Marc Benioff on Enterprise AI, Anthropic’s Funding and Klarna’s IPO

In this August 2025 roundtable, Harry Stebbings, Jason Lemkin and Rory O’Driscoll speak first with Salesforce co-founder and CEO Marc Benioff about enterprise AI. After Benioff leaves, they discuss Meta, reported Anthropic fundraising, Klarna’s proposed IPO and venture portfolio strategy. The discussion’s figures, reports and forecasts are presented in their original time context.

Benioff says he is skeptical of claims that artificial general intelligence is imminent. He regards current language models as useful technology but argues that their human-like responses should not be confused with human consciousness or compassion. These are his views about the technology, not a settled scientific conclusion.

Benioff warns that people can become overreliant on inaccurate AI output, citing articles about medical advice. He calls for practical use alongside awareness of limitations. His account is a general caution, not a clinical evaluation or a claim established about specific practitioners in the episode.

Asked whether Salesforce needs to match Meta’s reported spending on AI talent, Benioff says no. He says the company is focused on enterprise use cases and what works in practice. As Salesforce’s CEO and shareholder, he has a direct interest in the products and strategy he describes.

Benioff describes using Salesforce’s own agentic support system and says human support headcount declined from roughly 9,000 to 5,000 as people were reallocated. This is his account of staffing and product impact, not an independently audited measure of what AI alone caused.

He says Salesforce has accumulated many leads it could not previously contact and describes an agentic sales system intended to follow up and connect to human teams. The figure of more than 100 million leads is his cumulative company claim; the discussion does not establish their quality, consent status or conversion.

Benioff predicts that Salesforce’s software will increasingly incorporate agents working with people. He presents support, sales and Slack as examples of the direction. The breadth of future deployment and its effects on jobs and revenue remain forecasts.

Stebbings asks why AI has not yet produced a larger growth lift. Benioff disputes the premise and says Salesforce’s combined Data Cloud and AI business exceeded $1 billion in revenue. The combined figure should not be read as revenue from AI agents alone; the episode does not independently verify the breakdown.

Benioff says the data layer and agents must work together so software can use company information more effectively. He compares Salesforce’s opportunity with Snowflake, Databricks and Palantir. His customer, growth and competitive claims are those of a company executive discussing his own products.

Asked about Palantir, Benioff says its integration of data and analytics and its access to certain government customers prompted Salesforce to reassess where it sells. He praises its success while making competitive claims about Salesforce’s contracts, pricing and products that are not independently established here.

O’Driscoll asks whether forward-deployed engineers are a genuinely new idea. Benioff says Salesforce has long sent technical and service teams to customers, but finds it useful to begin building with a prospective customer before a deal is finalized. He sees a model the company could adopt more explicitly.

O’Driscoll notes the arithmetic of scale: 10% growth on $40 billion is $4 billion of incremental revenue. A new AI business can be large in absolute terms without dramatically changing the growth rate of a company Salesforce’s size. The figures are illustrative, not a forecast for a specific quarter.

Benioff says every order of magnitude in company size brings different products, customers and management challenges. He wants to learn from competitors without dismissing them. The guests agree that the consequences of AI may take longer to show up in large-company financials than enthusiastic predictions suggest.

Stebbings asks whether agents and protocols will make SaaS applications disappear. Benioff strongly rejects the claim that established applications will become mere databases. He expects people to continue using applications while agents and data interoperate with them; this is his product view and serves Salesforce’s commercial interests.

O’Driscoll reframes the question: even if Salesforce remains the system of record, other companies might own some of the interface used by salespeople. Benioff says an agentic layer and an ecosystem can connect applications and data. He hopes much of that ecosystem will be built on Salesforce.

Stebbings asks whether AI will eliminate sales-development representative roles. Benioff says automation can help address leads that were not previously contacted and allow people to move into other sales roles. Lemkin predicts substantial redeployment; O’Driscoll questions how easily every entry-level worker can make that transition.

The panel agrees that enterprise organization and job design may change, but disagrees about how smoothly workers will move into new roles. Benioff emphasizes additional opportunities for entrepreneurs and says Salesforce is hiring. These are perspectives on a possible transition, not measured employment outcomes.

Asked to choose between OpenAI and Anthropic at the valuations suggested by Stebbings, Benioff declines to pick and praises both. He discloses that Salesforce owns a stake in Anthropic, which he puts at 1%. The ownership figure is his statement in the episode and a relevant financial interest.

After Benioff departs, the guests discuss reports about Nat Friedman’s position in Meta’s reorganized AI division. Stebbings questions why an autonomous investor would take a role reporting through another executive. O’Driscoll says the organizational structure and personal motivation are separate questions and acknowledges that the promises made are unknown.

O’Driscoll says an operator may value direct involvement in a major technical project more than running a fund, while Lemkin prefers autonomy in the hypothetical. They cannot know Friedman’s reasons or the actual terms of his role. Their comments are opinions about career choices, not assertions of private conversations.

The panel considers Meta’s stock movement and its AI expenditure. O’Driscoll separates the cash-generating core business from the question of whether new investments will earn a return. He says an organizational announcement alone does not tell investors whether the strategy will succeed.

O’Driscoll cautions that expensive stocks can react sharply to modest changes in expectations, but says nobody knows when a market correction will arrive. The guests discuss volatility as a reason not to infer too much from one day’s price move. Their examples are market opinions, not trading recommendations.

Stebbings cites reports that Anthropic’s proposed fundraising grew from $5 billion to $10 billion and was heavily oversubscribed. The panel sees strong demand for leading AI companies, but the possible raise, allocation and oversubscription are reported or asserted figures, not completed financing verified by the episode.

Lemkin discusses why large venture firms might deploy more of their limited partners’ money in a sought-after round. O’Driscoll says the investment could work if enough operational progress follows, but investors must evaluate the price and eventual market size. Fund economics do not remove the risk to underlying investors.

O’Driscoll builds a hypothetical in which a company’s revenue run rate grows from $1 billion to $9 billion and then severalfold again. He says the result could support a very large valuation, but also notes that growth could decelerate much faster. The projected numbers are scenarios, not reported future revenue.

The central question, he argues, is the eventual market for foundation-model APIs and how much of that market any provider can capture. He is unsure it can reach the very large revenue levels implied by some valuations. Lemkin thinks labor-saving applications may unlock substantial new budgets; both acknowledge uncertainty.

O’Driscoll estimates what an AI sales agent could add to Salesforce revenue under hypothetical adoption and pricing, then assumes a share flows to model providers. Lemkin gives his own small company’s spending as a counterexample. Their ratios are illustrative, not Salesforce guidance or a market-size measurement.

The guests debate whether AI providers can capture enough of the value created for businesses to justify projected valuations. O’Driscoll points to actual growth as evidence against dismissing the opportunity, while maintaining that extrapolating extraordinary growth is risky. No future revenue outcome is established.

Stebbings worries about concentration in the largest public technology stocks. O’Driscoll says a good technology can also be overpriced and discusses his own risk tolerance and possible diversification. The comments reflect individual portfolio choices and contemporary views, not advice for a listener’s circumstances.

The panel observes improved results at some public software companies. Lemkin sees encouraging signs of AI-related reacceleration, while O’Driscoll notes that stock prices respond to results relative to expectations. A small beat at a low valuation can produce a sharp move without implying that every software business has returned to hypergrowth.

Lemkin says coding applications can generate demand for underlying databases and hopes established vendors benefit as well as startups. The panel debates whether fast-growing AI revenue will endure. Stebbings has an interest in Lovable, which is relevant when he argues for the durability of that category.

Stebbings raises Klarna’s IPO filing and cites a proposed valuation range around $13 billion to $15 billion, after a $45 billion financing and a later down round. Lemkin focuses on reported slowing growth; O’Driscoll says the minimum viable growth rate for a public offering depends in part on company scale.

O’Driscoll argues Klarna should be assessed as a financial-services business, including its lending and credit performance, rather than solely by software multiples. He expects a valuation reflecting a more mature business, but the IPO price and post-listing performance are unknown in the conversation.

The guests debate how earlier investors could fare after the valuation reset. O’Driscoll speculates about the effect of dilution and preference terms, while Lemkin says contractual protections might change the outcome. Both acknowledge they need to review the filing; their discussion is not a determination of any investor’s actual loss or rights.

The panel discusses Netskope’s filing using figures cited in the episode, including roughly $700 million in annual recurring revenue and growth above 30%. O’Driscoll offers a rough valuation range but says he cannot know where the stock will trade on day one. His estimate is a prediction rather than a realized price.

Stebbings cites an analysis counting 72 a16z seed investments versus 27 for the next large-fund investor. O’Driscoll says the difference suggests a platform operating at a different scale. The count is drawn from an analysis mentioned in the show; its time period and methodology are not established here.

O’Driscoll argues that a broad seed program may help a large platform identify exceptional companies for later, much larger checks. In his view, the platform’s overall outcome will depend heavily on whether it backs a few outliers at the right price. That is his interpretation of the strategy, not a reported fund return.

Stebbings introduces investor Martin Casado’s argument that non-consensus early investments can be dangerous because follow-on capital tends to move toward consensus. O’Driscoll says many good investments follow a broad technological shift, while a portfolio can still reserve room for less conventional ideas.

Lemkin adds a practical constraint: founders outside fashionable categories should not assume later investors will be available. A sound company can struggle to fund expansion if most potential backers concentrate elsewhere. His observations about the financing environment are investor experience, not a universal rule.

O’Driscoll describes a hypothetical fintech company growing quickly but receiving less attention because it is not AI. He says the investment can still make sense at an appropriate price and with capital-efficient operations. The absence of a fashionable label should affect the financing plan, not automatically disqualify the company.

He distinguishes betting outside the financial consensus from betting against a major technical shift. A company can be different without assuming the broader trend will fail. Lemkin stresses that high burn and scarce follow-on capital can make an otherwise attractive early deal difficult to sustain.

O’Driscoll says he regrets both missing some non-consensus opportunities and underestimating the scale of the AI trend. He also warns about survivorship bias: investors remember unconventional winners more readily than the many ideas that failed. The examples do not establish actual failure rates for a cohort.

The guests conclude that a consensus investment still requires a sensible price, while a non-consensus investment brings additional financing and execution risks. O’Driscoll reflects that he underestimated the capital flowing into AI infrastructure; the large spending numbers he cites are his estimates, not verified totals.

The panel closes by reflecting on its conversation with Benioff. Lemkin says he would like Salesforce to buy companies from his portfolio, a commercial interest relevant to his comments, while praising Benioff’s pragmatic framing of AI. The exchange ends with friendly debate over how hard the interviewers challenged their guest.