acquirer. Lemkin challenges the assumption that a large technology company will buy any plausible adjacent product. These are observations and forecasts, not evidence that all exits have disappeared.

O’Driscoll describes his firm’s portfolio framework as ‘fewer but bigger winners,’ with a diversified set of earlier investments and greater concentration in much larger companies. He uses rough revenue thresholds to illustrate how he thinks public-market requirements have risen. The thresholds are his judgment, not legal requirements for an IPO.

Lemkin predicts that some founders of mid-sized companies may merge with a trusted peer when they cannot see a credible standalone exit. He illustrates the idea with hypothetical companies and notes that combining them may not distribute cash to limited partners. O’Driscoll imagines operators consolidating businesses and improving cash flow, without claiming that any specific merger is planned.

Asked about private equity as an asset class, O’Driscoll warns against assuming a strategy cannot lose. Lemkin says he has seen a tentative offer for an AI-enhanced portfolio company and thinks some buyers are still active, though less frequently in his experience. He questions how durable new software businesses will prove and explicitly says he does not know the answer.

The guests debate whether large exits could compensate limited partners for many smaller companies that fail to exit. O’Driscoll says that is plausible but does not think the situation is quite that stark. Their discussion concerns portfolio scenarios and incentives, not a prediction that a fixed number of firms or companies will fail.

Lemkin welcomes Garry Tan’s call for greater precision in startup revenue reporting. He describes one investment that supplies several different ARR definitions, including core software, variable usage and committed revenue. He argues that those measures can diverge significantly and that concealing their definitions is more troubling than clearly disclosing them.

O’Driscoll says consistent definitions can help investors assess young companies and reduce disputes over overstated metrics. He also suggests that a market organizer benefits from maintaining trust in its ecosystem. His claims about prevalence and future adoption of guidelines are opinions, not established rates of misconduct.

The guests discuss a reported Thrive product and distinguish investments in enduring sports or media assets from a strategy of holding venture-backed public stocks for long periods. O’Driscoll says he is unsure of a specific early investment and does not claim to have verified the product’s full terms. Their descriptions are interpretations of the information they have seen.

Stebbings offers a countercase for sports assets: AI-driven changes in media consumption could affect rights revenue. O’Driscoll thinks ownership of live competition can remain valuable but agrees that media and athlete economics can change. Neither predicts a certain outcome for a particular team or fund.

The discussion compares US leagues with English football, where relegation can change a club’s economics. The guests use examples to discuss differing incentives and risks, while their political characterizations and estimates are personal opinions, not a comprehensive analysis of league rules or team valuations.

Stebbings asks about vehicles offering individual investors exposure to private companies. O’Driscoll discloses that he put the minimum $500 into one venture product to try its user experience, giving him indirect exposure to companies discussed on the show. He says he has not analyzed the underlying valuations and would do more work before committing a meaningful share of his assets.

Lemkin asks how to judge a quoted annual fee for an AngelList venture product. O’Driscoll contrasts fees on private investment vehicles with lower-cost public funds and says the comparison depends on underlying returns, expenses and carry. The quoted fee and product structure are the speakers’ description, not independently checked terms or a recommendation.

O’Driscoll notes that high gross returns are required for investors to justify a substantial fee drag, while past growth of a few private companies does not guarantee future returns for the asset class. Lemkin discusses investor incentives and says a leading fund’s decision not to increase its position can be a negative signal in his experience; that is not conclusive evidence about any company.

Stebbings discloses selling his Figma and Duolingo positions and says his Figma position was down about 40%. O’Driscoll first questions the impulse and then distinguishes the liquidity of public shares from a private investor’s longer commitment. Their discussion of when to sell is personal opinion, not advice tailored to a reader.

O’Driscoll says a public investor should understand why they hold a stock and compare it with an available alternative. Lemkin remains skeptical about how agents will use Figma and Duolingo, while wanting to see Figma’s response. These are their views of uncertain product and market developments, not established outcomes.

The panel closes with a brief mention of a reported AI lab in London and jokes about ending future episodes on a more positive note. The financing, restructuring and market outcomes discussed throughout remain subject to the underlying facts and future events.