Jason Lemkin | 20VC Summary

1. Key Themes

Anthropic Is Winning the Enterprise AI Battle — And OpenAI's Culture Is to Blame

The Ramp data showing 73% of new AI spending going to Anthropic is a leading indicator of a serious shift. But the panel argues the real issue is OpenAI's strategic inconsistency and wounded public posture, contrasted with Anthropic's focused identity.

"What I love about Anthropic is it's very consistent about its ICP and goals. It has been very consistent. We know what it stands for. We know what it's trying to do. OpenAI — I'm getting whiplash from everything." — Jason Lemkin 00:08:12

"There is a moment, there is a tide in the affairs of men as Shakespeare says, and this has been the last six months of coding — locking the recognition that coding is the motherlode app within the enterprise spend. And you're right, if you let Claude run away with that for another six or twelve months, you've probably sacrificed value that you'll never get back." — Rory O'Driscoll 00:13:52

The lock-in dynamic is accelerating. Once enterprises build scaffolding and AI agents on Claude, the soft costs of switching become prohibitive — not because of contracts, but because it works.

"Our AI VP of Marketing defines every day every single marketing activity... it runs on Sonnet 4.7. There is no way we're going to switch the model. It took us weeks to dial it in. There's no way we're going to switch them to Codex." — Jason Lemkin 00:14:42

The "Can You Charge for Your AI Product?" Test Is the New Litmus for SaaS Survival

The panel argues that legacy SaaS companies face existential risk if they cannot monetize AI — not because of competition per se, but because markets are rationally repricing the durability of their revenue.

"Why the hell would you not charge for your AI product today? If it's not good enough to charge for, it doesn't count. You're not an AI company if you can't charge for it. Very few public companies can effectively monetize AI. And that's why they're all in terminal decline." — Jason Lemkin 00:52:58

The 50% ARPU growth benchmark was offered as a concrete test:

"Is your ARPU 50% or higher than it was pre-AI? It's a really simple test... It appears that Notion has done that, in which case they passed the test." — Jason Lemkin 00:54:23

Rory adds the investor framing:

"If you're a software product and you don't think AI is going to disrupt not just how you build but what you build, then you actually probably want to actively short it." — Rory O'Driscoll 00:44:00

The Unicorn Exit Crisis: Acquirer-to-Unicorn Ratio Is at an All-Time Low

This is the most underappreciated structural risk in venture today. The panel argues that the exit math is fundamentally broken — and almost no one is discussing it seriously.

"I just worry there's some ratio of potential acquirers divided by unicorns and I think we're at the lowest ratio of our careers. I just don't believe the hyperscalers are going to buy these companies. Basically it's win or die." — Jason Lemkin 00:00:15

"We have outstripped any current ability for these companies to have any exit. So you are going all in on the IPO and not without any worry, rhyme or reason. And I think so many folks sub-Anthropic are going to get their arses burned because they'll end up in the dead zone — they'll have great companies without great IPOs and zero M&A opportunities." — Jason Lemkin 00:04:59

Rory adds a structural reason why the old acquirers can't step in:

"If you're Harvey and you're worth $10 billion, and the old practice management law software is only worth $2 billion, they're not going to buy you. Basically it's win or die." — Rory O'Driscoll 00:08:39

2. Contrarian Perspectives

Token Cost Optimization Is a Distraction for Most AI Applications

The conventional wisdom is that AI companies must ruthlessly optimize token costs as a core competency. The panel argues the opposite is true for a large and growing class of applications.

"There are so many applications that will deliver epic value on these LLMs where it's not worth it. You want to reduce my token cost from $2,000 a month to $1,500? Leave me alone. There are going to be more of those apps than we think." — Jason Lemkin 00:16:43

Rory adds a useful mental model for investors:

"I've been thinking about for us, for our software apps investments, is having this mental model of what's the token spend as a percentage of revenue... There's a ton of really interesting apps that you have for 5, 7, 8% of revenue on tokens are building huge value. Very different than the coding apps where you might be at 40 or 50%." — Rory O'Driscoll 00:17:01

Sales and Go-to-Market AI Adoption Is Essentially Irrelevant for Software Companies

While the industry celebrates AI-powered GTM, the panel argues this is noise compared to what actually matters — whether AI transforms the end product.

"If you're a software company, the number one question is how does AI change the end product you deliver your customers? That's what's going to determine success or failure... you could be using AI well or badly in go-to-market, it will catch up with you over time if you're not using it well, but that's not what's driving 30, 40% price declines." — Rory O'Driscoll 00:46:24

The Groq/NVIDIA "Aqui-hire" Structure Is a Sign That Antitrust Avoidance Has Become a Tax — Paid Either to Government or the IRS

Most commentary on the deal focused on the price. The panel identifies a deeper, more troubling dynamic: the structure itself reveals a perverse two-path system where acquirers pay either in lobbying or double taxation.

"You now have a process whereby the government makes the rules that enforces the antitrust... You either lobby extensively at the highest level of the administration and get a waiver from the top down, or you do it this way and pay double taxation. And the government wins either way. Wire me the money and I'll let you off, or wire me the money after it closes. Wire me the money either way." — Rory O'Driscoll 00:39:16

Being "Disciplined" in VC Has Actually Been the Wrong Strategy Recently

The panel is unusually self-critical here. The conventional VC virtue of ownership discipline and valuation rigor has been punished in the current cycle.

"I fuck it, will be at the end of the day here. My biggest regret with our Series A fund is not being more elastic, disregarding the Series A mandate and just saying, I'm going to leverage the brand and the access that I have to get into super hot companies like Eleven Labs, Lagoras, Lovables much earlier and being a momentum investor in a hot environment." — Harry Stebbings 00:57:09

"You've had players abscond with AUMs like never before, pay insane prices and get fast markups. And in a lot of cases, DPI because Zuck buys you an insane price." — Harry Stebbings 00:13:16

3. Companies Identified

Anthropic

AI foundation model company. Mentioned as the clear winner in enterprise AI adoption, with consistent brand identity, rapidly growing revenue (potentially $22B run rate), and a product quality step-function since Opus/Claude 4.5. Referenced as a case study in consistent ICP focus vs. OpenAI's whiplash strategy.

"The conclusion is real if Anthropic now is maybe a $22 billion run rate, the revenue does tie to these conclusions." — Jason Lemkin 00:06:43

Notion

Productivity and knowledge management SaaS. Cited as one of the few legacy SaaS companies successfully monetizing AI — doubling ARPU with AI add-ons.

"They have effectively doubled ARPU from it... Here's a way to look at an SMB product today: is your ARPU 50% or higher than it was pre-AI? It's a really simple test. It appears that Notion has done that." — Jason Lemkin 00:54:23

Intercom

Customer communications platform. Cited as the rare case study of a company that deliberately let its core business partially decline in order to bet fully on agentic AI — and succeeded.

"If you want to use it as a case study, it does illustrate what you have to do. Owen was clear: we let our core business go into partial decline. It's very hard, almost impossible for a public company." — Jason Lemkin 00:48:39

SpaceX / Starlink

Mentioned as the clearest example of step-function business building — each technical achievement unlocking a new value layer. Starlink's 53% profit margins make the $2T valuation discussion credible to the panel.

"If Starlink really has 53% profit margins and is wildly profitable, the fact that this extends the Starlink vision five orders of magnitude — it's actually a reason to say, if I believe in this at all, my DCF has gone up." — Jason Lemkin 00:23:58

Groq

AI inference chip company. Acquired by NVIDIA for ~$20B on sub-$100M ARR. Discussed as proof that strategic value to an acquirer can massively transcend standalone revenue multiples — and that AI inference has entered production at NVIDIA scale.

"They just announced at GDC last week that it's going into production. That's different. Jensen said within a year we can get this into production. That's worth billions." — Jason Lemkin 00:35:39

Palantir

Enterprise AI/data analytics. Cited as the aspirational model for how to respond to market doubters — by executing and eventually proving the bears wrong on earnings calls.

"Eventually it will turn. That's why drawdowns are hard. But the floor doesn't come because the market changes its mind. The floor comes if you execute." — Rory O'Driscoll 00:41:36

Ramp

Corporate spend management platform. Their transaction data (~0.5-1% of US GDP) is now treated as a leading economic indicator for AI spending trends. The 73% Anthropic figure came from Ramp data.

"I would argue Ramp is probably a pretty accurate statistical reflection of especially digital company spend in the US. They probably have decent data and good data scientists." — Rory O'Driscoll 00:05:52

4. People Identified

Jonathan (Groq Founder)

Founder of Groq. Mentioned for perseverance through years of being in the "desert and dark," taking the company from concept through $20B acquisition by NVIDIA despite heavy criticism along the way.

"Jonathan was in the desert and the dark for many years and he is a founder who's been a real, respectfully, cockroach who's gone through the hard times, who's gone through the criticism. He's also just a good dude." — Harry Stebbings 00:39:35

Yasmin (Spark Capital)

Partner at Spark Capital. Cited for the courage to lead Anthropic's early rounds at $4B pre-revenue — a bet that has paid off enormously and is now considered one of the best risk-adjusted venture investments of the era.

"I give all credit to, for example, Spark — Yasmin there at Spark — who broke all the rules and said $4 billion for a pre-revenue company, Anthropic. That's paid off." — Rory O'Driscoll 00:10:26

Matt Murphy (Menlo Ventures)

Partner at Menlo Ventures. Called out alongside Spark for consistently making the right call on high-conviction AI bets despite valuation skepticism from more "disciplined" investors.

"Do you vote for Matt Murphy at Menlo? Matt has consistently..." — Rory O'Driscoll 00:10:26

Ryan Smith (Qualtrics Founder)

Founder of Qualtrics, owner of Utah Jazz. Used as an example of how billionaires are drawn to environments where they feel welcomed rather than vilified — and Utah/Florida as the new centers of gravity for capital and entrepreneurship.

"Ryan Smith is beloved here in Provo where I am today. Sam Altman is vilified... who the hell wants to live where you're vilified?" — Jason Lemkin 00:32:50

5. Operating Insights

The "Installed Base as Trap" Framework for Product Allocation

Jason articulates a specific, actionable warning for operators of scaling SaaS companies: your existing customer base can consume 98%+ of engineering and product resources, starving your agentic/AI product investments. The Figma and Atlassian examples show this is not theoretical.

"The trap of earning that 35% can imperil your agentic growth. It can consume more than 100% of all the product and engineering and CS resources you have. Michael Cannon-Brooks, when he was on the show, he alluded to it — and then they did the layoffs. I have no people. It's a trap." — Jason Lemkin 00:49:56

The operating prescription: explicitly allocate to the new thing first, treat the existing business as residual. This is counter-intuitive but necessary.

Use Token Spend as % of Revenue as a Portfolio Health Metric

Rory offers a specific, under-used financial metric for evaluating AI application companies — not gross margin in isolation, but token spend as a percentage of revenue. High ratios (40-50%, as in coding) signal commoditization risk; low ratios (5-8%) signal durable margin and switching cost.

"I've actually meant to do this work... just looking at a couple hundred AI apps and literally looking at the AI token spend as a percentage of revenue across them all. I'd love to know what the pattern is because I totally see very different percentages depending on the token intensity." — Rory O'Driscoll 00:17:30

The 50% ARPU Growth Test for AI Monetization

A simple, CEO-level diagnostic: if your AI product hasn't driven at least 50% ARPU growth, it's not a product — it's a feature. Use this to benchmark board presentations and investment decisions.

"Is your ARPU 50% or higher than it was pre-AI? It's a really simple test. Can you drive ARPU up 50% or more? It appears that Notion has done that, in which case they passed the test." — Jason Lemkin 00:54:23

6. Overlooked Insights

The Open Router Explosion Signals a Hidden Market Bifurcation That Most Investors Are Ignoring

Jason briefly mentions that Open Router data has "exploded since the start of the year" — but the implications are barely discussed. This reveals a bifurcating AI customer landscape: one camp (cost-sensitive, model-agnostic, optimizing constantly) and another camp (quality-locked, sticky, building deep scaffolding on a single model). These are fundamentally different businesses with different competitive dynamics, different pricing power, and different investor risk profiles. Most AI app companies are being evaluated as if they're one thing, but they are actually two.

"There are a large set of customers who are optimizing when to use Kimi and when to use Haiku and when to use Mini... On the other hand, Claude Sonnet and Opus since 4.5 and 4.6 are so good, I want to stick there... There are these two things happening: on the one hand the soft costs are very low to pick a different model, but on the other hand there are high soft costs for managing the outputs and QAing it." — Jason Lemkin 00:11:26

The investor implication: companies building for the quality-locked camp have structurally better businesses — but they are not yet being differentiated in valuation from those serving the cost-optimizing camp.

The $1B Fund Minimum for Series A Is Now a Structural Floor — And Most Firms Haven't Adapted

This point was raised briefly in the context of Kleiner Perkins' fund announcement and quickly moved past — but it's a structural shift with massive LP and competitive implications. If a $1B fund is now the minimum viable size to lead Series A with proper ownership and reserves, the majority of existing "Series A funds" are structurally underpowered and will either get pushed out of lead positions or forced into concentration risk.

"You can't do early with less than a billion if you're going to compete to lead A's. A's now are 30 to 40 million. If you want to lead them, you need to be able to write 25 to $30 million checks. You need 20 across a firm. So if you need 20, $30 million checks, you're at $600 million — I would argue that your fund scale is as small as it could be to lead A's today." — Harry Stebbings 00:56:43

"The average round that we play in has crept up over the last year, year and a half... everyone wrestling with doing deals now is grumpy, stressed, and feeling the pressure." — Rory O'Driscoll 00:00:00 / 01:03:31

The implication for LPs: funds between $300M-$800M targeting Series A leadership are in an awkward no-man's land — too large to be seed-disciplined, too small to lead competitive A's with proper reserves.