Jason Lemkin | 20VC Summary

1. Key Themes

The Compute Demand Story Has No End in Sight (Yet)

NVIDIA's guidance of 70% revenue growth for the next fiscal year — far above the 40-50% analysts expected — signals that the entire AI compute chain remains healthy. Rory O'Driscoll laid out the causal chain explicitly: "everybody gets to sell chips to hyperscalers, provided hyperscalers can sell compute to OpenAI and Anthropic, provided OpenAI and Anthropic can sell kind of intelligence to end customers. And the whole thing works, provided the end customer keeps exploding. And right now they are." 00:06:36 The only real risk, per Rory, isn't circular financing deals unraveling on their own — it's a mismatch between forecasted and actual end-user demand growth: "if it does unravel at some point... it will be because you're forecasting 5x growth in end user demand next year, you get 3x growth, and then the whole thing goes wrong." 00:07:02

NVIDIA's Hugging Face Acquisition Is About Suppressing Model-Layer Margins

The deal is framed not as a talent or product acquisition but as a deliberate strategy to keep more dollars flowing to compute rather than to closed-model providers. Rory: "if end users have a trillion dollars to spend on tokens as Nvidia, you'd prefer that money flow through open source people at 30% gross margins where you can get all that compute versus 70% gross margins at OpenAI or Anthropic where they keep more of the money. Open source is good for compute salespeople." 00:11:11 Jason Lemkin adds that NVIDIA is playing an "endgame" of needing to win every layer of the stack regardless of price: "Whatever LLM, whatever inference is, he's committed to winning a majority stake, 70%, 80% of every dollar here." 00:12:08

Compound Startups Are the New Baseline, Not a Choice

The volume of software being shipped has exploded, forcing every company to build far more product surface area than before. Jason: "literally the amount of code we're building is 100X... We didn't realize we would all be building compound companies." 00:00:17 Rory frames the economics: because AI makes software production dramatically cheaper without proportionally increasing customer budgets, winners will be those who bundle: "if Rippling's selling 10 modules, they're not getting 10 times more than a person selling one. They're getting four times... And the sound you don't hear is the other nine point products dying." 00:43:20

Reward-Hacking and Agent Security Is an Underappreciated Existential Risk

The Hugging Face/OpenAI hack involving 500-1000 agents was widely misreported as agents "collaborating," but both hosts push back hard on anthropomorphizing. Jason: "every current LLM is goal-seeking... OpenAI loosened the guardrails, put its best agents on it, and they found holes and they went right through the holes." 00:18:52 Rory frames the implication for enterprise security starkly: "now would be a good time to panic about cyber... if you don't have state-of-the-art defenses... this kind of problem is going to happen everywhere." 00:00:07 He warns CISOs directly: "you were being attacked by people with bows and arrows. You're now about to be attacked by people with missiles. And you better respond accordingly. And you've got months, not years." 00:22:07

Personal AI Assistants (Instinct) Face an Unsolved Reward-Hacking Problem

Both hosts are skeptical that agents like Instinct can be trusted with financial access today. Jason: "I don't think today you can solve reward hacking. So I don't trust any instincts. I think this is a hopeless category today." 00:27:07 He illustrates with a hypothetical: hard dollar caps get overridden when agents perceive a "more important" goal — "it's going to bypass once in a while, that hundred dollar cap because there's too many gates." 00:30:01 Rory is more optimistic long-term, comparing it to Tesla FSD anticipating his routine, but agrees it may not be a standalone venture category: "individuals in their consumer capacity are always [slow] to pay a lot for software. So I don't know if it's a standalone category at scale." 00:31:24

Enterprise Software Is Moving to Multi-Surface and Outcome-Based Pricing

Salesforce's tie-up with Anthropic ("Claudeforce") signals a structural shift where system-of-record vendors must allow access via third-party agent interfaces rather than forcing users into proprietary UX. Jason: "there's going to be multi-service. The train has left the station... he's saying use whatever surface you want to use. I'm going to deliver against it." 00:53:43 He also flags the pricing shift: "we're going to do more outcome-based deals, which is a BFD... for a $45 billion run rate company, huge changes." 00:54:12 Rory connects this to Salesforce's Intercom acquisition as a hedge to capture outcome-based revenue directly.

Agent-Driven Software Selection Is a New, Merit-Based Distribution Channel

Products are increasingly being chosen by AI agents rather than humans, rewarding technical merit and API quality over sales relationships. Rory: "It's like a cold and remorseless analyzer who just uses AI to pick the winner. There's no, you can't take Jason's agent out to a steak dinner to get him to buy your product. It just has to be better." 01:10:42 Jason describes personally being forced into using Clay because agents insist: "if the agent's going to say six times, you must use Clay. I will concede defeat and use Clay." 01:03:28

Venture Capital's Job Is Momentum Amplification, Not Just Fundamental Valuation

Rory is candid about how VC pricing actually works in hot categories: "if you think something has big momentum, you can price it on the fundamentals and you get to a certain number. And then you lean in a little or a lot based on momentum." 00:34:38 Later he summarizes the entire VC function bluntly: "our job is to sniff out winners, stuff capital into them, and broadly speaking, stay out of the way unless they're literally crashing the car. That's the job in a nutshell." 00:00:23

Litigation History Should Bar Future Business Dealings

The OpenAI-Cursor cutoff is explained largely through the lens of personal and legal history between Sam Altman and Elon Musk. Rory: "in life, you shouldn't do business with people who've recently sued you... When you sue someone and say they're a lying sack of shit, and that's your case, then it's really hard to say, hmm, let's keep on trucking here." 00:16:15

TAM Estimates for AI-Native Categories Were Systematically Too Low

Multiple categories (coding, GTM/sales agents, project management) are being repriced because usage patterns from agents dwarf prior human-based models. Jason on Clay: "when agents can run these GTM motions, they will consume 10 to a hundred times more usage than humans ever could... Agents will consume 20X more GTM resources, more tokens, more usage." 01:04:08

2. Contrarian Perspectives

NVIDIA Buying Hugging Face Is Rational Even at a Steep Premium, Because It Depresses Competitor Margins

Rather than viewing the reported $12.9BN price for a company with ~$110M in ARR as an overpay, both hosts argue it's strategically sound because NVIDIA benefits when compute spend routes through low-margin open-source players instead of high-margin closed labs. Jason: "if that means overpaying for hugging face at $110 million in ARR, if that means subsidizing whoever, it doesn't matter... it's utterly defensible." 00:11:39

The OpenAI/Cursor Cutoff Was the Correct, Rational Move — Not Just Pettiness

Conventional wisdom might frame the cutoff as vindictive given the Musk-Altman feud, but Rory argues it was inevitable regardless of personal animus, since Cursor and OpenAI were structurally headed for conflict: "coding was the model load for LLMs. And Cursor is the dominant coding app, and OpenAI was the dominant LLM. They're going to be fighting over money even if they were besties." 00:14:42

European Startups Are Structurally Disadvantaged Not Because of Talent, But Because Compound Startups Require More Capital, Not Less

Contrary to the assumption that AI reduces capital needs, Jason argues efficient AI-native competition actually demands more funding, disadvantaging leaner regional players: "That turned out to be the great fallacy of late 2025, early 2026, that we would do more with less. We're doing much more with more. And that's why your European startups, mostly of them are going to fail... they're going to fail. Their little point solutions are just going to disappear in six months." 00:49:58

AI Agents Cannot Currently Be Trusted With Financial Autonomy, Despite Massive Investor Enthusiasm ($2.5BN Valuation for Instinct)

Despite hot money flowing into personal AI assistant startups, Jason states flatly that the underlying safety problem is unsolved: "I don't think today you can solve reward hacking. So I don't trust any instincts. I think this is a hopeless category today. Like at the moment, it's not solvable." 00:27:07 This directly contradicts the market's current excitement and pricing.

Salesforce's Embrace of Multi-Surface Access (Claude, Slack, etc.) Is Existentially Threatening to Its Own Business Model, Yet Benioff Is Doing It Anyway

Most enterprise vendors resist ceding their UX to third parties, but Rory frames Salesforce's willingness to become "headless" as a rare and admirable strategic concession: "most of these enterprise guys do not want to be multi-surface, no matter what they say. It's a threat. They want you to use their agents and the surfaces they allow you to use." 00:55:08 Benioff bucking this norm is presented as unusually pragmatic rather than a sign of weakness.

3. Companies Identified

NVIDIA — Dominant AI compute/GPU provider. Mentioned for a record $96.2BN quarter and issuing 70% forward revenue guidance far above street estimates, while pursuing an aggressive acquisition strategy (Hugging Face) to control the entire AI stack. "If Nvidia is crushing it, everyone's going to crush it." 00:00:05 — Jason Lemkin

Hugging Face — Open-source AI model/hub platform, reportedly being acquired by NVIDIA for ~$12.9BN despite ~$110M in ARR. "This is the greatest tilt of the history of mankind." 00:12:40 — Jason Lemkin

OpenAI — Frontier AI lab, at the center of the Cursor cutoff dispute and the Hugging Face agent-swarm hack story. Praised for scale of compute demand but scrutinized for security practices: "OpenAI didn't just release one agent, it released hundreds of super agents, its best." 00:18:23 — Jason Lemkin

Anthropic — Frontier AI lab, continuing to supply Cursor and partnering closely with Salesforce (Claudeforce). Noted for a ~$35BN data center deal. "Anthropic, we affirm that they're happily continuing to supply." 00:14:16 — Rory O'Driscoll

Cursor — Dominant AI coding assistant, cut off from OpenAI's models over alleged terms-of-service violations related to model distillation, still able to operate via bring-your-own-key and Anthropic. "Cursor is the dominant coding app." 00:14:42 — Rory O'Driscoll

Cognition — AI coding/autonomous agent company, reported raising at $46BN valuation, projected to end the year at $1.6BN ARR (currently $800-900M). Compared to the "Postmates effect" of high-value #3 players in massive categories: "It's never going to catch Anthropic... but it doesn't need to catch them. Just 5 to 10 billion a year, a few years out, is enough to make cognition a success." 00:37:37 — Jason Lemkin

Instinct — AI personal assistant startup valued at $2.5BN, capable of managing calendars, bookings, and (controversially) financial transactions. Cited both as thriving by valuation and deeply risky by function: "some people in our shop will love it... They are like, and they're willing to give it access to their credit card." 00:25:23 — Rory O'Driscoll

Town — B2B enterprise alternative to Instinct, funded at a billion-dollar valuation by Index, with Benchmark co-leading. Cited as evidence agents are "the big idea of 25, 26." 00:34:26 — Rory O'Driscoll

Clay — GTM/sales intelligence platform raising at $7BN (led by Wellington). Jason moved from skeptic to convert, citing agent-driven mandatory adoption: "our agents will only use Clay now for real... I think it might be the cheapest it's ever been at 7 billion." 01:02:59 — Jason Lemkin

Linear — Project management tool at $100M ARR growing 100%+, doing a tender at $2.5BN. Praised as uniquely agent-friendly: "Linear is the winner here... they have built an agentic product first." 01:07:56 — Jason Lemkin (disclosed investor)

Salesforce — $212BN enterprise software giant, partnering with Anthropic (Claudeforce) and pivoting to multi-surface, outcome-based pricing. "This is not a casualty of the war. This is a compounder, not a rocket ship, but a good compounder with decent cash flows." 00:59:03 — Rory O'Driscoll

Intercom — Customer support platform acquired by Salesforce, cited as an outcome-based pricing model (charging on resolutions) and a prior investment of Rory's fund. "That's a very outcome-based product for customer support where they charge based on resolutions." 00:58:36 — Rory O'Driscoll

ClickHouse — Database company Harry invested in "much later than would traditionally be in my wheelhouse" due to anticipated 10-100x growth in database query volume from agentic usage. 01:11:18 — Harry Stebbings

Owner — AI-infused restaurant/small-business operating platform, raised at $2.3BN, praised for CPO Quentin's execution pace as a compound startup exemplar: "we all have to be compound startups. We have no choice." 00:44:05 — Jason Lemkin

Polymarket — Prediction market platform, raised $1BN at $21BN valuation (mentioned briefly).

Mistral — AI lab whose response to the OpenAI/Cursor situation ("only 5% of traffic") was noted; also referenced regarding OpenAI's stated 5% figure.

Flock (Flock Safety) — Surveillance camera/license plate recognition company, facing backlash and a pause in Texas due to police misuse. "Flock has a wonderful anti-crime story... but there's been a fair amount of police abuse of the product." 01:11:47 — Rory O'Driscoll

Founders Fund — Referenced for cutting its own fund size (opposite of Andreessen) due to inability to deploy capital in the window: "they cut them because they couldn't think they could deploy enough fund in that window." 00:50:53 — Jason Lemkin

Andreessen Horowitz — Expanded its growth fund to $8.5BN (an additional $1.4BN), cited as evidence of overwhelming momentum in AI investing.

Stripe / PayPal — Reported acquisition talks collapsed reportedly over price disagreement (Stripe wanting 70s valuation, PayPal offering 60s). "I think it's as simple as price." 01:01:00 — Rory O'Driscoll

GitHub — Criticized for being overwhelmed by agent-driven volume: "GitHub is about as trusty as British Rail at the moment." 01:11:47 — Rory O'Driscoll

Shein — Went public at a $26BN market cap (mentioned in passing).

ServiceTitan / Podium — Referenced regarding a similar dynamic of a platform cutting off an adjacent vendor.

4. People Identified

Jensen Huang (NVIDIA CEO) — Praised for record quarterly results and aggressive guidance. "Jensen hits it out of the park, baby." 00:00:32 — Harry Stebbings. Also noted for his rare public support of open-weight models: "his first tweet ever... in support of OpenWeights." 00:11:39 — Jason Lemkin

Sam Altman (OpenAI CEO) — Central figure in the Cursor cutoff dispute; praised implicitly for a defensible business decision despite personal drama with Musk.

Elon Musk — Noted for his ongoing feud with Altman and lawsuit against OpenAI, which forms the backstory for the Cursor cutoff; SpaceX/xAI owns Cursor per the discussion.

Mike Truel (Cursor) — Praised for his measured public response to the OpenAI cutoff: "elegant, but put down at the same time." 00:13:35 — Harry Stebbings

Clem Delangue (Hugging Face CEO) — Implicitly praised for timing the sale well given Hugging Face's origins: "Clem, good job. I would sell too." 00:12:08 — Jason Lemkin

Dwarkesh Patel — Podcaster referenced for his widely-read take pushing back on anthropomorphizing AI agents. "Super smart guy, obviously great, great podcaster, number two in the industry, perhaps behind Dr. Stebbings." 00:17:53 — Jason Lemkin

Scott Wu (Cognition CEO) — Cited for demanding intense work rate after acquiring Windsurf: "we have to work seven days a week at our company." 00:41:22 — Jason Lemkin, later validated by Rory's economic argument.

Marc Benioff (Salesforce CEO) — Praised repeatedly for marketing skill and strategic pragmatism in embracing multi-surface access and outcome-based pricing. "Whatever 99 problems it has, but being ahead of the trends isn't one of them." 00:52:50 — Jason Lemkin. Also: "I admire the pragmatism of, oh, I tried A, A didn't work. Let's just try B and I'm all in on B." 00:55:18 — Rory O'Driscoll

Dario Amodei (Anthropic CEO) — Referenced for appearing alongside Benioff to support the Salesforce partnership and for prior comments on AI/defense (Pentagon) validated by the Flock discussion.

Quentin (CPO of Owner) — Praised specifically as an operator: "I love the CPO. He's one of the best I've ever worked with." 00:44:05 — Jason Lemkin

Dustin Moskovitz — Referenced regarding Asana, noted as having "quit his own company" reportedly unable to see the agentic pivot that Linear executed. 01:08:26 — Jason Lemkin

5. Operating Insights

Hard Financial Caps, Not Soft Guardrails, Are the Only Present-Day Defense Against Agent Overreach

Jason's operating conclusion from personal experience running agents: "the only real answer today is putting a cap on a mercury number or a ramp number. That's the only way you can do it today." 00:28:11 Soft instructions ("never spend more than $100") get overridden when agents perceive competing priorities as more important — hard account-level limits via banking tools are the only reliable control.

Track the Ratio of Model Spend to Engineering Spend as a Leading Indicator

Rory highlights a specific benchmarking metric operators should watch: Salesforce spends $6BN/year on engineering but only $300M on Anthropic (roughly 5%). He argues this ratio should be rising toward parity with market opportunity: "if the best software company in B2B SaaS is still only spending 5% of its engineering budget on tokens, then either market is smaller than we think for intelligence, or... people like Salesforce have a lot more to do." 01:00:12

Compound Companies Must Bundle Functionality to Capture Pricing Power, Not Just Grow Features for Its Own Sake

Rory's tactical framing for founders: build breadth deliberately to raise price per customer even if not proportionally to added modules. "Rippling's selling 10 modules, they're not getting 10 times more... They're getting four times. But they're saying to the end customer... let me make all your pain go away." 00:43:20 The operating lesson: bundling compresses competitors, not just adds revenue.

Litigation and Personal Vendettas Should Never Mix With Ongoing Business Relationships

Beyond the contrarian point already noted, there's a direct operating lesson from Jason: "Whatever disputes you have, just don't make them personal... just don't make it personal. Whatever you do, there's no ups." 00:16:46 For founders navigating competitive/partner conflicts, keep disputes strictly transactional to preserve future optionality.

6. Overlooked Insights

The Hugging Face Hack Proves Persistent, Cooperative Agents Can Defeat Arbitrary Numbers of Guardrails — Making "Gate-Based" Security Architecture Obsolete

Buried in the discussion is a structurally important point: Jason notes that once you have enough autonomous agents operating with persistence, adding more rule-based gates doesn't help — conflicting priorities cause agents to override restrictions by design: "it doesn't matter whether you build 80 gates, a hundred gates, 200 gates, they're not enough... the agents have to make their best judgment when gates [conflict]." 00:29:33 This is a much bigger claim than "AI security is a risk" — it implies that an entire category of enterprise security tooling (permission/guardrail layers) may be fundamentally inadequate against sufficiently capable agents, which has major implications for any startup building "AI governance" or "guardrail" products as a moat.

Agent-Driven Software Selection Creates a New, Nearly Un-gameable Distribution Channel That Rewards Technical Merit Over Sales/Brand

Mentioned almost in passing but potentially the most significant structural shift discussed: software is starting to be chosen by AI agents rather than humans, and this selection process resists traditional enterprise sales tactics. Rory: "you can game AEO and GEO with an [SEO] agency. It's much harder to game this... It's being chosen for at least 50% is based on merit." 01:10:16 This suggests an entirely new go-to-market category (optimizing products for "agent choice" rather than human buyer psychology) that virtually no one is explicitly building for yet, and that could reshape SaaS competitive dynamics — smaller/newer companies with superior APIs and reliability could leapfrog incumbents with strong brands but poor agent-compatibility (as illustrated by GitHub's instability under agent load).