Jason Lemkin | 20VC Summary

1. Key Themes

The Talent War Is Existential for AI Labs — and Anthropic Is Winning It

The departure of Noam Shazeer (Character AI) and John Jumper (Nobel Prize winner, AlphaFold co-creator) from Google DeepMind to Anthropic within 48 hours signals a momentum shift. Top researchers want unconstrained research environments, and Anthropic is threading the needle between frontier research freedom and product execution — something Google cannot offer while defending an existing business.

"The amazing thing about someone like Anthropic is they're able to do both. They're able to create these wonderful research environments for the leave me alone, let me do research people while at the same time, they're obviously executing tactically brilliantly in terms of product, making stuff happen. That's momentum. When you start winning, everything starts going your way." — Rory O'Driscoll 00:13:31

Open Source Is a Geopolitical Subsidy, Not a Market Phenomenon

The open source AI ecosystem is largely funded by the Chinese government as a sovereignty play — not by idealistic developers. DeepSeek's $7.4B Series A at a $50B valuation, where only the Chinese government receives voting rights, makes this explicit. This "free" competition is actually a state-funded assault on US closed-source model economics.

"Open source is a bit of a fake because China's paying for all the training... It is an existential sovereignty thing. The Chinese do not want to be reliant on Anthropic and OpenAI to run the next generation economy. It's very smart. And whatever investment the government has quietly made to subsidize all of these providers, it's a drop in the bucket at the sovereign level. What does an aircraft carrier cost? A lot. It's easier just to subsidize an open source model and pretend the training costs are 10 million." — Jason Lemkin 00:16:55 and 00:22:55

The $725 Billion Math Requires 7-8% of the US Labor Force to Be Replaced

Goldman Sachs projects $7.6 trillion in cumulative AI CapEx from 2026–2031. Rounding to $1 trillion in annual spend (CapEx plus electricity), and given that total US labor spend is sub-$20 trillion, the revenue math only works if AI displaces a large share of white-collar labor. This is not speculation — it's arithmetic.

"You're really talking about 7-8% of the labor force being replaced by tokens for the math to work. When you look at that, you go, it's a dauntingly high bar." — Rory O'Driscoll 00:38:00

The "Flabby Middle" of Closed-Source AI Is the New Existential Risk

OpenAI and Anthropic dominate at the frontier (Opus, GPT-5.5) and at the low end (Haiku, Mini), but the mid-tier workload market — too expensive for frontier models, too important for toy models — is being hollowed out by Chinese open source alternatives that are roughly 2x cheaper. This threatens their revenue just as they prepare for IPOs.

"I firmly believe we need these middle level models from the closed source providers for the business to work. Not everyone can afford to run every workflow on Opus, and Haiku and Mini are too small. This is where open source is going to disrupt the market... just when they're all ready to go IPO, they have a brand new existential risk, the flabby middle." — Jason Lemkin 00:19:19

The 2027 ROI Reckoning: Token Maxing Era Is Ending

2025 and early 2026 were the "just go build it" era — CIOs handed out AI budgets without accountability. By 2027, enterprise buyers will demand hard ROI. Departments that can show measurable output — headcount reduction, revenue growth — get more tokens; those that can't get cut off. This is a structural shift in how AI spend gets justified.

"I think the big story of 2027 in AI and the enterprise is show me the ROI next year. If you have ROI, if you were able to lay off 20% of your department, or you have the highest growing division in our company, we will give you more tokens. And this group that can't ship or can't get anything done or is in decline, we're just not going to give you the tokens." — Jason Lemkin 00:36:52

Seats Are Dead — Variable Pricing Is the Only Model That Survives

Every company selling seats is being crushed in public markets. Every company selling usage-based or variable pricing is winning. This isn't coincidental — seat-based models cannot reflect the economic reality of AI-driven productivity, while variable models capture the upside.

"When I kind of did the other day, I was just trying to do my own little analysis to oversimplify the public markets — everyone selling seats for the most part is getting crushed. Everyone selling variably, one way or the other is winning." — Jason Lemkin 01:03:17

"There's only one thing worse than a seat-based model, Jason, and that's a model that's based on bodies. If your business as Accenture is I bill out a hundred people, I pay them 200 grand a year, I bill them out at 500 grand — if I only need 40 people and some AI, that makes my head hurt." — Rory O'Driscoll 01:05:56

LLM Lifting Is Destroying Switching Cost Moats

The historical moat of data lock-in and switching costs is being dismantled by LLM-powered data migration. Databricks promises a full data lift in 30 days for any customer. Salesforce migrated SaaStr off Marketo in weeks with no human involvement. These are not edge cases — they are a structural change in the competitive dynamics of enterprise software.

"Your moat can be LLM lifted away... it's a moat destroyer when LLMs will lift you from one vendor to the other. I literally was doing a pitch this week and the founder was going on and on about their moats and I immediately didn't want to invest." — Jason Lemkin 01:05:17

AI for Systems Integration (SI) Is a Massive Underexplored Investment Theme

The consulting and systems integration market — led by Accenture — is being disrupted from below by AI-native SI companies. The work (requirements gathering, SOPs, deployment scripting) is precisely what LLMs excel at. Accenture's 19% single-day drop (40% YTD) is the canary. New entrants can bid 15-20% of what incumbents charge.

"We think that's a super interesting place to invest. There are companies like Tessera and Conduct doing AI systems integration for SAP deployment. Companies like Swantide doing it for Salesforce. In all these markets, Accenture was probably billing 20-40 million dollars for an SAP implementation. And today they might only get 20 million because 20 million of that can be done using LLMs." — Rory O'Driscoll 01:01:36

The New Startup Operating Model: Small, Office-Based, Relentless

The optimal startup structure has fundamentally changed. Winners are running small, highly-paid, equity-rich teams working 6-7 days a week in-office. The companies built on WFH cultures with bloated headcounts cannot compete with focused, AI-augmented small teams. This is not a management preference — it's a survival requirement.

"I want small high paid teams that work in the office over six days a week. That's what I'm interested in investing in. It's not because I don't have empathy. It's because they're going to fail... You don't get to make 10 million for working 18 hours a week. You get a watch. You want an Omega or you want to be rich? Make your choice, boys." — Jason Lemkin 01:09:41 and 01:11:09

2. Contrarian Perspectives

Anthropic's Prompt Caching Push Is Actually a Defensive Move Against Open Source — Not a Customer Benefit

Anthropic sent unprompted emails to customers about low prompt cache hit rates. On the surface it looks like a helpful cost optimization tip. In reality it is Anthropic competing on price against open source inference by making their product cost-competitive — not against other closed-source labs. The email is a war dispatch, not a service notice.

"This is Anthropic's shot across the bow for open source, trying to get you to cache your prompts, which are very expensive. They offer such a massive discount on cached prompts that it can actually be cheaper than open source... This is an email they sent saying, cache your prompts so it's cheaper than open source. That's why I think it's just hard for number three." — Jason Lemkin 00:17:49

The US Government Is Behaving More Like China Than Anyone Wants to Admit

While everyone criticizes China's state control of DeepSeek, Anthropic was blocked from shipping its most recent model until it satisfied US government concerns. The pattern of national governments taking control of frontier AI companies is identical — just dressed differently.

"We can't give China grief for this kind of thing when we're doing the same thing ourselves. Both governments are feeling this technology is pretty existential and trying to figure out, do you regulate it? Do you take over it? Do you stop it from being used by other people?... I got to give Leopold his 2024 situational awareness credit here — he called it that about now national governments are getting involved." — Rory O'Driscoll 00:25:53

OpenAI Building the Jalapeno Chip Is the Wrong Move at the Wrong Time

Vertical integration two levels down the stack — past data centers into chips — is a massive distraction from the actual battle OpenAI needs to win: enterprise customer acquisition and defending against open source in the mid-market. Every hyperscaler and chip company is already competing to provide cheap compute to OpenAI. Building your own chip when you could just leverage that competition is strategically questionable.

"OpenAI and Anthropic have discovered the single best tech market in terms of consumer demand in the last 20 years. And they should put all their effort into meeting that demand. Vertically integrating backwards down the stack two levels down doesn't strike me as the highest and best use of resources. You've got Oracle, Google, Microsoft, CoreWeave — a whole bunch of vendors dying to do business with you. The whole reason OpenAI and Anthropic models work is because other idiots have spent the $300 billion on their behalf." — Rory O'Driscoll 01:14:29 and 01:16:19

Productivity Gains from AI May Never Show Up in Profitability — for Anyone

Even if AI genuinely makes every white-collar worker 20% more productive, if every company adopts it simultaneously, the productivity gain competes away entirely. No company will be more profitable; they'll just be forced to cut costs to match competitors who also cut costs. The only companies that win are the AI vendors — not the adopters.

"You can have an improvement in productivity in industry by virtue of an enabling technology like AI, and no improvement in profitability if everyone adopts the technology. If every bank adopted ATMs at roughly the same time, the profit didn't change. The same could happen here. Everyone adopts AI and everyone's cost function reduces by the same amount. Which is why you have to lean in." — Rory O'Driscoll 00:39:53

Number Three in Closed-Source AI Has No Viable Path — Even With Google's Balance Sheet

The standard VC wisdom is that Google is fine at #3 because it has infinite capital. The contrarian case: open source is so heavily subsidized by China that it's functionally "free," and this ceiling makes the #3 position uneconomical regardless of backing. Google cloud survived as #3 by being cheapest. That lever doesn't exist in AI because open source is cheaper still.

"If there is a compelling alternative to this entire set of competitors that's 5x cheaper, which is what open source is, then it's going to grind everyone down. What tends to happen is the number one guy makes a little less money, the number two guy makes quite a lot less money, and the number three guy goes bust. In this case, obviously not bust because they've got Google behind them, but it's a real powerful downward pressure on profitability." — Rory O'Driscoll 00:19:23

3. Companies Identified

Anthropic Leading closed-source AI lab. Mentioned as the momentum winner in talent, research environment, and product execution — recruiting Nobel Prize winners and attention paper authors away from Google DeepMind, while also sending aggressive prompt caching emails to compete on price with open source.

"Anthropic is able to create these wonderful research environments for the leave me alone, let me do research people while at the same time executing tactically brilliantly in terms of product." — Rory O'Driscoll 00:13:31

DeepSeek Chinese open source AI company. Raised $7.4B at a $50B valuation; founder committing ~$3B personally; fewer than 10 investors including JD.com; Chinese government retains sole voting rights. Intentionally crippled within China (no web search, different training data). Viewed as a state sovereignty project, not a commercial enterprise.

"The government's the only one getting voting shares. It says all you need to know. It is an existential sovereignty thing." — Jason Lemkin 00:22:29

Menlo Ventures 50-year-old VC firm. Just raised a $3B fund. Named as one of the best AI investors of this wave via early Anthropic bet, plus investments in Agora and Lovable. Deliberately conservative fund size discussed as strategically smart — preserves ability to raise SPVs for outlier deals.

"They've done an amazing job with Anthropic and they've done an amazing job lasting 50 years." — Rory O'Driscoll 00:54:50

Databricks Data and AI platform. Growing ~80% at $6B ARR, accelerating. Mentioned for a highly specific and underreported capability: promising customers a full LLM-powered data lift in 30 days — compared to 5-year Accenture-led migrations previously.

"Databricks claimed they can do that lift in 30 days for their customers... compare that to a five-year lift with Accenture to go to Salesforce." — Jason Lemkin 01:04:48

Cursor AI coding environment. Cited as the canonical example of a company that won by grabbing market share even with negative gross margins during the first wave — and is now worth ~$60B as a result. Used as a template for when bad unit economics are the right bet.

"It may be that next generation they're seeing a little more focus on gross margins now. But there's no doubt that to date, the bet that my gross margins are shit but my growth will cover it — even though it sounds stupid when you say it — has in fact been 100% true. It paid Cursor to grab the ground." — Rory O'Driscoll 00:47:14

Kalshi Prediction market platform. At $2B revenue run rate, prepping for IPO. Achieved regulatory arbitrage by being classified as a prediction market (CFTC jurisdiction) rather than a sports betting house, allowing it to avoid state-by-state licensing. ~80-90% of volume is sports betting.

"Kalshi found a way to pretend it's a prediction market, found a way to get US jurisdiction from the CFTC. 90% of what they do is sports betting. They found a regulatory arbitrage to a wildly popular pursuit." — Rory O'Driscoll 00:55:33

Zhipu AI (Z.ai / GLM) Chinese AI company, publicly traded in China at ~$100B market cap (1,000x revenues). Its GLM 5.2 model beats GPT-5.5 on coding benchmarks. Cited as evidence that 6 Chinese open source models are at or near frontier US performance.

"Zhipu, Z.ai, is actually public now in China and it's trading at $100 billion or something like that, a thousand times revenues." — Rory O'Driscoll 00:24:57

Devv (Devon) AI company, up 30% YTD, up 3x from its IPO bottom. Named as one of the clear winners in the variable pricing vs. seat-based pricing split in public markets.

"Devon up 30% this year, one of the big winners... up three X from the bottom." — Jason Lemkin 01:03:17

Tessera / Conduct AI-native systems integration companies doing SAP deployments. Named by Rory O'Driscoll as examples of the "AI for SI" investment theme — bidding a fraction of Accenture's price by using LLMs to do the work.

"There are companies like Tessera and Conduct that are doing AI systems integration for SAP deployment." — Rory O'Driscoll 01:01:36

Swantide AI-native Salesforce integration company. Named as an example of AI displacing Accenture-style consulting for CRM deployments.

"AI systems integration for Salesforce integration, there are companies like Swantide and others doing that." — Rory O'Driscoll 01:01:36

CoreWeave GPU cloud provider. Named as one of the vendors competing aggressively to provide OpenAI and Anthropic with cheap compute — undercutting the rationale for OpenAI building its own chip.

"You've got Oracle, Google, Microsoft, CoreWeave — a whole bunch of vendors breaking their picks to provide you with cheap compute and taking on all the capital risk." — Rory O'Driscoll 01:15:26

Cerebras AI chip company. Stock dropped 16% on the day OpenAI announced the Jalapeno chip — its primary customer. Used as the real-time market signal that OpenAI's chip announcement is being taken seriously as a threat to third-party chip providers.

"Cerebras went public, had a decent quarter going really well, and their big traction is a $20 billion chip order from OpenAI. Did the stock go down? It went down 16%. So there you are. The market said, poor old Cerebras, OpenAI is going to build that chip instead." — Rory O'Driscoll 01:18:48

Bill.com Finance automation platform. Mentioned as one of the tools integrated into SaaStr's AI-built VP of Finance agent. Aurora (presumably a fund or investment vehicle) is noted as an investor.

"It logs into Bill.com where Aurora is an investor. It sends the invoice, it follows up on the invoice." — Jason Lemkin 00:41:33

Algolia Search and discovery platform. Nicholas Dessaigne, Algolia's former founder, now at Y Combinator, authored the insight that the most common reason good companies fail to raise Series A/B right now is margin, not growth.

"Nicholas Dessaigne at Y Combinator, formerly founder of Algolia — a most common reason I see good companies fail to raise their Series A/B right now isn't growth, it's margin." — Harry Stebbings 00:46:10

Perplexity AI search company. Named as a Framer customer (sponsor mention), but noted for brand recognition as a fast-moving AI company.

Eleven Labs AI voice platform. Named as a Deel customer, cited as an example of a fast-growing company trusting Deel for global hiring infrastructure.

4. People Identified

Noam Shazeer Co-author of the original Transformer ("Attention Is All You Need") paper. Previously left Google, co-founded Character AI, was acqui-hired back to Google for a reported ~$2 billion. Left Google again (with ~50% unvested) to rejoin the AI frontier. Departure signals Google's inability to retain foundational researchers.

"Noam obviously was at Google, did the original attention paper, left, did Character AI, got bought back to Google. A couple billion dollars. So a good slug of that was going to him. So assuming four-year vesting, he's probably left half on the table. That's a lot." — Rory O'Driscoll 00:07:47

John Jumper Nobel Prize winner (Medicine). Co-creator of AlphaFold at DeepMind. Joining Anthropic. First job outside academia was DeepMind. His move is interpreted as a signal of Anthropic's emerging next-generation science research initiative.

"He has a Nobel Prize for medicine and is still relatively young. These are not top one percent, they're top point zero zero zero one percent people." — Rory O'Driscoll 00:13:02

David Cahn (Sequoia) Sequoia Capital partner. Author of the "$600 billion question" piece on AI CapEx vs. revenue, which the hosts credit as prescient — and note that since publication, conviction and spending have only doubled.

"I thought the David Cahn piece a year ago was really good. And then you step back and since then, all that's happened is people have doubled their conviction and doubled their willingness to spend." — Rory O'Driscoll 00:31:18

Leopold Aschenbrenner Author of the 2024 "Situational Awareness" paper. Credited for correctly predicting that national governments would begin directly intervening in frontier AI companies by approximately this period.

"I got to give Leopold his 2024 situational awareness credit here. He called it that about now national governments are getting involved." — Rory O'Driscoll 00:25:53

Nicholas Dessaigne Former founder of Algolia, now at Y Combinator. Credited with articulating the insight that the #1 reason good companies are failing to raise Series A/B right now is gross margin — not growth rate.

"Nicholas Dessaigne at Y Combinator, formerly founder of Algolia, said: a most common reason I see good companies fail to raise their Series A/B right now isn't growth, it's margin." — Harry Stebbings 00:46:10

Andrej Karpathy Former OpenAI/Tesla AI researcher. Cited for claiming he moved from 20% to 80% AI-assisted coding capability in a 6-month period, used as evidence that capability curves in non-coding domains (legal, accounting) may follow a similarly rapid trajectory.

"Where Andrej Karpathy says I moved from 20% to 80% in a six months period — if you're able to see that same advancement in legal, whoa." — Harry Stebbings 00:40:49

Ryan Petersen CEO of Flexport. Made a viral claim ("work from home is white collar fraud") that prompted discussion. Used as a case study of an older-model startup CEO struggling to retrofit a distributed team for the intensity required to compete in 2025.

"Ryan Petersen at Flexport said, kind of jokingly, work from home is white collar fraud... I think he's struggling with trying to modernize his team and being competitive with the way startups are today." — Harry Stebbings / Jason Lemkin 01:07:26 and 01:09:18

Brandon from McCall (unverified name) Posted that "training agents will be the largest job category in five years time." Cited as a forward-looking claim about where human labor shifts — though Jason Lemkin draws a comparison to the discredited "prompt engineer" prediction from 60 weeks prior.

"Brandon from McCall posted nine hours ago: training agents will be the largest job category in five years time." — Harry Stebbings 00:43:00

5. Operating Insights

Build Your AI VP of Finance Before Hiring One

Jason Lemkin built a fully functional AI director of finance — remotely, from China, in a single-digit number of hours — using a conversational prompt connecting Bill.com, QuickBooks, Salesforce, and Brex. It outperformed every human on the team, caught unbilled revenue, and closed transactions end-to-end. The specific prompt was essentially: "Build me an AI VP of finance, connect it to Bill, QuickBooks, Salesforce, and our other apps, and automate all billing processes."

"We built an AI VP of finance while we were in China. It creates the quote, it builds the contract, it ships the contract, it gets it signed. It updates the opportunity in Salesforce. It logs into Bill.com, sends the invoice, follows it up, gets it paid... closes out the entire transaction. Then it logs into QuickBooks and for the first time in 10 years our books are accurate." — Jason Lemkin 00:41:05

Master Agents, Not Prompts — and Know Where They Go Lazy

The skill that matters now is not prompt engineering (that's worthless) but "agent mastery" — understanding where agents goal-seek, where they skip steps, where they get lazy. Specifically: Claude Sonnet rapidly goal-seeks and will skip complex behaviors. When an agent admits it "didn't fully read a contract," the right response is to redesign the agent's instructions to require full reads — not to abandon it.

"Being a master of agents is not throwing your monitor out the window when you hear that. It's like, okay, I get what happened. This is running on Sonnet. Sonnet rapidly goal seeks. It tries not to finish complicated behaviors. I need to work with the agent to change how we do it and make clear all contracts must be read from beginning to end." — Jason Lemkin 00:44:35

Use the Parity Tax Framework to Decide Which AI Investments Are Worth Making

Before projecting ROI on any AI initiative, apply the parity tax test: if every competitor will have the same tool within 12 months, the productivity gain is real but the profitability gain is zero. The correct framing for mandatory AI adoption isn't ROI — it's survival cost. Use this to prioritize investments that create differentiated capability over those that merely restore competitive parity.

"If you're the only one using a tool, you can achieve certain types of productivity. But then all my competitors deploy Anthropic and all of a sudden it doesn't show up because we have parity. We may reach a situation in 2027 where we just cannot prove any of these productivity gains. Everyone may just need to lay off 10% of their company in addition to all other layoffs." — Jason Lemkin 00:39:03

6. Overlooked Insights

Anthropic's Enterprise Inference Is Already Highly Profitable — And That Is Exactly What Open Source Is Targeting

This point was made briefly and almost in passing, but it is structurally significant. Anthropic's enterprise API customers — not the subsidized Max plan users, not the free tier — are generating 40-70% gross margins on inference alone. This is the target market for Chinese open source models. The "cheap open source" narrative is not aimed at hobbyists; it is a precision strike on Anthropic and OpenAI's highest-margin customer segment, right before their IPO windows.

"They have an enterprise base, which has between 40% and 70% gross margin just on inference... That's what open source is attacking, that enterprise customer, the lucrative customer." — Jason Lemkin 00:35:05

Sovereign AI Blockades Are Already Live — and Gemini Is the Unintentional Winner

Almost no one noticed this: OpenAI and Anthropic proactively block access from China and Hong Kong (intentional IP blocking, not Chinese government censorship). Gemini does not. This means that in the world's second-largest economy, Google Gemini is the default Western AI — by default, not by strategy, not by sales effort, not by product excellence. In a world of sovereign AI fragmentation, the company willing to operate in restricted geographies without political constraints may capture entire markets that its more cautious competitors have voluntarily abandoned.

"Anthropic and OpenAI don't serve there intentionally for security. They don't allow you to access Anthropic or OpenAI. It's not China — it's them blocking it. But Gemini doesn't. So when I'm in China for two weeks, it's DeepSeek and Gemini. Those are my friends. It's a parallel universe." — Jason Lemkin 00:21:35