Summary

Marc Benioff’s core distinction is that today’s LLMs are powerful enterprise tools, not conscious beings or evidence that AGI is imminent. They combine improving but finite algorithms with relatively finite internet data, can feel intelligent as ELIZA did to him at 16, and create dangerous “hypnosis” when people outsource judgment. His operating conclusion is pragmatic: ignore the talent frenzy and make every Salesforce product agentic.
Salesforce’s own deployment supplies the episode’s strongest evidence that agents already change enterprise economics. An omni-channel supervisor helped cut human support agents from 9,000 to 5,000, with staff rebalanced elsewhere, while agentic sales can finally contact more than 100 million historical leads Salesforce never had enough SDRs to call. Data Cloud plus AI has exceeded $1 billion in revenue and is Salesforce’s fastest-growing cloud product in 26 years.
Benioff rejects the claim that SaaS applications collapse into CRUD databases, instead underwriting a three-layer market of data, applications, and interoperable agents. Humans still need applications in their flow of work, while an open agentic layer can coordinate them through ecosystems such as Slack and AppExchange. The discussion leaves room for third-party interfaces and agents even if Salesforce remains the system of record.
Palantir has become both a competitive benchmark and a valuation provocation for Salesforce. Benioff called Foundry’s integration of data and analytics “very inspiring,” credited forward-deployed engineers as smart pre-contract customer commitment, and said Palantir’s pricing makes Salesforce look cheap. Yet he emphasized the scale gap: roughly $4 billion versus Salesforce’s $41 billion of revenue, while asking, “How do I get that 100 times revenue multiple?”
Anthropic’s expansion of its round from $5 billion to $10 billion, reportedly four-times oversubscribed, is defensible only if foundation-model demand becomes enormous. Revenue moving from roughly $1 billion toward $9 billion or $10 billion creates a trajectory where even severe deceleration could produce $40 billion to $50 billion next year. Rory’s unresolved fork is whether API demand is closer to $50 billion or $500 billion—and whether agents can command $20,000 to $40,000 per worker rather than $2,000.
The model-provider TAM looks less obvious after tracing enterprise agent revenue through to inference costs. Rory’s Salesforce thought experiment turns a 30% uplift on an estimated $12 billion Sales Cloud into $3.6 billion of agent revenue, but only $720 million for model providers at a 20% cost share. Jason Lemkin’s counterexample—roughly $500,000 spent on 11 agents by a tiny team—shows why the round is cheap if unusually high agent-to-software spending ratios persist.
Public-market risk comes from expectations and concentration, not from one volatile trading day. Meta’s core business is throwing off cash while Mark Zuckerberg commits a discussed $60 billion to $70 billion to an AI business whose economics remain unexplained; at elevated multiples, “everything that goes wrong, no matter how tiny, gets magnified.” Conversely, MongoDB’s 27% jump and rebounds at Box, Okta, and Zoom show how modest reacceleration can rerate SaaS names once pessimism is already priced in.
Venture capital is bifurcating between industrial-scale consensus investing and capital-efficient bets the follow-on market may ignore. Andreessen completed 72 seed deals against Sequoia’s 27, treating seed like “cheap milk in the supermarket” that sources a few outliers into which it can invest billions. Martin Casado’s warning lands because 10 deals reportedly absorb 40% of venture capital: non-consensus founders can still win, but must price on fundamentals and assume, “Don’t expect any money.”

Deep dive

1. Benioff rejects AGI hypnosis while embracing useful AI

Benioff’s opening objection was semantic and strategic: an “AGI head” sounds like “an oxymoron.” He is not saying machine intelligence could never arrive—“we’ve all seen those movies”—but that the technology available in 2025 does not justify claims about what is imminent.
His model of current LLMs has two finite inputs: a set of algorithms that improved incrementally over roughly five years, and a relatively finite body of data drawn from the internet. Their outputs can feel uncannily intelligent, but that feeling is not evidence of consciousness.
The analogy was personal: ELIZA on his TRS-80 Model 1 felt like a person when he was 16. Yet an LLM “doesn’t have a childhood,” has not suffered, and lacks compassion; physicians becoming intellectually lazy and giving bad advice after over-trusting AI illustrated the practical danger of confusing simulation with judgment.

2. Salesforce is using itself as the agentic-economics test case

Benioff dismissed the need to join the extravagant AI-talent bidding war: “We’re not.” His principle is that “tactics must dictate strategy over time in enterprise software,” with customer deployments—not claims about AGI—defining Salesforce’s next architecture.
At help.salesforce.com, an omni-channel supervisor coordinates human support staff and digital agents. That system helped reduce human support agents from approximately 9,000 to 5,000; Benioff corrected the hosts’ later misstatement of the numbers and stressed that the headcount was rebalanced into other growing areas.
Salesforce accumulated more than 100 million inbound leads over 26 years that it lacked enough SDRs to call back. Agentic sales can now contact those people, conduct conversations, and connect them into the company’s new sales product, which Benioff said would be shown at Dreamforce.
His categorical product call: “I don’t think that there will be a piece of software that we sell that will not be agentic.” The promise extends beyond Sales Cloud and Service Cloud into Slack, with humans and agents working together rather than one simply replacing the other.

3. Data quality, not model mystique, anchors Salesforce’s AI thesis

Benioff tied AI accuracy to a federated Data Cloud that harmonizes enterprise information, offering Salesforce’s Informatica acquisition as part of that foundation. Loading Salesforce’s entire website into Data Cloud let Agentforce handle as many customer interactions as its support agent, replacing navigation with conversation.
Data Cloud plus AI now exceeds $1 billion in revenue and, according to Benioff, is Salesforce’s fastest-growing cloud product in 26 years. Agentforce had not been announced a year earlier and did not ship until the previous November, yet already had thousands of customers and deployments.
Against Snowflake, Databricks, and Palantir Foundry—each described as occupying the $3 billion to $4 billion revenue range—Benioff framed the gap as attainable: “They’re in my sights.” His Star Wars operating metaphor was a TIE fighter pilot’s instruction to “stay on target.”
Asked to choose between OpenAI at “300” and Anthropic at “170,” Benioff declined the valuation contest, called both great companies, and disclosed that Salesforce owns 1% of Anthropic. Anthropic’s enterprise orientation was the only differentiation he volunteered.

4. Palantir has changed Salesforce’s product, pricing, and delivery calculus

Benioff called Palantir’s combination of Foundry and analytics “very cool and amazing” and “very inspiring.” Salesforce’s corresponding data foundation comprises Data Cloud, an agentic Tableau, Informatica, and MuleSoft, alongside a need for government certifications and a renewed focus on segments it historically did not serve.
The US federal government is already Salesforce’s largest customer, with the Veterans Administration and GSA cited alongside a recent US Army contract won against Palantir. Benioff nevertheless acknowledged that Palantir reaches groups Salesforce traditionally has not targeted, and that its dealmaking got his attention.
Palantir’s public price lists produced an unexpected reaction: “My prices are too low.” Benioff said Salesforce’s prices were lower and its products easier to use, while openly coveting the valuation: “How do I get that 100 times revenue multiple?”
He saw forward-deployed engineering as both old and genuinely new. Salesforce already has sales engineers, professional services, and partners, but Palantir branded and operationalized the willingness to start building before a contract exists: “We’re gonna make a bet that we’re gonna start doing business together.”

5. SaaS applications survive, but the interface and workforce both change

Benioff called predictions that SaaS becomes a collection of CRUD databases “one of the greatest disservices” done to CIOs and software CEOs. His rebuttal was blunt: users still need applications, and “I need apps and I need agents and I need them to work together.”
Rory separated that argument from the harder ownership question: Salesforce may remain infrastructure while salespeople use superior third-party interfaces or agents above its data. Benioff accepted an ecosystem model comprising persistent application functionality, an interoperable agentic layer, and open connectivity patterned on Slack and AppExchange.
For small companies, Benioff expects amplification rather than contraction. Hearing that Jason’s organization already had agents listening to calls, coaching staff, and delivering Slack updates led him to predict “an order of magnitude more SMBs,” because entrepreneurs can operate with capabilities once reserved for enterprises.
The employment dispute remained unresolved. Benioff said Salesforce could rebalance SDRs into sales roles; Harry doubted that unskilled entry-level SDRs could simply be redeployed, while Jason estimated Salesforce might move 70% into enterprise selling or forward-deployed engineering. Benioff’s broader point was that the future enterprise software company will not be staffed or structured exactly like the SaaS company of the past.

6. Meta’s AI reorganization is rational, but its economics remain opaque

Rory viewed Meta’s four-part structure—pure science, foundation models, applications, and infrastructure under Alexandr Wang—as sensible after roughly $20 billion of talent acquisition. His soccer analogy: once the rich club buys every star, a manager still must decide “who’s gonna play forward, who’s gonna play striker, and who’s gonna play fullback.”
Harry’s actual confusion was why Nat Friedman would surrender fund autonomy to report to Wang rather than Zuckerberg. Rory understood an exceptional operator leaving venture, but questioned the level of the transition; Jason’s judgment was that a few elite rooms might not compensate indefinitely for giving up one’s own shop and potential carry.
The stock problem is the disconnect between Meta’s cash-generating core and its new AI program. Rory said current advertising optimization appears more rooted in older AI than LLMs, leaving investors to determine whether $60 billion to $70 billion of spending creates a business or simply consumes cash.
Zuckerberg’s record permits opposite outcomes: success could resemble the value attributed in the discussion to Instagram and WhatsApp; failure could resemble the Metaverse. Jason cautioned against reading too much into a 6% move when Meta’s beta was cited at 1.59 and Nvidia’s at 2.3: “The beta’s too high.”

7. Anthropic’s trajectory makes both the bull case and the slowdown unprecedented

The round reportedly doubled from $5 billion to $10 billion and was four-times oversubscribed. Rory’s explanation was scarcity: public managers seeking pure AI exposure have few at-scale choices beyond OpenAI and Anthropic, while the companies have an obvious use for proceeds—buying GPUs.
Jason offered a possible fund-economics explanation: Iconiq and Lightspeed may be able to tap very large pools of LP and sovereign capital while retaining substantial economics on the deployment. He asked whether, at the GP level, deploying $6 billion rather than $2 billion could involve similar risk with much greater upside.
The operating case is not facially absurd. If Anthropic moves from roughly $1 billion to $9 billion or $10 billion of revenue, even slowing from 9x or 10x growth to 3x could imply more than $30 billion next year; Jason said $40 billion was possible, while Rory’s thought experiment reached $50 billion-plus.
That creates two extraordinary possibilities: revenue becomes enormous at unprecedented speed, or growth decelerates faster than almost any company’s before it. Rory invoked his corporate version of Newton’s law—“things in motion stay in motion”—while conceding he still expects a sharper slowdown than the market may assume.

8. The $100 billion model-revenue thesis needs very expensive agents

Rory’s central valuation question was whether foundation-model API demand is closer to $50 billion or $500 billion. In the larger market, two leaders capturing a substantial share can justify today’s valuations; in the smaller one, “a lot of these people are gonna be sad.”
A $100 billion revenue target is approximately 2.5 times Salesforce’s current $40 billion scale despite Salesforce’s dominant CRM position. Rory’s “big aha” was that agents cannot merely cost $2,000 per engineer: they must remove large pieces of labor budgets and become worth $20,000, $30,000, or $40,000 per worker in enough use cases.
His Salesforce thought experiment began with an estimated $12 billion Sales Cloud. A 30% AI-SDR uplift adds $3.6 billion, but if LLM costs absorb 20%, model providers receive only $720 million—roughly $1 billion after generous rounding—from one of software’s largest, most relevant deployments.
Jason supplied the bullish counterexample: his small organization has four Salesforce seats but is nominally spending about $500,000 on 11 AI agents. Rory’s answer was conditional: if even a fraction of that ratio persists broadly, Anthropic’s round is cheap; if incremental agent spending only matches core software spending, the model-provider share remains modest.

9. Expectation gaps are reviving SaaS while exposing expensive AI stocks

Rory would not predict when elevated AI markets break: “You’ll know when it’s happened because it’ll hurt.” High valuations magnify every disappointment, leaving only two resolutions—growth fills the earnings gap, or prices fall—without a knowable timetable.
He nevertheless believes concentration will revert toward the mean and was considering a roughly 5% allocation to core commodities, not a wholesale exit. His formulation was the episode’s cleanest market warning: “Something can be amazing and still overpriced.”
MongoDB’s 27% jump, alongside strength at Box, Okta, and Zoom, reflected the inverse setup. Once “SaaS is dead” pushes a company toward roughly 5.5 times revenue, beating expectations by a few percentage points can cause a sharp rerating; a hypothetical Salesforce acceleration from 10% to 13% would be similarly explosive.
Jason connected the rebound to AI infrastructure demand: Lovable and Replit-style applications spin up Neon or Supabase databases at enormous rates. Databricks buying Neon for $1 billion now made more sense to him, while Wix’s $80 million Base44 acquisition looked extraordinary after the product reportedly generated $1.2 million in the preceding week—though that velocity also raised durability questions.

10. Klarna and Netskope illustrate two very different IPO setups

Klarna filed around a valuation of $13 billion to $15 billion after a SoftBank-led $45 billion round and a later repricing near $6.5 billion. Growth slowed from 24% to 20%, which Jason called the IPO “hard deck”: “Pull up. Pull up. Pull up.”
Rory argued that transaction size changes the threshold; at roughly $14 billion, Klarna is large enough to go public at 20% growth. But its million dollars of revenue per employee signals maturity, and investors should treat it as a scaled financial-services company whose valuation depends on lending quality—not a software company making “software noises.”
His scorecard was unsparing: investors at approximately $6 billion were right, those at $45 billion were wrong, and “venture is a game played by 6,000 people, and in the end, Sequoia wins.” SoftBank is not necessarily wiped out; Rory guessed that Sequoia would not have left a block in the documents, but Jason noted that a ratchet could exist and that the S-1 needed to be checked. Absent such protection, Rory expected SoftBank’s position to convert and trade at perhaps 30 to 40 cents of its original purchase price.
Netskope offered the cleaner growth story: about $700 million of ARR, reaccelerating from 30% to 33%, against a 2021 valuation around $7.4 billion. Rory could envision demand moving an initial $5 billion to $6 billion price toward $7 billion to $8 billion, but rejected forecasts of another Figma-like first-day bounce.

11. Andreessen’s 72 seed deals are sourcing, not the economic product

Andreessen completed 72 seed investments against 27 for second-place Sequoia, making it a different game “ipso facto.” Rory described Andreessen as the successful quantity provider across stages and, he thought, the largest Silicon Valley capital raiser, with Insight excepted as a more later-stage comparison.
Whether the strategy works will not be decided by aggregate seed returns. Seed is “cheap milk in the supermarket”: a loss leader that attracts the rare outlier, after which Andreessen must invest perhaps $1 billion—as it did in the Databricks example—at a price that lets the winner carry the platform.

12. Consensus protects financing, but price still determines returns

Martin Casado’s thesis was that non-consensus alpha is dangerous early because follow-on capital becomes increasingly consensus-aligned. Rory agreed: OpenAI may have been non-consensus in 2016, but perhaps 90% of similarly contrarian bets failed, while consensus themes such as SaaS and public cloud generated investable waves over long periods.
Jason supplied the current concentration data: 10 deals consume 40% of venture capital, and investors who once funded B2B now focus almost exclusively on AI. His advice to strong B2B-plus-AI founders outside that center is stark: “Don’t expect any money”; roughly 80% of his usual referrals may not even take the meeting.
Harry’s investment committee had just seen a fintech company reach $5 million of ARR in a year, yet interest stalled because “it’s not AI.” Rory’s answer was not to reject it, but to buy at a fundamentals-based price and operate capital-efficiently because the next round will not arrive with “magic pixie dust.”
The crucial distinction is technical consensus versus valuation consensus. Agentic software may be the industry’s direction for 20 years, making a technical bet against it unwise; that does not excuse paying “beyond the dreams of man.” Consensus can remove one axis of risk while simultaneously encouraging catastrophic overpayment.

13. The missed AI call was underestimating capital formation itself

Looking back, Rory wished he had made more on-trend AI investments while also regretting three or four non-consensus deals he declined. The corrective is remembering “the 90 non-consensus bets” that disappeared, while consensus companies survived long enough to learn because other investors’ capital buoyed them.
His largest analytical miss combined scaling laws with Sam Altman’s ability to inspire belief in them and unlock roughly $600 billion of annual CapEx. That wall of money benefited foundation models, Nvidia, inference providers, and nearly anything attached to “making AI.”
Had he foreseen that financing capacity, Rory would have “broken glass” on conventional financial models to own more model and inference exposure even at prices that looked high. His final self-correction preserved the ambiguity: perhaps scaling laws were not consensus when the decisive investments had to be made—which is precisely why investing cannot be reduced to a consensus-versus-contrarian slogan.

Can Rate Cuts Save Crypto’s Bull Market?Untangling Neural Network Mechanisms: Goodfire's Lee Sharkey on Parameter-based Interpretability