Jason Lemkin | 20VC Summary

1. Key Themes

Anthropic's Structural Advantage Over OpenAI Is Compounding in a Frightening Way

The revenue gap between Anthropic and OpenAI is widening, but the more alarming signal is the cost structure differential. Anthropic reached $30B ARR (up from $9B at the start of the year) while spending a fraction on training.

"Their training costs are a quarter of OpenAI. It really feels like the investors in OpenAI got a much worse deal in the last round than the Anthropic ones did." — Jason Lemkin 00:07:13

"They're growing faster than us. The gross margin economics are roughly the same slash slightly better. And their cost below the line and to a rounding error costs are compute and scientists to run the compute for training are better. And that's a bad fact pattern." — Rory O'Driscoll 00:08:47

Rory made the most analytically crisp observation: if both were public stocks, hedge funds would be running a long/short trade — long Anthropic at $370B, short OpenAI at $820B — as a pure relative performance bet with AI risk hedged out.

"If you short the one and long the other, you're kind of diversifying away the AI overall risk and you're just making a relative performance bet. That would be an interesting one." — Rory O'Driscoll 00:11:57

AI-Driven Cyber Threats Are About to Overwhelm Unprepared B2B Companies

This was framed as an emerging but inevitable crisis. The Mercor hack was presented as a preview of what's coming at scale. The key insight: AI commoditizes attack infrastructure, turning what used to require hacker farms of humans into 24/7 automated assault.

"With AI, you can hack anybody you want... it will use every possible way to penetrate that, to steal the data, to create malicious acts. Everything that is possible on the face of the earth that can be delivered with software." — Jason Lemkin 00:16:52

"It used to be hacker farms of people in fill in the blanks, the Philippines, Russia. Now it's going to be hacker farms of AI agents cranking 24-7. It's going to be miserable." — Rory O'Driscoll 00:16:59

Rory explicitly called out that the security market sell-off following the Anthropic announcement was irrational — security spend should be accelerating, not declining.

"I think the response of security stocks going down to the Anthropic announcement was absurd. I think anyone who's cutting back on the security budget in 2026 is missing the point." — Rory O'Driscoll 00:15:41

The Agentic Era Is Already Creating More Databases Than Humans — Supabase Is the Sleeper Winner

This is not a future thesis — it has already crossed the threshold. Agents are spinning up more databases than human developers, and Supabase is the default infrastructure layer for that activity.

"I know more databases are being created by agents than humans. So that is the trend you're betting on." — Jason Lemkin 00:59:37

"Whether you even realize you have a database, all the millions and millions and millions of vibe coded apps have a database in the background. So with Supabase, they get to monetize them all." — Jason Lemkin 01:00:05

Rory drew the MongoDB parallel explicitly — Supabase is playing the same role MongoDB played in the SaaS/cloud era, but on an accelerated timeline. The window to lock in developer loyalty before commoditization closes is measured in years, not decades.

"You want to be the right database for vibe coded and agent apps in 2026, 27, or 28." — Rory O'Driscoll 01:01:01

2. Contrarian Perspectives

Most of OpenAI's Mega-Round Was Not Actually Cash — A Sign of Weakness Disguised as Strength

The conventional narrative around OpenAI's massive funding rounds is one of dominance. The contrarian read: the structure of the round tells a different story.

"The SoftBank money comes in tranches. They have to borrow money to pay it. The Amazon money is tranched in part on IPO or AGI. And the NVIDIA money is almost all not money. It's almost all offsets and compute. It is not a sign of strength where the majority of the round is not cash up front." — Jason Lemkin 00:09:42

Buying Media Assets Almost Never Works — And OpenAI's Acquisition of TBPN Is a Vanity Project Masquerading as Strategy

Andreessen's thesis about controlling the media narrative has been widely adopted, but Rory called it out directly: OpenAI is already the most-covered company on the planet. More media exposure solves nothing.

"Owning a media asset invariably takes way more time than you think for way less money than you expect. See Jeff Bezos for details." — Rory O'Driscoll 00:22:48

"You don't launch a Code Red edict and a Focus edict and no more side projects edict, and then within the space of a week do something that's so obviously a side project." — Rory O'Driscoll 00:16:40

The "Perfect LinkedIn" Executive Hire in Turmoil Has Only a 30% Success Rate

Bringing in seasoned executives with impeccable credentials during a crisis is a common playbook — and Jason argues it rarely works.

"In my experience, that has about a 30% chance of success, just roughly. That bringing in the perfect LinkedIn, giving them a massive portfolio... when you're in tumult, there's not a lot of time to learn everything, right? There's not a lot of time for the get-to-know-you tour." — Jason Lemkin 00:14:58

Many VCs Are "Enablers" Running a Pre-AI Playbook That's Actively Harmful to Their Portfolio Companies

The instinct to be supportive of founders in tough times is increasingly dangerous when the competitive landscape is shifting weekly.

"I see too many VCs running a pre-AI enabler playbook where when folks do fall behind a tick or two, you see kumbaya activity instead of code red activity... enablers can enable a death spiral that you feel good about as you approach the event horizon and your startup implodes." — Jason Lemkin 01:05:57

The Most Instructive Marketers Right Now Are the Fraudsters and Dodgy Operators on the Frontier

The GLP-1 two-person, $1.8B revenue company was held up not as a model to copy, but as a 12-month preview of what AI-powered marketing will look like when it goes mainstream.

"We are watching the future. And if you don't adapt, you're going to be killed by folks running the old playbook." — Jason Lemkin 01:19:12

"The most entrepreneurial people on the planet out there right now are the hackers and the kind of dodgy marketers that are on the front line of pushing things, but the tactics that they're using and the way they're leveraging AI, everyone's going to be doing within a year or two." — Rory O'Driscoll 01:21:31

3. Companies Identified

Anthropic Leading AI lab. Mentioned for surpassing OpenAI in revenue ($30B ARR), having training costs at one-quarter of OpenAI's, and being compute-constrained despite explosive growth.

"Getting to 30 billion up from 9 billion at the start of the year... Anthropic got there in five. But maybe they really got there in three." — Jason Lemkin 00:04:30

Supabase Open-source Postgres-based database platform. Mentioned for becoming the default database infrastructure layer for agentic and vibe-coded applications, raising at a $10B valuation.

"They built a product that could basically self-deploy a Postgres database. And it's what every agentic product needed, right? They needed to spool up a database without humans and they leaned the hell into it." — Jason Lemkin 00:59:11

OpenRouter LLM routing marketplace. Mentioned for elegant dynamic model selection, rapid growth (10M to 50M ARR in ~6 months), and for its potential Stripe/Twilio-like infrastructure positioning.

"Twilio was an interface between an app builder and all the complexities of telco. And these guys are an interface between an app builder and all the complexities of LLMs." — Rory O'Driscoll 00:48:07

RevenueCat Mobile subscription management platform. Mentioned as a comparable case study for low-take-rate, high-volume infrastructure businesses — praised for 50% market share in mobile subscription management and recent AI-driven growth acceleration.

"They have about 50% market share in managing mobile subscriptions. If you have a mobile app that is paid, 50% chance they have RevenueCat deployed." — Jason Lemkin 00:51:58

SpaceX Launch and satellite company. Mentioned for confidentially filing for IPO at a target $2T valuation, $15-16B in 2025 revenue, $8B EBIT, and for being one of the "Big Three" whose combined IPO value may exceed all IPOs of the last 20 years.

"The big three, SpaceX plus OpenAI plus Anthropic, assuming they all IPO... their value at IPO will exceed every other IPO for the last 20 years combined. All of them." — Jason Lemkin 00:30:27

4. People Identified

Doug Leone Sequoia Capital founding partner. Mentioned for returning to active investing at Sequoia (not in leadership capacity), praised as a deal-closer and gravitas provider in an increasingly competitive talent market.

"When Doug Leone goes to that meeting with them, whether it's Christian Hecker at Trade Republic in Germany, or whether it's the team at Wizz, he fucking closes the deal." — Harry Stebbings 00:39:54

Denise Dresser Former CEO of Slack, now leading go-to-market and related functions at OpenAI. Mentioned as OpenAI's big leadership bet — praised for credentials but flagged as a high-risk appointment given the timing and turmoil.

"You take Denise Dresser, who was CEO of Slack, who came from Salesforce just a couple months ago, and you put her in charge of basically everything go-to-market and related, right? That's a good bet on a seasoned executive." — Jason Lemkin 00:14:30

5. Operating Insights

The "Small Chip / Big Chip" M&A Framework Is an Underrated Forcing Function

Jason described Adobe's internal M&A discipline: every senior exec gets one "big chip" deal (bet your job on it) and one "small chip" deal (lower stakes, one per year, no career consequence if it fails). This creates a portfolio approach to corporate M&A that manages risk without paralyzing action.

"Everyone got a small chip — what could be like a 50 to 200 million dollar deal. And you had to justify it and you didn't get five as a forcing function. You got one, but you really weren't challenged that much to do the smaller chip. You picked one a year and you didn't get fired if it didn't work out." — Jason Lemkin 00:28:32

Time and Management Turnover Are the Silent Killers of M&A Deals — Default Yes to Good Deals When the Champion Is in the Room

M&A windows close fast and permanently. The lesson from the TBPN deal: the champion who wanted the deal in January is gone by April. The probability that the same executive, with the same priorities, is still there 12 months later approaches single digits.

"The odds that the VP that wants to do the deal is there in 12 months and that their priorities have not changed, you know, approaches single digits." — Jason Lemkin 00:24:33

Being "Founder-Honest" and "Founder-Fact-Based" Is More Valuable Than Being "Founder-Friendly" — And Requires Direct Product Knowledge

Rory reframed what good board membership looks like in the AI era: not cheerleading, not being a jerk, but grounding every conversation in a factual competitive assessment based on hands-on product experience.

"You have to be very clear-eyed with your companies on where the competition is... ideally with hands-on experience of the products. And then you find a way, without being a jerk, to keep the company honest about where they are relative to the competition." — Rory O'Driscoll 00:06:53

When a Liquidity Window Opens in a Private Company, Always Take Some Off the Table

Rory's consistent message to employees at high-valued private companies: the window may not reopen. Don't assume the current price is a floor.

"When the liquidity window opens, take it seriously. Because it mightn't open again for a while... The alert reader should say, if you're planning to buy that house in San Francisco, you might need an extra few million just based on what I'm seeing in the market." — Rory O'Driscoll 00:12:26

6. Overlooked Insights

Chinese Open-Source Models Are Quietly Subsidizing the American Independent Software Vendor Ecosystem

This was mentioned briefly and almost in passing, but the structural implication is enormous. The dominant models on OpenRouter are Chinese open-source models (Qwen, Kimi, etc.). This means that Chinese state-backed AI investment is effectively reducing inference costs for American startups — creating a hidden subsidy to the US developer ecosystem.

"Right now, amazingly, all these open source models are primarily Chinese open source models. And ironically, the Chinese Communist Party is effectively subsidizing the American small independent software vendor by providing cheap open source models. God bless them." — Rory O'Driscoll 00:51:11

The investment implication: any infrastructure company (like OpenRouter) that routes across models benefits directly from this dynamic. If Meta or a US company ships a competitive open-source model, the cost advantage narrows — but until then, this is a structural tailwind hiding in plain sight.

Compute Constraints at Anthropic Mean Pricing Power Is About to Dramatically Increase — And the Playbook Is Already Being Executed

The removal of open-source Claude API access from base plans isn't just a product decision — it's the first visible move in a deliberate capacity rationing strategy. As Rory explained, when you're supply-constrained, rational economics dictate you allocate capacity to highest-revenue uses. This signals that Anthropic's effective pricing per token is going to increase significantly before new data centers come online.

"What you start doing is allocating capacity based on money. One of the first things they've figured out is these folks using these open Claude type agents are consuming vast amounts of tokens on fixed price plans. And they probably want to stop that, which is what they've done." — Rory O'Driscoll 00:05:56

The overlooked implication for investors: any company whose core product is built on Anthropic API at fixed or near-fixed pricing is facing a margin compression event in the next 12-18 months that is not yet priced into their models or valuations.