Jason Lemkin | 20VC Summary

1. Key Themes

Price discovery on the supply side of AI, not a verdict on AI demand

Rory frames the OpenAI down-round as pure supply-side mechanics: a liquid, larger, profitable competitor is about to list, so the illiquid, smaller, loss-making one has to price below it. The real signal to watch is demand. Rory: "I would dispense with the whole AI wave cracking narrative... the other company is bigger, growing faster, profitable, and about to go public. So you can buy it liquid at 1.5 to 2 trillion. What would you pay for the smaller company growing less quickly, still losing money, that's illiquid? Answer: less than 1.5 trillion." [00:04:55]. On how to read the stress in the debt markets: "If the demand for AI is as strong as people think, then everyone's going to need every damn computer, GPU, and data center they want. And those bonds will be money-good because someone will buy the damn company" [00:06:50]. Jason counters that the wobbliness is real: "CoreWeave debt is paying 15%. That's wobbliness... Broadcom having to guarantee Anthropic's debt is a little wobbly" [00:05:36].

The Anthropic growth-rate debate: same facts, two readings

Rory lays out the GAAP numbers: "Roughly 5 billion in Q1, roughly 11 billion in Q2, roughly 17 billion in Q3" [00:11:06]. The positive read is roughly 50% quarter-on-quarter, "which is still 5x year-on-year if you compound it." The negative read is that growth fell from 100% to 50% in one quarter. He adds that "they're probably going to do 56 billion in GAAP revenue this year... a 12x" growth rate, with about 120 billion next year, "from 12x growth rate to 2x growth rate in a year" [00:15:29]. Jason's thesis is pricing pressure: "OpenAI is basically running around selling somewhat equivalent products for 30% off" [00:10:24]. Harry notes enterprise penetration is tiny: "they're 0.1% of Fortune 500 spend" [00:14:04].

Real-world token deflation is finally showing up

Jason offers a first-hand data point: "When I switched it myself to Haiku 55 and Jev, total costs went down by 70%, 72%" for the same task at higher quality [00:16:30]. He adds that "nothing got cheaper until Haiku 55 plus Jev" [00:16:30]. Rory's historical frame is that digital goods fall 50-60% a year in cost: "The bet you take over and over again is digital goods get reduce in price 50, 60% year-on-year. If people like the shit and they get hooked on it, over 2 or 3 years, the math gets better" [00:23:07].

Launch before the economics work, or lose to someone who does

Rory argues venture capital makes waiting irrational: "You can't wait until the economics work perfectly to launch a product cuz someone else, armed with venture dollars, will launch that product 12 months earlier when the economics don't quite work, grab all the mind share, and then they'll basically lean into the cost curve" [00:25:11]. "Almost everything that's new and digital will look crazy for a while in terms of economics. And the only question is, can the economics get better before the money runs out?" [00:25:11]. Jason's warning sign: "Elon just had to cap it [Grokbot]... the era of these free subsidies for the agents are coming to an end" [00:26:29].

Established SaaS and consumer web brands are eroding faster than expected

Jason churned Canva after 12 years: "It's just you just don't need it. The agents don't need it" [00:18:31]. He is surprised at how far it extends: "The real world isn't buying much of this crap anymore... Canva, a Squarespace, which just launched 20% layoffs, a Wix... Shopify" [00:20:50]. Rory's view of the endgame for non-adapters: "If you decide to just not play that game, I think over the medium term, you're just heading to quiet obsolescence" [00:19:42].

Agents as a disintermediation threat to aggregators

The delivery and ride-hail aggregators face agentic risk, but of different severity. Jason on DoorDash: "I think agents are an existential risk to DoorDash because their number three competitor is nobody" [00:31:25], arguing an agent can route you to the restaurant's own delivery or pickup. Harry thinks ride-hailing is protected because a conversational interface is "a worse user experience" [00:32:37], and that food is about discovery.

Customer-driven M&A: the SVP "punch card" and the $100M ARR threshold

Jason describes how large acquirers like Salesforce run deal flow: "You go around the table with all your SVPs, and they get their best idea... None of these deals are Marc betting the future of the company. They're SVP chips" [01:02:06]. On scale: "If they want to do an M&A where they want to put their resources against it, their GTM machine, their platform, they'll all tell you this, they want at least 100 million of revenue" [01:04:53]. Salesforce's Cresta ($2.7B, 10x on 200M+ ARR) contrasts with Listen Labs ($2B on about $20M ARR).

Robotics foundation models and the TAM-doubling logic for the labs

Rory on Physical Intelligence at $20B: "Technology is amazing. The business model has got a lot of developing to do... There are 81 million knowledge workers in the U.S. We have an LLM, and we sell to them... [then] a robotics foundation model, we can address [physical workers]. I've doubled my TAM" [00:57:56]. He expects acquisitions: "I think all the big foundation models company buy one of these things."

Cyber risk is the under-covered story of the next 12 months

Jason on the ArtX pen-testing tool: "Some really smart developers just a couple months ago in China built an open source thing... it became very easy to turn it around and just attack other folks with agents" [00:39:23]. Rory: "You built a gun, and you hoped it was in good people's hands. And now, it's ended up in bad people's hands" [00:40:14]. His worry: "The entire U.S. Fed wire goes down for a day" [00:40:33].

Growth-stage selection frameworks: multiple on cash vs. dollars out the door

In the IC game, Harry picks Serval for upside multiple but admits fund economics push toward Cognition: "The multiple is way better here... they only raised 150 million. And if I'm Andreessen, I want to put 500 million, a billion out the door" [00:44:31]. Jason's constraint: "10 billion is the lowest possible exit where we're allowed to bring it to IC" [00:57:08].

2. Contrarian Perspectives

The OpenAI "crack" is a liquidity-pricing artifact, not an AI-cycle signal

Most commentators read the 1.4T to 1.25T cut as a verdict on AI. Rory rejects the framing entirely: "No projection about AI positively or negatively, just there's two securities that are going to be sold in the next 2 months. And which one do you want to buy?" [00:04:55]. His implication is that the real test is demand growth, and that even stressed credit like CoreWeave bonds will be "money-good" if demand holds.

Anthropic's IPO is at risk because pricing power has hit a "local maximum"

Despite 12x revenue growth, Jason argues the market will punish it: "I actually think the Anthropic IPO is going to suffer for it. I think there's evidence of substantial pricing pressure" [00:09:27]. His thesis is that "pricing pressure hit them in Q3... Maybe we've reached the point where Anthropic ability to command a price premium has hit a local maximum" [00:14:20]. Demand and market share (via OpenRouter) have not visibly declined; the problem is price.

Vertical AI is going to produce the biggest "tears" in 2027

Jason is avoiding the category altogether: "2027 will be the year where we find out a lot of vertical AI plays are nifty little startups, but didn't in the end expand TAM enough... you're just slicing up the same old TAM 11 times... I just view it as tears in five years" [00:48:06]. Rory pushes back that Serval is cross-functional and expanding the TAM, but the thesis that most vertical AI merely re-slices incumbent spend stands.

Subscription AI products at $30-50/month may be structurally unprofitable for persistent agents

Jason: "I think [persistent agents] could cost as much as 20 bucks a month per user to service... Then it's a moronic business model at $40 or $50 per user per month" [00:21:48]. He contrasts it with Replit and Lovable's tuned on/off VMs at "$4 to $5 a month." Harry agrees power users are "the worst" economic profile.

The human-plus-AI law firm model (Crosby) has a ceiling below IC thresholds

Jason on Crosby's bid for Weil's assets: "I don't think I can take this to IC. It's a lot that has to happen for the corporate company... to be worth $10 billion, because it has to suck so many billions of years of profits that the lawyers could otherwise take home as partners" [00:56:20]. Rory half-agrees: "if 70% of the dollars come from the high rainmaker, irreplaceable by AI stuff... you're left with the low-margin parts of the business" [00:56:33].

3. Companies Identified

Anthropic: Frontier AI lab, heading to IPO. Mentioned for explosive growth and as the central debate on growth deceleration vs. pricing pressure. Rory: "4.5 billion in GAAP revenue last year full year, exiting run rate of 9 billion. And they're probably going to do 56 billion in GAAP revenue this year" [00:15:29]. Harry noted FTX's stake: "FTX's $500 million position from Anthropic would be worth 39 billion at the 1 1/2 trillion" [00:37:59]. Jason, on coding: "The one thing that is working that never stops working is coding agents" [00:41:41].

OpenAI: Frontier AI lab raising at $1.25T (down from $1.4T), $20B round (cut from $30B) with only ~$11B committed. Jason: "OpenAI is basically running around selling somewhat equivalent products for 30% off" [00:10:24]. Rory describes it as "smaller, growing less quickly, still losing money, that's illiquid" relative to the competitor [00:04:55].

Alphabet / Google: Fell ~12% in a decade-worst drop. Google Cloud "doubled to like 130 billion, growing 103%," with Search slowing to 12% and negative free cash flow (Harry [00:08:40]). Rory: "These guys are saying they can sell compute as fast as they can make it. And that's the number one thing, and that's where the growth's coming from" [00:08:40].

CoreWeave: Neocloud and GPU data-center provider. Jason flags "CoreWeave debt is paying 15%" [00:06:50]. Rory: it "is going to be the first person who's going to suffer when the demand is not there" [00:06:50].

Broadcom: Semiconductor and infrastructure company. Jason cites it guaranteeing Anthropic's debt as "a little wobbly" [00:06:50].

DoorDash: Delivery platform buying Instacart for $17B (under half its 2021 private price). Jason: "If Instacart's got 20% of the market and DoorDash has 60... you can get to 80% market share, this is how you dominate networks" [00:27:46]. Notable disclosure: AI agents have sent "only several hundred orders against 3 million deliveries a day" (Harry [00:29:34]).

Instacart: Grocery delivery platform; DoorDash target. D1 owns 9.7%, Sequoia 8.1% (Harry [00:27:37]). Rory: "They clearly tried to do it before. It clearly makes sense. There's also probably an element of do it now while you can" [00:27:37].

Uber / Uber Eats: Harry: "I'm with Bill Ackman, I think Uber is a banger of a buy right now... I think it's so cheap" [00:32:17]. Rory: the two risks are "Waymo disruption and agentic disruption" [00:32:22]. Uber's earlier Postmates acquisition worked better than Jason expected [00:28:12].

Kalshi: Prediction-market and sports-event exchange raising $1B at $40B (2.5% dilution), up from ~$22B 3-4 months earlier. Rory: "Kalshi is a money deal. The thing is working. It's growing like a weed... we've long since converted into sports gambling markets" [00:33:28]. Utah has charged founders.

Cohere: Canadian LLM company; Canadian government investing up to ~$1.14B. Rory: "give Cohere huge credit. They are one of the survivors of that first group of LLMs" [00:33:28], but notes it is a sovereign-AI play, not a momentum play. Jason: "Cohere not disclosing revenue is hardly a flex" [00:36:44].

Mistral: European LLM company cited as the parallel "you're not America" option (Rory [00:36:44]).

Sequoia: Led Kalshi's last round "right after felony charges were against the founders" (Jason [00:37:59]), and invested in both Listen Labs and Cresta (Rory [01:04:33]).

Binance: Crypto exchange; ~$6.6B reportedly stolen, restored by the company. Rory: "The most compelling fact about the Binance thing was that they could pony up $6 billion to replace the money that was taken... what a good business" [00:38:11]. Jason doubts they had $13B on hand [00:38:53].

FTX: Cited as a counterfactual: "same VC, same risks... would FTX still be around today?" (Jason [00:37:59]).

Cognition: Coding agent (Devin, Windsurf team), raising at $70-80B vs. $48B in September. Jason: "Cognition at 80 billion, it seems market correct... They lucked out buying the Windsurf team. They lucked out buying IP. They lucked out from open weights" [00:41:41]. Rory: "coming into large enterprises and basically saying, 'We will help you effectively rewrite your entire code base'... just an ability to command way more dollars" [00:46:16]. Harry: "you're fourth in the market. You got Codex, you got Claude Code, you got Cursor, and then you got them" [00:51:10].

Serval: AI ITSM/service-management app at $3B (up from $1B in December), raising ~$150M. Harry: "Jake and Tatiana at Serval are incredible. The product's amazing. The enterprise adoption is better than I've ever seen" [00:44:37]. Rory: "If there is a next-generation ServiceNow, the probability of it being Serval is probably 80, 70% plus" [00:50:07].

ServiceNow: Incumbent in ITSM, worth ~$110B (Harry [00:50:52]). Bill McDermott leads it. Rory: "Until recently ServiceNow was, I think, the second-largest market cap SaaS company" [00:46:16].

Baseten: Inference provider at $30B, up 2x from June. Rory: "If you want revenue velocity right now, nothing gives you revenue velocity like inference" [00:00:09]. Harry would rather back Modal or Fireworks.

Modal: Inference/compute platform. Harry: "Modal's product suite is unbelievably strong. Like, Erik is incredible for that" [00:49:45].

Fireworks: Inference provider. Harry: "Lynn and the team at Fireworks, by far, I think have the best inference product" [00:49:45].

Cursor: Coding agent. Rory: "I'd take the 60 billion and call it a win myself" [00:24:00]. Jason claims it "would have died" selling Anthropic tokens at a loss, and refers to a SpaceX acquisition at $60B [00:42:54].

Replit / Lovable: Jason: tuned VMs reach "about $4 to $5 a month" [00:22:05]. Replit CEO Amjad said "the era of these free subsidies for the agents are coming to an end" [00:24:06].

Muse: Meta's AI product. Jason: "Muse can afford it for now, but what's the scale of the Muse subsidy" [00:24:06]. Rory: a $10/month subsidy is trivial to someone "spending $70 billion on CapEx" [00:25:09].

Grok / xAI: Jason notes Elon "had to cut back Grokbot use because it's too expensive" [00:21:48]. Rory suggests he prefers selling compute to Anthropic for "50 bucks an hour" [00:27:01].

Automattic (WordPress): Jason: "I think Automattic will stay private forever. I don't think it'll ever go public, and I don't think it'll ever get acquired" [00:19:22].

Canva: Jason churned after 12 years; reports vary on growth (zero to 20%) [00:19:02].

Shopify: "just fell 15%... because they came up a smidge short, even though they raised guidance" (Jason [00:09:27]).

Squarespace / Wix: Squarespace "just launched 20% layoffs, a Wix, which is going to be bailed out by Base44" (Jason [00:21:46]).

Salesforce: On an M&A spree: "Salesforce is on an epic M&A roll. I mean, they've hoovered up everybody this year" (Jason [01:00:17]). Bought Cresta for $2.7B (Harry [01:01:22]) and earlier Braze, after "losing to Braze in marketing for years" (Jason [00:29:10]).

Cresta: Contact-center AI, acquired for $2.7B with 200M+ ARR (Harry [01:03:37]). Rory: "I think there's a smart deal for Salesforce. You get critical mass" [01:04:36].

Listen Labs: AI research startup, sold for ~$2B at ~$20M ARR (Harry [01:03:37]); Jason calls such deals "the dumbest bets" at big acquirers [01:05:51].

Intercom Fin: Customer service agent (sponsor). Jason noted Fin's agentic business "had to hit 100 million" to be acquirable [01:05:51].

Crosby: AI-native law firm, raised $85M from Sequoia, bid for Weil's assets. Rory: "That was chutzpah" [00:53:13]. Jason: "I don't know that the business model is sustainable... It's too human-based" [00:55:37].

Weil Gotshal: Elite law firm that collapsed from partner disputes. Rory: "If four or five key lawyers... the guy who ran all the PE practice, which is probably 20, 30% of a company like this, left, took his book of business" [00:52:40].

Physical Intelligence: Robotics foundation model, talks at $20B (Aveneer and Lux). Rory: "I've done a lot in robotics over 10-plus years. When you see what the new models can do, it is truly amazing" [00:57:56].

Generalist: Robotics company Rory's firm holds a small investment in [00:58:00].

Hugging Face: $12.9B deal drawing DOJ scrutiny (Jason [01:00:17]).

Creandum: Harry: "probably the most exciting fund that Europe's produced in the last three years, raises 750" [00:57:33].

Dropbox: 9% staff cut, CTO leaves. Rory: "Company with billions of revenue ain't worth a lot... revenue multiples don't get you anywhere" [00:00:09].

Factory: Coding-agent startup Harry has invested in [00:51:29].

Base44: AI app builder (sponsor); Jason noted it would "bail out" Wix [00:21:46].

AlphaSense: Research platform (sponsor), launching SuperAnalyst agent [00:02:05].

WorkOS: Breached via the ArtX tool (Jason [00:39:23]).

ArtX: Chinese open-source pen-test tool weaponized for attacks (Jason [00:40:14]).

Atlassian: Cited as a TAM slice vertical AI startups chase (Jason [00:48:32]).

Higgsfield / Opus / Haiku 55 / Jev: Tools and models Jason uses in his agent stack [00:21:46].

Vinted: Cited as exposed to agentic disintermediation (Harry [00:30:29]).

Lyft / Free Now: Alternatives an agent could price-optimize across (Harry [00:30:29]).

Waymo: Autonomous-ride competitor to Uber (Rory [00:32:31]).

OpenRouter: Cited by Jason for market-share data showing demand hasn't declined [00:14:20].

Slack: Harry's benchmark for growth: "0 to 10 in 18 months was like the gold standard" [00:16:23].

Facebook: Rory's analogy for early server-cost angst that proved overblown [00:23:40].

YouTube: Harry: "one of the best businesses in the world" despite early economics [00:23:42].

Gamma, Asana, crypto.com: Named Fin customers in the sponsor read [00:03:19].

4. People Identified

Brian Chesky (Airbnb CEO): Jason compares the DoorDash agent comment to Chesky's past remark that "there's no need for AI for rentals and housing" [00:29:34].

Parker Conrad (Rippling CEO): Jason: "another generational CEO, said there's no need for AI in B2B 18 months ago" [00:29:34]. Both are cited as examples of leaders dismissing AI disruption.

Tony Xu (DoorDash CEO, referred to as "Tony"): Reported that the Instacart deal had been discussed for years (Jason [00:29:10]).

Matt Mullenweg (Automattic): Jason: "I think Automattic will stay private... and I think Matt is fine with it" [00:19:22].

Amjad Masad (Replit CEO): Said "the era of these free subsidies for the agents are coming to an end" (Jason [00:24:52]).

Sarah Guo (Conviction): Jason: "I wouldn't mind being Sarah Guo and investing in Instinct at 50 or 100" [00:24:06].

Jake and Tatiana (Serval founders): Harry: "Jake and Tatiana at Serval are incredible" [00:44:37].

Bill McDermott (ServiceNow CEO): Harry: "an old guard... a sales-led CEO" [00:44:37].

Mark (Uber president): Harry: "Very clearly was most worried about agentic usage of these platforms because it would disintermediate Uber" [00:29:59].

Bill Ackman: Harry cites him on Uber as a buy [00:32:22].

Elon Musk: Capped Grokbot use due to cost; Jason: "He subsidizes everything round-trip... This is a man that would not want to cap Grokbot if he did not have to" [00:26:29].

CZ (Changpeng Zhao, Binance): "Luckily they, he got pardoned" (Jason [00:37:59]).

Sam Bankman-Fried ("Sam"): Jason: "That's why Sam's in jail and CZ is out" [00:38:33].

Stewart Butterfield (Slack): Harry's growth benchmark reference [00:16:23].

Marc Benioff (Salesforce): Jason: "Bless Benioff and friends, right? We need these deals" [01:01:22].

Lynn (Fireworks) and Erik (Modal): Harry praises both teams' inference products [00:49:45].

Cliff (Cliff Obrecht, Canva): Harry notes he messages saying hosts are harsh on Canva [00:20:48].

Michael Grinich (WorkOS CEO): Posted about the ArtX breach (Harry [00:40:14]).

Ryan (Crosby): Harry: "Ryan just like, 'Yeah, we'll take it'" on the Weil bid [00:53:16].

5. Operating Insights

Don't buy a scaled incumbent when you can only absorb a mature one: the $100M ARR acquirer threshold

Jason: if a large acquirer is going to put its GTM machine behind a deal, "they want at least 100 million of revenue... The Borg otherwise will just, they can't even assimilate you. You're too small" [01:04:53]. Founders hoping for strategic exits should model the buyer's materiality test, not just the multiple. Below that line, you are priced as a team/IP deal.

Use model-swaps as a cost lever for production agents

Jason cut costs 70-72% on a B2B matching workflow just by switching models: "When I switched it myself to Haiku 55 and Jev, total costs went down by 70%, 72%" [00:16:30]. Operators running agentic workflows should re-benchmark model choice every release cycle, since deflation arrives unevenly.

Run a subsidy-end-date model on any agent product

Jason's calculus on persistent-VM products: tuned on/off VMs get "$4 to $5 a month" versus persistent agents potentially "20 bucks a month per user" [00:22:05]. Replit's CEO signalled the subsidy era is ending. Rory's counter is that the bet only works "before the money runs out" [00:25:11]. Founders should model unit economics at unsubsidized token pricing while planning to ride the cost curve down.

Evaluate whether a challenger expands wallet or just displaces it

Rory's diligence test for Serval-style disruptors: compare ACV like-for-like. "Would that same company [spending 200 grand on ServiceNow] be willing to spend 300 or 400 grand on Serval? And that's what you need to see to get TAM expansion" [00:48:06]. It is a concrete metric for separating real expansion from price-matched replacement.

Prepare for agent-routed customers with a direct channel

Jason's point that an agent will route around aggregators to direct delivery or pickup ("Lemkin's Curry down the street, actually, it will deliver for $2") [00:31:24] implies marketplaces need loyalty, pricing, and direct-ordering paths before agents commoditize them. Merchants should build their own ordering rails now.

6. Overlooked Insights

The "Q-on-Q arithmetic trap": adding the same absolute dollars on a larger base is itself the deceleration

Rory's aside deserves more attention: "They added exactly the same number Q in Q3. But by definition, because they're starting at 11, now you're adding the same absolute number... but on double the revenue size. So the Q-on-Q growth rate has, quote, unquote, plummeted to roughly 50%" [00:11:06]. The implication is that "flat net-new revenue" is the true steady state for a maturing hypergrowth company, so the diagnostic question is whether net-new dollars (not growth rate) hold or fall. Jason's separate claim that Anthropic's net new bookings fell versus Q2 [00:06:50] is the harder number to explain, and it points to pricing, not demand.

A government-backed AI lab is a customer-acquisition problem no capital can solve

Rory's single line about Cohere has wide implications for sovereign-AI plays (Mistral, Canadian and EU programs): "Governments can make things happen for a fair while, but what they can't do is make the private sector use the product" [00:36:02]. Combined with Jason's observation that Cohere declined to disclose revenue ("hardly a flex"), the sovereign-AI cohort may be funded on political demand rather than commercial pull, which matters for anyone evaluating late-stage or secondary exposure to these names.