AI Spending, Software Stocks and IPO Speculation: A 20VC Earnings Roundtable

Harry Stebbings, SaaStr founder Jason Lemkin and Scale Venture Partners’ Rory O’Driscoll review technology earnings, AI investment and software valuations before discussing a reported Anthropic fundraise, Sierra, litigation and management. Their company figures and transactions are presented as reports or speakers’ accounts; forecasts and legal interpretations are opinions, not established outcomes.

Stebbings calls the week a technology earnings ‘Super Bowl,’ citing $540 billion in combined revenue, $700 billion in AI capital spending and an Alphabet cloud backlog of $462 billion. Lemkin highlights rapid cloud growth and says fears that AI would immediately destroy Google’s search advertising business have not appeared in the economic results he is discussing. These figures are the panel’s cited figures, not independently reconciled here.

Lemkin says SaaStr’s own search traffic has increased and argues that Google is simultaneously defending its existing business and investing in the future. He sees allocation of compute among internal products, partners and customers as an important constraint. His small-site experience does not establish a broader trend for all publishers.

O’Driscoll calls the quarter unusually aggressive: large companies are growing at scale while committing more of their cash flow to capital expenditure. He observes that established incumbents are investing forcefully rather than leaving the opportunity to startups. Those choices may strengthen existing leaders, but they are not proof that the resulting investments will all pay off.

O’Driscoll thinks much of the hyperscalers’ early AI growth comes from supplying compute to model developers and distributing their products. In his view, that creates a question about where long-term intellectual-property value will accrue. He treats Google’s own Gemini model as a different position from merely hosting another provider’s models.

He compares reported Gemini token production with his estimates for private model providers and argues that coding adoption may be more concentrated than benchmark results imply. The figures are not independently verified or necessarily comparable in the conversation. Lemkin points to Palantir CEO Alex Karp’s claim that models are interchangeable for his company, while acknowledging he is not sure whether that view will hold.

The guests agree that Palantir’s enterprise offering may capture value at a different layer from model hosting, but disagree about how durable the distinctions will be. O’Driscoll sees model providers as potentially retaining the most attractive part of the value chain; Lemkin says the relationship among application, model and infrastructure layers remains fluid.

Lemkin argues that large public companies have sophisticated financial teams and understand the risks of their AI investment, even if the bet later proves wrong. O’Driscoll replies that confidence during a capital-spending boom does not guarantee correct allocation. Both distinguish potential overinvestment from a proven bubble and note that established businesses would remain if spending were reduced.

O’Driscoll claims Microsoft’s business outside the AI-related areas he identifies would be flat or slightly down and says this makes AI growth important to its valuation. That is his calculation and depends on how business segments and growth are attributed; the interview does not independently substantiate it. Stebbings cites $37 billion in AI annual recurring revenue and $190 billion in capital spending as figures for discussion.

Lemkin argues that when investors permit a company to spend aggressively, management gains a window to invest that may close when growth slows. O’Driscoll agrees markets influence those decisions but warns that spending because permission exists, rather than because returns justify it, increases the risk of groupthink and poor allocation.

Lemkin recounts talking to a software CEO focused on shaving model costs and argues that a company can miss the chance to build a better product by prioritizing short-term margins. O’Driscoll agrees that positive-return AI spending can be worthwhile, while preferring a more measured approach. Their disagreement is about strategy, not evidence that every cost-saving effort is harmful.

O’Driscoll cites Google’s reported increase in Gemini token production and contrasts it with his estimate of Anthropic’s growth. He interprets the comparison as evidence of stronger coding traction at other providers, but the discussion does not verify the numbers or establish equivalent measures. Product benchmarks, usage and revenue can tell different stories.

Stebbings cites Meta revenue, earnings and a higher capital-expenditure range and asks why the market reacted differently from Google. O’Driscoll thinks Google’s cloud revenue is more directly attributable to compute demand, whereas Meta’s prospective payoff is harder to model. That is his interpretation of investor reaction, not a demonstrated explanation of a stock move.

O’Driscoll says Meta has discussed improving advertising performance with AI but argues that a proper comparison would require controlled tests to estimate incremental lift. He speculates that management is also investing for future consumer experiences. Neither his estimate of the benefits nor his explanation of management’s intentions is independently established here.

Lemkin says investors can model infrastructure depreciation and cloud revenue more easily than indirect future benefits at Meta. O’Driscoll agrees the uncertainty could lead to a greater discount, while acknowledging the investment could ultimately succeed or fail. Both describe a valuation problem, not a conclusion that Meta’s strategy is necessarily wrong.

Asked to choose among the large companies, O’Driscoll tentatively selects Amazon over Microsoft, describing it as a weakly held, on-the-fly view. He points to AWS distribution and access to model products, and cites his own interpretation of Microsoft’s non-AI growth. These are time-specific opinions, not recommendations or guarantees.

Lemkin argues that an increase in the number of applications could benefit cloud infrastructure in addition to model providers. He speculates that this software-building boom may give Amazon and others another source of demand, while seeing less direct benefit for Meta. The scope and timing of that boom remain uncertain.

Stebbings cites Palantir’s reported $4.45 billion remaining performance obligations, growth in that measure and a high Rule of 40 figure. Lemkin says he heard management describe rapid acceleration, although he could not fully reconcile the comments with forward guidance. The company figures and growth interpretations are presented as the panel’s account.

O’Driscoll argues that large enterprise leaders need a credible company-wide AI initiative, not just a collection of small individual subscriptions. He thinks Palantir’s experience with large customers and ability to assemble enterprise projects may fit that demand. His account is a market thesis, not proof that any specific deployment delivers the results promised.

Lemkin recalls that at Adobe a small number of large initiatives consumed significant budgets and took years to deploy. He says Palantir management described a faster buying process in which multiple stakeholders now attend meetings together. Those examples are his recollections and interpretation of the company’s comments, not a measured average across enterprises.

O’Driscoll adds that strong enthusiasm can compress evaluation but also suppress skeptical voices and lead to misallocation. Asked about Palantir’s valuation, he says a great business can still be expensive and that the share price assumes substantial future growth. His scenarios for doubling and valuation are speculative, not an investment result.

The guests argue that outside expertise can be valuable when organizations lack AI implementation skills. O’Driscoll describes calling a specialist when he is stuck rather than spending days solving a problem alone. Lemkin predicts demand for practitioners able to combine domain work with AI tools, but his salary and employment examples are conjectural.

O’Driscoll notes Apple’s reported strong results despite what he views as a less visible AI story. The guests then discuss memory-chip costs: he suggests that higher component prices may raise capital budgets even when the physical amount purchased stays the same. His prediction that costs could reach consumer devices is a scenario, not a confirmed future price change.

Lemkin discusses a reported Mac mini price difference and says he is unsure about the configuration comparison. He worries that higher prices and expensive housing can affect households differently from highly paid technology workers. The examples are observations and concerns, not a verified measure of the AI sector’s effect on inflation.

Stebbings cites gains in Atlassian, Twilio and Five9 shares and asks whether the ‘SaaS apocalypse’ is over. Lemkin welcomes the reported acceleration at Atlassian and Twilio but cautions that other companies may report weaker results. A rebound in selected stocks does not establish an industry-wide recovery.

Lemkin distinguishes Atlassian’s reported success selling AI features to its existing base from Twilio’s potential new customer growth. He notes uncertainty in the available Twilio disclosure and suggests that monetizing existing customers is valuable but may not answer whether the customer base is expanding. His figures and causal explanations require confirmation.

O’Driscoll says a company priced as if its business will disappear can rebound if fundamentals remain sound, without becoming a new frontier AI company. He uses hypothetical revenue multiples and growth rates to explain a bounded valuation range. His examples are investor opinions, not targets for any stock.

Lemkin recounts former Twilio CEO Jeff Lawson’s view that AI companies would need communications APIs. He thinks Twilio’s existing infrastructure may benefit as new products scale, although a superior competitor could emerge. His observation that specific tools use Twilio is an example, not proof of future revenue growth.

Asked which other software companies might recover, Lemkin looks for both AI monetization and new customers. He points to HubSpot’s stated plan to make its platform accessible to agents and says its results will be a useful test. He predicts possible reacceleration if the plan works, but recognizes execution and pricing remain uncertain.

O’Driscoll expects reacceleration to be the exception rather than the rule, while noting that selling new products to existing customers can still matter even in a mature market. He and Lemkin agree that several stronger examples would help investors distinguish viable software businesses from those struggling to adapt. Their expected valuations and outcomes are forecasts.

The panel turns to a reported $44 billion Anthropic revenue run rate and asks whether enough developer demand exists to sustain growth. O’Driscoll says the key unknown is long-run token spending relative to engineering salaries in an AI-intensive organization. His scenarios use assumptions about task automation and are not verified financial projections.

O’Driscoll illustrates how a 20%–30% token-to-salary ratio would imply a much larger potential market than a 5% ratio. He thinks coding currently offers clearer automation opportunities than some other knowledge work, but labels his wider estimates as his own judgment. The relationship between token purchases, salaries and model-company revenue is not established by these examples.

The guests consider a counterintuitive possibility: if AI makes developers more productive, growing companies might hire more of them rather than fewer. O’Driscoll frames this as a hypothetical return-on-investment effect; Lemkin agrees it depends on growth and access to capable people. Neither claims this will happen in every company or occupation.

Lemkin says SaaStr’s two semi-autonomous marketing and customer-success agents cost about $254 in a month, much less than the other guests guessed. He presents it as his own operating example and believes the agents are useful. The discussion does not independently measure their output or establish that the figure includes every associated expense.

O’Driscoll says his team’s sample of engineering spending and Lemkin’s example raise questions about assumptions for token intensity outside coding. He wonders whether heavy users are early examples of wider adoption or are spending inefficiently. He explicitly says he does not know the answer and is surveying portfolio companies.

The guests note that cheaper tokens could make more tasks economical even as consumption rises. Lemkin imagines frequent code review as one possible new use; O’Driscoll offers a rough estimate of price declines and admits the long-run result is hard to forecast. Their examples do not establish a fixed future rate of cost reduction.

Stebbings raises reports that Anthropic is seeking $50 billion at a $900 billion valuation and asks what it means for a possible IPO. O’Driscoll says he previously expected the company to go directly to an offering but now thinks a private raise could increase flexibility. The financing, proposed terms and any listing are not independently confirmed in the discussion.

O’Driscoll sketches why fast growth could force a model company and its partners to commit capital well before revenue arrives. His multiples of revenue and compute spending are an illustrative model, not Anthropic’s disclosed obligations. He argues that a larger balance sheet could reduce financing pressure if the estimates prove broadly right.

Lemkin thinks a large private round could make an IPO that year less urgent, particularly if a competitor also waits. O’Driscoll agrees financing would increase optionality but says favorable conditions could still prompt a listing. Their views of timing, shareholder rights and investor demand are guesses; no offering date or transaction outcome is established.

Stebbings cites reports that Sierra is raising $950 million at a $15.8 billion valuation and has $150 million in annual recurring revenue. He contrasts the implied high multiple with an estimated $400 billion customer-service labor market. These are reported figures and market-size estimates, not verified financing terms or proof of addressable revenue.

Lemkin questions whether customer-service spending will expand enough to support a very large outcome, even if agents replace some tasks. He suggests that sales and upsell could enlarge the opportunity, but says he does not know whether Sierra will become a $100 billion company. O’Driscoll distinguishes replacing existing software from potentially capturing labor expenditure.

O’Driscoll argues that customer-service applications can add workflow and domain value beyond the underlying language model. He guesses model costs may be less than 10% of Sierra’s revenue, explicitly presenting this as an estimate. Both guests warn that a roughly 100-times revenue multiple assumes an aggressive future and does not guarantee returns.

The panel debates why investors might pay high prices for sought-after assets. Lemkin suggests some investors perceive downside protection, then acknowledges that venture investors can overstate that protection and potential acquisition outcomes. O’Driscoll likewise cautions that historical experience includes worse outcomes than optimistic exit scenarios.

Stebbings introduces the Musk–OpenAI litigation and cites purported testimony and estimates of executive holdings. O’Driscoll discusses possible issues including limitations periods and standing, but his account is a lay interpretation of an ongoing case, not a judicial ruling. His descriptions of testimony, private motives and likely outcomes are not independently verified here.

Lemkin cites a reported Coinbase reduction and interprets CEO Brian Armstrong’s comments as a push for leaders who can contribute directly as well as manage. He strongly favors hands-on work with AI. The conversation does not establish the proportion of managers able to do this, or prove that AI caused a particular reduction.

O’Driscoll says he is generally skeptical when executives attribute layoffs to AI because overhiring, slower growth, capital spending and changing skill needs can also play a role. He nonetheless finds Armstrong’s emphasis on a hands-on culture persuasive. That assessment of the executive and the reasons for cuts is his opinion, not a verified causal finding.

O’Driscoll argues that leaders should use new technology enough to understand frontline decisions, rather than relying entirely on reports. Lemkin wants an even more direct role: he says a SaaStr customer-success agent contacted sponsors and argues an executive should understand and improve that workflow. His example describes his organization’s experience, not an independently measured comparison with all human teams.

Lemkin argues that marketing leaders should be able to direct AI-assisted campaigns themselves. O’Driscoll asks whether those who do so will actually perform better; he says adoption would spread if that turns out to be true. Their predictions about roles and successful managers are opinions, and the episode provides no evidence for the percentages guessed in the exchange.

Stebbings closes by arguing that working from home on Fridays can become an extended weekend. O’Driscoll describes his own firm’s preference for Monday and Friday office meetings, with limited exceptions, while Lemkin emphasizes that employees may reasonably prefer different ways of living. The panel offers personal workplace preferences, not a controlled comparison of remote-work productivity.

The conversation ends with the guests acknowledging that people choose different trade-offs between work and personal life. Their debate reflects their views as founders and investors; it does not establish that remote employees work less or that a single management model fits every company.