Jason Lemkin | 20VC Summary

1. Key Themes

The Capital Wall Is Real: Only Five Entities Can Finance a Frontier Model

The Poolside acquisition illustrates a structural truth — venture capital ran out of runway for frontier model building years ago. The only funders capable of sustaining state-of-the-art models are Microsoft, Google, Amazon, and now NVIDIA, with OpenAI and Anthropic themselves as recipients rather than funders.

"The only people capable of financing a state-of-the-art frontier model in the United States of America has been the hyperscalers themselves. The only reason Anthropic and OpenAI exist is because Microsoft, Google, and Amazon gave them enough money to play because no one else on the planet has enough money. And the only other person who now has enough money is NVIDIA." — Jason Lemkin [00:12:30]

NVIDIA Is Now a Venture Fund With a Chip Business Attached

NVIDIA is systematically deploying its enormous cash flow — estimated at $50B+ in free cash flow — across every layer of the AI stack: chips, open-source models (Poolside/Nemotron), data providers (Mercor), and inference products (Perplexity). The strategic logic differs by investment type: vendor financing for some, ecosystem expansion for others.

"The best use for it is to spend the money with people who will in turn enhance the ecosystem... If you invest right, it's brilliant. You get a customer. Not only do you get circular revenue, but you ensure the success of your ecosystem, the viability. You get a twofer out of it." — Rory O'Driscoll [00:20:09]

Anthropic Has Lapped OpenAI, and OpenAI Has No Choice But to IPO

Anthropic is at a $60B annualized run rate growing dramatically faster than OpenAI's ~$25-30B. OpenAI's Q1-to-Q2 growth of 18% quarter-over-quarter was shocking relative to expectations. The competitive math forces OpenAI's hand on going public — they cannot wait.

"If danger can be described as the absence of choice, they were now in danger. And I think for OpenAI, what you're seeing here is the absence of choice is starting to pile up." — Jason Lemkin [00:34:38]

The Token Addiction Is Real and Will Force Corporate Budget Reckonings in 2026-2027

Employees who have experienced high-volume AI agent use will not go back. But the cost is real — users spending $200/month are consuming $8,000–$12,000 worth of tokens — and CFOs are already panicking about under-budgeted token spend. The next tension will be between retaining AI-native talent and controlling AI spend.

"Jesus Christ, if 30% of my company leaves to go work for Harvey, I'm dead in the water." — Rory O'Driscoll [00:47:00]

"I need my 10 subagents running 24 hours a day to do my job or I quit. I would quit. Take away my agents. I quit." — Rory O'Driscoll [00:55:43]

Intelligence Must Be Managed Like Capital, Not Software Seats

The Stripe annual letter articulated a framework shift: intelligence is fungible like money and must be allocated, not simply licensed. This has profound implications for how enterprises budget, who gets token allocations, and how SaaS pricing models evolve.

"What they're saying is seat-based SaaS. I sold five seats to Harry's organization. I'm done. And Harry's done allocating it too... But to Jason's point, if you're buying intelligence on an uncapped basis, in theory, your employees could go on spending that forever... The analogy of saying it's like money. You have spending controls on your money, but you also recognize money is the lifeblood of your business." — Jason Lemkin [00:56:15]

Coding Is the Only Sentence That Matters — Anthropic Won Because of It

The single strategic insight that explains Anthropic's dramatic outperformance of OpenAI is its focus on coding as the primary use case. Coding has the highest ROI, fastest adoption S-curve, and highest willingness to pay of any AI application category.

"What on Anthropic's guard is, it's all about code. That's the sentence. That's the only sentence that matters... Coding is the fastest adopting market. It's the highest ROI market. It's the mother lode. It's as simple as OpenAI was focused here and Anthropic focused there." — Jason Lemkin [00:37:54]

AI Wealth Is Hyperly Concentrated Geographically — and That Has Compounding Consequences

California captured three-quarters of all venture dollars, largely because Anthropic and OpenAI alone took ~60% of total dollars. San Francisco — a city of only ~780,000 people — is experiencing 1999/2007-level asset inflation. This creates a cost spiral that prices out everyone except the AI elite.

"The whole Bay Area is only seven million. This is a wall of money falling into a tiny place that's a peninsula with sea on three sides and a little bit of mountains on a tiny valley called Silicon Valley on the fourth side. Property is not plentiful and it's hard to build." — Jason Lemkin [00:50:55]

The AI Cycle Is Less Than One-Third Complete

Despite all the noise about corrections and volatility, the structural view is that the AI wave has barely started. Even the cloud computing wave lasted nine years, and AI is arguably larger in scope.

"I think you have to believe we're less than a third of the way through this cycle. Even that little tiny cloud thing lasted nine years. We're just getting going here." — Rory O'Driscoll [00:52:35]

Open-Weight Models Are Winning Token Volume — But Closed Frontier Models Will Keep Revenue

Vercel data shows 68% of tokens going to open-weight models, a share that is increasing. But the revenue premium stays with frontier closed-source models because they command more value above inference cost. This creates a structural squeeze on OpenAI's second-place position.

"I think the vast predominance of tokens processed will be open weight. And obviously the significant majority of revenue will still be frontier, state-of-the-art because it can command more value than just the price of inference." — Rory O'Driscoll [00:32:46]

2. Contrarian Perspectives

$9 Billion Is Not a Good Exit for a Seed Investor in 2026

The conventional reaction to a $9B acquisition is celebration. The panel argues it is mathematically insufficient for a seed fund to generate a fund-returner, because effective seed entry prices (with dilution) are now so high that you need ~$63B+ exits for a 50x return.

"9 billion doesn't clear the bar for seed investing in 2026. That's the irony." — Rory O'Driscoll [00:11:04]

"If you get 15x on your failures in venture, you'll die a rich man." — Jason Lemkin [00:12:00]

Customer Support and CX Software Is Already Dead as an Investment Category

Against the consensus enthusiasm for AI customer support companies, both Harry and Rory argue the category is essentially over: large enterprises are building their own systems, only one or two vendors will win meaningful share, and the surface area is collapsing into general agents.

"A classic CS and CX won't even exist in 24 months. There's really going to be commodity cheap products, but we won't even need it. It's already dying and merging into marketing, sales, everything's becoming one agent." — Rory O'Driscoll [01:17:15]

Defense Tech Will Disappoint Venture Investors Despite National Importance

Jason Lemkin argues that defense is strategically important for America but structurally bad for VC returns — because account control and portfolio breadth matter more than individual product excellence, meaning two or three platform players (like Anduril) will consolidate the category.

"I think there's an element of that business that you have to have account control. And I think the two or three largest companies like Anduril will end up doing a bunch of scooping up over the next two decades... I think it's a portfolio of products that it takes to survive the interaction with the Pentagon." — Jason Lemkin [01:15:08]

Consumer AI Subscriptions Are One of the Worst Business Models Ever Built

Selling $200/month subscriptions that deliver $8,000–$12,000 of token value is economically irrational and unsustainable. The consumer AI pricing model, as currently structured, would destroy any standalone business.

"The consumers selling $10,000 worth of tokens for $200 is one of the worst business models of our lifetimes. If that was the only business, these guys would be dead in the water." — Rory O'Driscoll [00:38:49]

AI Agents Still Cannot Be Trusted — And This Problem Is Not Close to Solved

Despite the hype around agentic AI, the panel notes repeated incidents of agents leaking confidential data, making unauthorized purchases, and behaving erratically. Multiple products — OpenClaw, GrokBot, Instinct — have all had the same failure modes. The goal-seeking nature of LLMs makes full trust unlikely in the near term.

"We still can't trust agents today. We can't trust them with anything... The new entrants can't solve the goal-seeking nature of the LLMs they're running on." — Rory O'Driscoll [01:09:17]

3. Companies Identified

Poolside AI model company building a US open-source alternative to Chinese models; acquired by NVIDIA for $6B license + $1B equity investment at a $12B pre-money valuation, with 109 engineers moving to Nemotron.

"They had no choice but to fail up for $6 billion plus topping off a billion at $12 billion." — Rory O'Driscoll [00:05:36]

Nemotron (NVIDIA) NVIDIA's open-weight model effort, now supercharged by the Poolside acquisition and engineering team. Strategic value: open-source models are a complement to NVIDIA chips, pulling compute share away from closed frontier models.

"If I'm the maker of compute, it's awesome for NVIDIA if there's a viable US open source model running on NVIDIA chips, taking market share away from the frontier models at the margin." — Jason Lemkin [00:09:53]

Mercor AI data labeling/training data company raising at a $20B valuation led by General Catalyst, with NVIDIA reportedly participating significantly.

"Gross margins above 30% don't matter anymore... if you value Mercor at eight classic multiples for 80% gross margins and growing, it's not expensive." — Rory O'Driscoll [00:17:24]

Perplexity AI search company raising at a $30B valuation with NVIDIA investing. Harry and Jason are investors.

"There's been a lot of cynicism, skepticism around Aravind Srinivas and Perplexity... We're a small investor in the company. So NVIDIA is actually propping up most of my portfolio." — Harry Stebbings [00:26:20]

Anthropic Closed-source frontier AI lab. Now at $60B annualized run rate, reportedly heading toward an IPO in months, claiming a $30 trillion TAM. Gross margins moved from -91% to +30% in one year.

"Anthropic went from negative 91 to positive 30 in a year." — Jason Lemkin [00:24:56]

OpenAI Frontier AI lab, inventor of ChatGPT. Now positioned as number two to Anthropic by growth rate. CFO Sarah Fryer confirmed 2027 IPO target.

"The ranking has changed, and you're now number two. So you will go out later then at a lower price than the other guys, unless you change the trajectory massively." — Jason Lemkin [00:29:56]

Stripe Payments infrastructure company, accelerating to 41% growth with billings up 71%. Buying back shares while private. Acquired OpenRouter.

"You literally have to argue with an agent to get it not to use Stripe." — Rory O'Driscoll [01:04:38]

OpenRouter Model routing infrastructure acquired by Stripe. Growing 15% month-over-month at $150M ARR when acquired.

"OpenRouter was growing 15% and $150 million... This is the moment when all of a sudden every open weights and these models went from experiments to mainstream. Sell, baby." — Rory O'Driscoll [00:42:56]

Cursor AI coding tool, discussed as a case study of surviving negative gross margins through talent, market positioning, and model development. Now valued at ~$60B.

"The VCism is these are some of the smartest kids ever. They have a strong market position. The wind is at their sails. The kids will figure it out." — Rory O'Driscoll [00:23:35]

Hugging Face Open-weight model hub reportedly in acquisition discussions at $13B, despite only ~$150M in revenue. Strategic value as a neutral platform for enterprise open-weight model access.

"If you think about assets that an IT company might want to own if they were trying to build a counterbalance to the closed weight frontier models, this would be a super interesting asset." — Jason Lemkin [00:42:00]

Harvey AI legal platform cited as a talent magnet pulling best engineers and employees away from incumbent enterprises.

"If 30% of my company leaves to go work for Harvey, I'm dead in the water." — Rory O'Driscoll [01:00:47]

Anduril Defense technology company cited as the likely category consolidator in AI defense, hoovering up smaller point-solution defense tech companies.

"I think the two or three largest companies like Anduril will end up doing a bunch of scooping up over the next two decades." — Jason Lemkin [01:15:08]

Palantir Enterprise AI/data company cited as one of the very few public companies whose growth approaches that of the private AI leaders.

"Nothing except Palantir approaches these, right? Even Cloudflare isn't this good." — Rory O'Driscoll [01:05:58]

Databricks Data/AI platform company expected to IPO soon, cited as one of the assets that will reorder the public market leaderboard.

"When OpenAI and Databricks go public at 80% growth and Stripe accelerating 41% and 71% billings, nothing except Palantir approaches these." — Rory O'Driscoll [01:05:58]

Instinct New AI personal assistant/chief-of-staff agent that went semi-viral among VCs before a data security incident. Described as a next-generation agent with access to email and calendar.

"I would like to invest in the instinct that actually can honestly solve these issues. That one I would do at 600 pre. But it's got to actually solve existential issues that no one else at the moment can solve." — Rory O'Driscoll [01:09:42]

Locus Robotics Warehouse robotics company; Jason Lemkin is a board member. 15,000 robots in the field. Cited as a positive example of purpose-specific robotics succeeding.

"I'm on the board of Locus Robotics. We have 15,000 robots in the field, but it's a specific purpose robot." — Jason Lemkin [01:15:55]

Superhuman AI-powered email productivity tool, now part of Grammarly. Harry and Jason are investors.

"You've got Superhuman, which is now part of Grammarly. Harry and I are proud shareholders there." — Jason Lemkin [01:12:51]

Vanta Compliance automation platform. Trusted by 16,000+ companies including Ramp, Harvey, and Writer.

Framer AI website builder used by Perplexity and Miro.

Deel Global IT and HR platform, managing hardware, software, and access across 130 countries. Used by Eleven Labs.

Citadel Ken Griffin's hedge fund, cited for unwinding 80% of its AI-related fund position (the "Folio" book associated with Leopold Aschenbrenner) after a successful short-term trade.

4. People Identified

Jensen Huang (NVIDIA CEO) Orchestrating NVIDIA's ecosystem investment strategy, reportedly with a $70-75B annual budget to deploy across the AI ecosystem.

"Jensen says there's a budget. I don't know what it's going to be, $70 billion this year maybe. And we're going to spend it all on our customers and ecosystem." — Rory O'Driscoll [00:17:42]

Aravind Srinivas (Perplexity CEO) Praised despite widespread skepticism about his company. NVIDIA is now a significant investor.

"There's been a lot of cynicism, skepticism around Aravind Srinivas and Perplexity. It's a company that actually people like to dislike for whatever reason." — Harry Stebbings [00:26:20]

Sarah Fryer (OpenAI CFO) Confirmed to employees that OpenAI will be a public company in 2027, seen as a forced hand given Anthropic's momentum.

"CFO Sarah Fryer told employees at OpenAI, we will be a public company in 2027." — Harry Stebbings [00:26:49]

Dario Modebi (Anthropic CEO) Cited for both extraordinary execution (taking Anthropic to $60B run rate, transforming gross margins from -91% to +30%) and for some perceived grandiosity ($30 trillion TAM claim, asking job candidates if they'd be happy with Anthropic going to zero).

"Dario is nutso a lot of the time, right? Apparently interviewing folks today, asking if they'd be happy joining Anthropic if it all went to zero. Cool question, actually, but nutso." — Rory O'Driscoll [00:35:52]

Sam Altman (OpenAI CEO) Seen as more likable and approachable than a year ago, but questioned on what OpenAI's differentiated mission is.

"Sam is a much more approachable CEO now. He got through all the Sam Altman crap and he's got the sweaters out and he's the more likable guy." — Rory O'Driscoll [00:35:52]

Ken Griffin (Citadel founder) Cited admiringly for unwinding 80% of a leveraged AI position at profit — demonstrating market-making discipline.

"Never fight with Ken Griffin. Man will come out on top." — Harry Stebbings [00:44:47]

Leopold Aschenbrenner Former OpenAI researcher, wrote the "Situational Awareness" essay that Citadel reportedly built a fund position around, before unwinding 80% of it.

"Citadel unwinds 80% of Leopold Aschenbrenner's four-soul book. I'm like, man, never fight with Ken Griffin." — Harry Stebbings [00:44:47]

Elad Gil Investor cited for advising AI asset owners to sell now during the peak open-weight transition moment.

"Elad Gil said, sell if you have an AI asset right now." — Rory O'Driscoll [00:42:56]

Ben Thompson (Stratechery) Cited for the insight that Silicon Valley perpetually forgets the average American is not trying to optimize their life.

"Ben Thompson and Stratechery has one great comment is, Silicon Valley forgets every three years that the average American is not trying to be efficient." — Jason Lemkin [01:12:23]

Satya Nadella (Microsoft CEO) Cited for his comments that every enterprise needs to own its own knowledge and not cede it to frontier model providers — ironically, given Microsoft's deep investment in OpenAI.

"It started with the Satya comments on every enterprise needs to have its own knowledge and not give it up to the frontier. To which you want to say, no shit, Satya. Well, thanks for funding OpenAI for three years." — Jason Lemkin [00:43:43]

5. Operating Insights

Manage Intelligence Spend Like a Capital Allocation Problem, Not a Software License

The Stripe letter framework — intelligence is fungible like money and must be actively allocated — gives operators a concrete playbook. High performers should get more token budget; low ROI users should be capped. This is the CFO conversation happening right now, and companies without a framework will either blow their budgets or lose their best people.

"You're going to have to control intelligence in a way that's more like how you control money and less like how you think about software licensing." — Jason Lemkin [00:56:42]

Reduce Share Count While Private — It's the Best Signal of Financial Health

Stripe is buying back shares while private, reducing its total share count from levels four years ago. This is a concrete, actionable benchmark for any well-capitalized private company. Fewer shares outstanding means more per-share value in any future exit or IPO.

"Another stunning fact on the Stripe letter, their share count is down on three years ago, four years ago, which means they've been buying back stock. They're doing everything a public company can do while private." — Jason Lemkin [01:07:08]

Retain AI-Native Talent by Giving Them the Tools — The Cost of Not Doing So Is Existential

CFOs are caught between token budget overruns and catastrophic talent flight to AI-native companies. The operating insight is that token spend is not a cost to cut but a retention tool to invest in selectively. Cutting AI access will cause your best people — exactly the ones building leverage — to leave.

"If I don't give these people what they need for AI, I'm going to lose all of them. And it is true. You'll just be stuck with the folks that are still AI skeptics. Your organization will be full of the moldy oldies of SaaS." — Rory O'Driscoll [01:00:47]

When Acquiring a Platform/Marketplace, Don't Touch It for 24-36 Months

Any acquirer of a neutral platform (specifically discussed in the context of Hugging Face) risks destroying the entire value by commercializing it. The operating discipline is hands-off integration.

"If you mess with this marketplace for 10,000 models, even if you put a little ad at the top, you destroy it. I'm almost confident if anyone actually spends $3 billion, let alone $13 billion, they're going to 95% leave it alone for 24 to 36 months." — Rory O'Driscoll [00:44:31]

Force Your Team to Ask "What If It All Goes Right?" as a Structured Exercise

As an operating discipline against natural pessimism and risk-aversion in investment or product decisions, framing the upside scenario explicitly as a team exercise changes the decision calculus.

"I challenge my team, what if it all goes right? Entry price for any of these winners doesn't matter." — Rory O'Driscoll [00:25:30]

6. Overlooked Insights

The Real Rate-Limiting Factor for the Entire AI Wave Is Corporate Revenue Demand, Not Capital Supply

The hosts briefly surface this but don't fully dwell on it. NVIDIA won't blink. Hyperscalers won't blink. The IPOs of OpenAI and Anthropic will unlock the next turn of marginal public capital. But the actual question — whether corporate America can absorb $200-400B of AI spend in revenue terms by 2027-2028 — is the single variable that determines when the cycle ends. This is more actionable than anything else discussed: investors and operators should be tracking enterprise AI adoption rates (not just hype) as the true leading indicator of cycle health.

"The only two things that stop it are you run out of capital or you run out of demand... Anthropic is talking about 200 billion of GAAP revenue in 2028. Is there 200, 300, 400 billion of demand for this stuff in corporate? That to me is the question that will determine when the train stops." — Jason Lemkin [00:53:52]

Even Failed Frontier Model Bets Produce Enormous Returns — Which Means Every Non-Frontier AI Company Is Being Systematically Undervalued Relative to Its Acquirer Value

This point was made quickly but has enormous implications. Poolside "failed" to raise capital for its next training run — it could not sustain itself as a standalone — and still returned 15x to seed investors and $1B+ to founders. The structural implication: in a market where first prize is $1-2 trillion, fifth prize is still $9 billion. Any company that moves meaningfully in the right direction in AI, even if it cannot achieve standalone viability, has enormous embedded acquisition value to the four or five entities that can actually deploy the capital. This creates a different risk calculus for AI investing than any prior technology wave — the floor on well-executed but undercapitalized AI companies is dramatically higher than it appears.

"If first prize is a trillion dollars, turns out fifth prize is still $9 billion... The potential was there for that bet, which is why at the time it was a rational bet. And then the capital markets were such that you couldn't get it, and you still got a 15x." — Jason Lemkin [00:14:50]