Anthropic's $10B Round, Klarna's IPO, Inside a16z's 72 Deal Seed Investment Machine ft. Marc Benioff

Summary

• Marc Benioff opened by torching the AGI narrative: “we have all been sold a lot of hypnosis around what’s about to happen with AI.” His mechanism — LLMs are “a finite set of algorithms,” only incrementally better over five years, trained on “a relatively finite set of data” — and his warning is doctors already “giving their patients bad advice and becoming intellectually lazy” from over-reliance. Yet the same man makes the categorical call: “I don’t think that there will be a piece of software that we sell that will not be agentic.”
• Salesforce as customer zero is the tradeable proof point: agentic support cut human agents from ~9,000 to ~5,000 (redeployed, not fired), and agentic sales is now calling back the 100 million leads Salesforce never had the people to call over 26 years. Data Cloud plus AI has crossed $1B in revenue — “our fastest growing cloud product ever in 26 years” — and Benioff put Snowflake, Databricks and Palantir Foundry (all $3-4B revenue) “in my sights.”
• Benioff on Palantir is envy plus adoption: “Have you seen their price list?… I’m like, whoa, my prices are too low” — he beat them on a huge US Army contract, openly covets the “hundred times revenue multiple,” and says the forward-deployed-engineer model of building before the deal is signed is something “we can all embrace and adopt.”
• The Meta reorg — Nat Friedman reporting to Alex Wang amid a hiring freeze within 30 days — split the panel: the structure is sensible (“you spent $20 billion on talent, you now need to tell them what position to play”), but the puzzle is why a would-be Microsoft CEO gave up autonomy. Jason’s verdict: “there’s only so many rooms I need to be in — I might rather have more carry than show up to ringing the bell.”
• The episode’s sharpest math caps the AI trade: Anthropic went 1→9B in revenue, so either it lands in the tens of billions next year (unprecedented) or “they slow down faster than anything slowed down ever.” It all reduces to whether foundation-model API demand is $50B or $500B — and Rory’s back-of-envelope says even Salesforce turning on an AI SDR across its $12B Sales Cloud drips only ~$720M to LLMs: “you have to sell a lot of labor replacement to get to 100 billion.” Jason’s counter: he pays ~$12K/year for Salesforce and nominally $500K for 11 AI agents — a 20x ratio that, if it holds, makes the round cheap.
• On the pricey tape: Mag 7 concentration is at all-time highs and Rory does believe in reversion to the mean — 2022 saw cloud valuations fall 66% even as growth persisted — “something can be amazing and still overpriced.” But the cavalry arrived the same week: [company name unclear] up 27%, Box, Okta and even Zoom reaccelerating on AI, and “reacceleration at scale is always epic.”
• Klarna filing at $13-15B off SoftBank’s $45B round illustrates the “hard deck”: below ~20% growth it can be hard to file, and hyping $1M revenue per employee is “a coded message… we’re mature.” SoftBank doesn’t get washed — “they just do what’s called losing money” — and Rory’s epigram carries the round: “venture is a game played by 6,000 people and in the end Sequoia wins.”
• Martin Casado’s tweet — that non-consensus early-stage investing is dangerous because follow-on capital is consensus-aligned — got endorsed roughly 70/30: betting against the megatrend is “like doing a client-server deal in 2002.” Rory’s confessed miss: underestimating scaling laws and Altman’s ability to “unlock $600 billion of capex spend a year” — the one trend “you just almost could not have had too much on.”

Deep dive

Benioff Rejects AGI Hype

• Asked about Amazon’s AGI head saying only a thousand AI engineers matter, Benioff went straight at the premise: “AGI head — that sounds like an oxymoron. You’re talking to somebody who is extremely suspect if anybody uses those initials.” His mechanism: LLMs are “a finite set of algorithms” — better, “but incrementally better over the last five years” — applied to “a relatively finite set of data that has come off the internet.” Feeling intelligent isn’t being intelligent: “it kind of felt that way when I was using Eliza when I was like 16 years old on my TRS-80 Model 1.”
• The line to keep verbatim: “it’s not a person and it’s not intelligent and it’s not conscious and it doesn’t have a childhood and it hasn’t suffered.” Not that AGI couldn’t happen one day — “we’ve all seen those movies” — but “that isn’t the state of technology today.”
• His warning case, from articles he’d tweeted about: doctors over-reliant on inaccurate AI “are giving their patients bad advice and becoming intellectually lazy at the same time. I think that is a huge warning sign for all of us.”
• On the talent frenzy — billion-dollar offers to his peers “with bluntly very little to show for it… other than the team” — Benioff was flat: “No. And we’re not.” Salesforce is focused on “what is the next generation of the enterprise,” where “tactics must dictate strategy over time.”

Salesforce Reallocates Support Headcount

• help.salesforce.com is Salesforce’s agentic support layer, with an “omni-channel supervisor” arbitrating between human and digital agents. Result: human support agents down from ~9,000 to ~5,000, with the headcount rebalanced “into other parts of my company where I need more help… because we’re still growing.”
• The story Benioff broke on the show: over 26 years, “more than 100 million people contact us that we’ve not been able to call back. We just have not had the people.” Agentic sales is now calling all of them and feeding a new agentic sales product debuting at Dreamforce — against a base of 15,000 salespeople.
• The categorical call, which Rory framed as the 2025 equivalent of “never on-prem again” in 2000: “I don’t think that there will be a piece of software that we sell that will not be agentic.” And the deeper point Benioff seized from Jason: “the fundamental architecture of an enterprise software company in the future is not exactly as it was in the past” — the apps changed over 25 years of SaaS; now the companies themselves do.

Data Cloud Challenges Palantir

• Pushed by Harry on AI not showing up in the numbers (“it’s so untrue”): Data Cloud plus AI is now more than $1B in revenue, “our fastest growing cloud product ever in 26 years,” from a product not shipped until last November. The Agentforce agent on Salesforce’s own website “has done as many customer interactions as our support agent” — because “we put our whole website into our data cloud.”
• Rory’s defense of Benioff on the growth charge: at $40B, 10% growth adds $4B — “an entire Palantir every year” — and the nine figures of revenue added on AI deals last quarter is “a $400 million AI-only startup, which would be freaking amazing if we all owned it.” People expecting AI to transform $100B market caps in a week “are just way overestimating what it takes.”
• On the $3-4B-revenue data clouds: “They’re in my sights… I am like the guy in Star Wars in my TIE fighter — stay on target.” Palantir’s growth is “very cool and amazing and very inspiring,” but its price list stunned him: “I’m like, whoa, my prices are too low. I’m automating the whole VA at this price. What would they be charging?” The US federal government is already Salesforce’s largest customer, and “we just won a huge US Army contract — we beat Palantir.”
• On forward-deployed engineers, an honest both-and: Salesforce has always had systems engineers and professional services in the customer, “but we don’t have that branding” of building the product before the deal is signed — “I think that idea is very cool… something that we can all embrace and adopt.”

Benioff Defends SaaS Apps

• On unnamed executives — “great people actually, and great executives” — saying SaaS apps become CRUD databases: “if you really think that, wow, you are really wrong. That is crazy talk… I need apps and I need agents and I need them to work together.” His kicker: “Why does Microsoft have 3% CRM market share? Because of nonsense.”
• Harry’s sharper framing — discard the replacement talk entirely, but “assume Salesforce is the infrastructure”: maybe the rep gets a better front-end tool, or an agent not owned by Salesforce does the work and coordinates on the back end. Who owns that real estate? Benioff’s answer is a three-layer stack — apps stay in the flow of work, data underneath, and “an agentic layer that’s going to interoperate with those applications and that data” — open, ecosystem-fueled, “and I hope that it’s going to be built on Salesforce.”

Salesforce Reassigns SDR Headcount

• Jason — “basically the grim reaper” on this show, per Harry — has argued a mass exodus of the SDR class in 12-24 months. Benioff’s counter is redeployment into segments Salesforce couldn’t serve before, and Jason half-converted himself: “I bet you redeploy 70% of that headcount into enterprise reps or forward-deployed engineers — that headcount just becomes more valued with Agentforce sales.”
• Harry’s mechanism for why this works at scale: “Mark’s budget’s fixed — he’s got 80,000 heads on a spreadsheet… if you can move those heads up the value chain, Salesforce can be a much more efficient company.” And Benioff’s relentlessly additive framing — “an order of magnitude more SMBs because SMBs can do more than ever” — is, per Rory, “entirely the only way you’re going to sell this AI revolution, otherwise there’ll be another freaking revolution.”
• Harry stayed skeptical: “I think it’s grossly overly optimistic to think that you can redeploy 25-year-olds” at entry level. Benioff’s rebuttal was blunt: “this narrative that we’re not going to hire any more kids out of college — this is also [expletive]. You can go to our website and see who we’re hiring.”
• Harry’s unfair closer — OpenAI at 300 or Anthropic at 170, which would you buy? — got a masterclass dodge: both “great companies,” Anthropic “very focused on the enterprise,” and the tell: “Salesforce owns 1% of Anthropic.” Harry: “I don’t know what you’re paying your media training person, but you should pay them more.”

Meta Reorganizes AI Leadership

• Jason’s read on Nat Friedman reporting to Alex Wang: the consensus two weeks ago was it’s fine to give up “billions of potential carry” to be in the game — “but then essentially getting undermoted in a reorg… hiring freeze in a total reorg within 30 days, it’s a lot to process. I might rather be running my own fund.”
• Jason found the structure itself sensible — one person in charge, four divisions (science, foundation models, applications, infrastructure) — via a soccer analogy worth keeping: “you spent $20 billion on talent, you now need to tell them what position to play and who’s going to play forward, who’s going to play striker.”
• Harry’s confusion, made precise by Jason: not the structure but why someone “in the grooming position to be the next CEO of Microsoft” would report to someone who reports to Zuck. Jason’s only rationale: “Elon goes to Zuck when he wants to buy OpenAI… it’s pretty cool being in that room.” Jason’s rebuttal: “there’s only so many rooms I need to be in… I might rather have more carry than show up to ringing the bell” — though his standing advice to great operators is don’t do venture: “your highest and best use is operating.”
• On Meta down 6%: the core business is extraordinary but has “a very tenuous link” to the AI initiative, so valuing the stock is “like Kremlinology… who lines up in Red Square.” Until Zuckerberg explains the $60-70B — Instagram/WhatsApp-right or metaverse-wrong — “we’re guessing.” And with Meta’s beta at 1.59 (Nvidia 2.3), “I don’t think we can read anything into these ups and downs.”

Anthropic Tests AI Demand

• On the 5-to-10-billion raise, reportedly 4x oversubscribed: demand is “pretty damn high” partly because public-market AI purity doesn’t exist — Apple has nothing, Amazon little — so “if you’re a Fidelity-type manager… there’s two obvious at-scale candidates.” Jason’s GP logic on ICONIQ leading after Lightspeed: “it’s the same amount of risk… I lose two billion, six billion — what’s the difference? I can make so much more money.”
• Rory’s trilemma, run out loud: revenue went 1→9B in a year, so either it lands in the tens of billions next year — “unprecedented because the amount of revenue would just be so big” — or “they slow down faster than anything slowed down ever.” Even growth halving still gets you to a trajectory that supports the price. “I always joke that Newton’s law of motion applies to companies.”
• The reduction: “it boils down to your assessment of is there $50 billion of demand for foundation model APIs or $500 billion. If it’s the latter, they’re probably going to get 40% of it… if it’s the former… a lot of these people are going to be sad.” Rory’s own guess, hedged as stated: “My guess is no. And it slows more than you think. But it’s not a crazy call.”
• His supporting math: agents must be “worth 20, 30, $40,000 almost ahead to the enterprise for the math to work” — “if all it is is $2,000 an engineer, I don’t know if you get there… I just run the math and I can’t find the town. But I could be wrong.”

AI SDR Economics Constrain TAM

• Rory’s worked example against himself: say Dreamforce’s AI SDR is a 30% uplift on the ~$12B Sales Cloud — $3.6B of new revenue — at a rich 20% LLM cost, that’s $720M to the model layer. “You’ve just had the second largest software company on the planet turn on the most labor-saving device for their core marquee product and… round it up to a billion. You have to sell a lot of labor replacement to get to 100 billion.”
• Jason’s live counter from his own P&L: four Salesforce seats at ~$12K a year versus “nominally, nominally $500,000 for 11 AI agents” — a 20x-plus ratio. “I don’t know whether that makes sense long-term… but if a portion of that ratio were to hold, then it’s a pretty cheap round.”
• The catch both conceded: even if the spend comes, “Salesforce may not capture that incremental 120 billion. Workday may not capture it. Palantir appears to be capturing it… the big guys mostly don’t seem to be capturing this agent dollar” — which was exactly Harry’s original point to Benioff.

AI Reaccelerates Public SaaS

• Harry’s confession on Mag 7 concentration — “I really hope there’s not a blip here, dear Lord” — got Rory’s translation: “I don’t have the stomach to sell, crystallize my gains… I’m just going to let it ride and pray a little. It’s what I’m doing too.” But Rory does believe in reversion: he’s eyeing a core-commodities ETF — “the only things that survive the 70s” — as a 5% play, and his frame is “something can be amazing and still overpriced.” Memory check: 2022, cloud companies still growing, “the valuations fell 66%. It was brutal.” His timing answer is honest: “How the hell would I know?… you’ll know when it’s happened cuz it’ll hurt.”
• Same week, the cavalry: [company name unclear] up 27% on amazing numbers, Box and struggling Okta up, even Zoom reaccelerating on AI. Jason: “thank God… the cavalry is coming just in time” — and “reacceleration at scale is always epic… because it’s so rare.” Harry’s mechanics: stocks move on expectation gaps — at 5.5x revenue with the “SaaS is dead” story priced in, beating by a couple of points bounces hard; “when all the good news is priced in… you fall fast.”
• Durability evidence from the vibe-coding boom: Lovable and Replit at “three or four hundred million” of ARR are spooling up Neon and Supabase databases at loads “they’ve never seen” — “Neon got bought by Databricks for a billion. I didn’t even understand why at the time; now I get it.” And Wix bought eight-person Base44 for $80M — now reportedly doing $1.2M a week — “deal of the century.” The meta-point: it’s “heartening” to see [company name unclear] benefiting, “because it means maybe the revenue is a little more durable” — though Harry’s instinct stands: “all this revenue feels fragile.” Rory: “No, it feels very durable, thank you very much.”

Klarna Tests IPO Growth

• Klarna filing at $13-15B, down from SoftBank’s $45B round and Sequoia’s ~$6.5B reprice, with growth decelerating 24%→20%, is flying at what Harry calls the hard deck: “there is a level of growth below which it’s hard to file… it’s 20%.” Rory’s caveat: the threshold scales — at $10B revenue “they’ll happily take you public with a 7% growth rate” — and at $14B “it’s a perfectly doable deal.”
• The tell in the marketing: hyping $1M revenue per employee as growth decelerates “is implicitly saying we’re finding our rule of 40 in the bottom line… We’re mature.” Rory’s blunter version: “this is the classic fintech company trying to make software noises. Let me give you a clue: you’re a fintech company… you’ll get the medium-growth fintech valuation and everyone with the last round will make money.”
• Does SoftBank get washed? No — the 6B down round took only ~10% dilution and “preserved [the investment] by keeping the company alive.” Rory guesses no block (“Sequoia are not dumb people”), so SoftBank will “trade at 30-40 cents of what they originally paid… they just do what’s called losing money.” His epigram, via Gary Lineker: “soccer is a game played by 22 people and in the end the Germans win. Venture is a game played by 6,000 people and in the end Sequoia wins.” Jason pushed back that SoftBank did have a ratchet in a similar deal (WeWork) around the same valuation — “it is knowable… we will feed the S-1 into ChatGPT and we’ll know in an hour.”
• Netskope, by contrast: $700M ARR reaccelerating 30→33% (“may sound modest, but it’s a lot of work”), last priced ~$7.4B in 2021 — Rory’s gut is it walks up to trade “7–8-ish” — roughly a flat round, and no Figma bounce, “the largest bounce of any large-cap IPO since I think 2000… I believe in reversion to the mean.”

Consensus Bets Attract Capital

• Martin Casado’s tweet — non-consensus investing at early stage is “actually quite dangerous” because “follow-on capital tends to be more and more consensus-aligned” — was, per Rory, “more spot-on than people give credit for.” His frame: roughly 70% consensus megatrend, 30% brave-new-world (an IVP concept from 20 years ago); in 2016 the non-consensus bet was OpenAI, “but probably 90% of non-consensus bets would have failed entirely.” And betting against the technical consensus — that “most software is going to be agentic for the next 20 years” — is “like doing a client-server deal in 2002.”
• Rory’s practitioner corollary, with 10 deals now consuming 40% of venture capital: he’s done several B2B-plus-AI deals he loves, “and the advice I give to all those founders is don’t expect any money… 80% of the folks I can refer you to are not going to take your meeting.” Harry’s live example: an IC that day on a fintech at $5M ARR in a year with a great founder, stalled because “it’s not AI” — Rory’s rule: do it, “at the right price, because you’re not going to get the magic pixie dust next round,” and run it capital-efficiently.
• The same logic explains Andre’s 72 seed deals versus 27 for the number-two mega-fund: “by definition it’s a different game.” Rory: whether Andre works out “won’t be because of their seed program” — it’s “basically like cheap milk in the supermarket, it brings in the crowds,” a loss leader that pays if they stuff a billion into the few outrageous outliers at the right price, as at Databricks.
• Harry’s synthesis and Rory’s confessed miss close the show: the consensus bet means “you’re probably right on direction; you might ludicrously overpay. The non-consensus bet, you could be way wrong on is it even going to work… but if you get it right you’ll have a high-ownership, low-capital, N-of-one outcome.” Rory’s regret: underestimating scaling laws and “the ability of primarily Altman… to inspire belief in those scaling laws and unlock $600 billion of capex spend a year” — “you’d have broken glass on your financial model” to get exposure, “the trend that you just almost could not have had too much on in the last year.” As always: “investing is hard and you can’t just paint the numbers and collect 100 million bucks.”