Jason Lemkin | 20VC Summary

1. Key Themes

The Chinese Open-Weight Model Threat Is Real But Nuanced

China shipping near-frontier open-weight models (Kimi K2, Qwen from Alibaba) is not new news, but it is accelerating. Half of OpenRouter traffic already runs through China-created models. The significance is the potential scale shift from niche tech-forward users to mainstream adoption.

"If you look at OpenRouter data, half the traffic's through China-created models... you know, in a year could be everybody. And that's material." — Jason Lemkin 00:06:11

"Kimi K2 is, I think, a 2.8 trillion parameter model. It's a huge honking thing... much more comparable in size and therefore in terms of compute capacity to some of the US frontier models." — Rory O'Driscoll 00:07:27

The Agentic Token Consumption Explosion Is Driving Model Cost Pressure

Token usage at regulated enterprises is exploding not from raw usage, but from agentic architectures requiring supervisor models to watch agents. This structural demand for cheaper equivalent models will only intensify.

"The reasons are having supervisor models track the agents so the agents don't mistake... the more regulated you are, the less forgiving you are of an error in an agent. And so it's like four times the agentic use just to have multiple agents regulating agents. If it's already grown that much in the first half of the year, the quest for equivalent models at a cheaper price is just going to keep going up." — Jason Lemkin 00:10:56

Everything In AI Ultimately Comes Down to OpenAI and Anthropic's Growth Rate

The entire AI ecosystem — hyperscaler CapEx, NVIDIA valuation, data center commitments, training data companies — is a derivative bet on whether these two foundation model companies can sustain their growth trajectory in 2026 and 2027.

"The only thing that matters is the OpenAI and Anthropic growth rate in 26 and 27." — Rory O'Driscoll 00:19:43

"If you're growing 10x year on year and you have any kind of positive and improving gross margin, it just covers all the nut. The minute that growth rate stops... if the only way you can keep that growth rate up is by lowering your price per, then your gross margins start to deteriorate instead of continuing to improve." — Rory O'Driscoll 00:46:46

The Infrastructure Layer Is Where All the Real Money Is — Apps Are a Rounding Error

The gap between infrastructure spend and application revenue is staggering and barely discussed. The entire AI application layer outside of Cursor is smaller than what Anthropic and OpenAI spend on training data alone.

"There's probably more money being spent on training data for the foundation models, you know, the Scale AI surges and that, than pretty much any app company outside of Cursor. In fact, probably the sum of all the apps companies outside of Cursor are probably less than the amount that Anthropic and OpenAI are spending on training data." — Rory O'Driscoll 00:36:53

"I'm waiting for the era of the application layer in AI and making bets and seeing some good stuff, but I don't believe it's here yet." — Jason Lemkin 00:35:19

Open-Weight Inference Providers Are the Unexpected Winners

Companies like Fireworks, Base10, Together, and Fly.io are benefiting massively because open-weight model adoption and inference demand are inseparable. Whatever compute they've already locked in is now being charged at premium rates, driving margin expansion.

"Whatever compute you own now, you can charge way more, which means what looked like a low-ish gross margin business has now probably become a very attractive business. So not only are they probably growing 5x to a billion, but they're probably growing 5x to a billion with expanding gross margins." — Rory O'Driscoll 00:33:52

Custom/Domain-Specific Models Will Relegate Generic Models to Prototyping Only

Jason's hands-on experience building an agentic recruiting app demonstrated that data labeling and domain-specific fine-tuning produces order-of-magnitude better outputs. Generic models will increasingly be for prototyping only.

"The outputs are just literally an order of magnitude better once you do it. So everyone wants your own model... Generic models are great, but it is amazing how much better you can do than them for any specific workflow. You can do epically better." — Jason Lemkin 00:39:31

"Every mediocre history professor, every mediocre doctor that doesn't even know what caused your runny nose is in the LLM. But if you get the best people training it on the best answers, it's a step function." — Jason Lemkin 00:40:30

The Stripe/PayPal Deal Is Already Pre-Scripted Theater

The board's rejection is a standard fiduciary move, not a real rejection. The 28% premium offer is deliberately below the mid-30s average for take-privates, creating room for the scripted negotiation to land around 35%.

"The average take private like this is in the mid-30s. Now, average does not control any deal, but that is the perfect amount of back and forth, 28 to 35. It's already prescripted." — Jason Lemkin 00:57:44

"A company buying something one and a half times its size for... a lot less of its market cap. If they pull it off, they will look back and go, wow, that was an amazing deal." — Rory O'Driscoll 00:51:06

Late-Stage Growth Investing Has Been the Best Risk-Adjusted Play in Venture

Series B/growth rounds have been massively underpriced relative to Series A valuations on an absolute multiple basis. Early-stage prices are richly valued while late-stage companies with proven revenue trade at surprising discounts.

"Series A, $2 million in revenue at $300 million price, which is the going rate for a hot AI company at Series A... or would you rather stick money into Fireworks, which says they're going to hit $2 billion by the end of this year, at $17.5 billion? You're paying less than 10x revenue." — Harry Stebbings 00:03:24

"Growth for the last two or three years has been a very attractive place to make money. Those kinds of deals at one, two, and three billion have been subsequently marked up a lot." — Rory O'Driscoll 01:05:30

2. Contrarian Perspectives

Banning Chinese Open-Weight Models Would Be Terrible Policy

The conventional wisdom in DC (and pushed by OpenAI's own policy team) is that Chinese models are a national security threat that should be restricted. The contrarian view — held by David Sachs, Emil Michael, and Bill Gurley simultaneously, an unlikely coalition — is that a blanket ban would harm American startups with zero justification.

"If you're a Decagon and you're a startup doing inference on customer support queries for a very boring consumer product, there is no reason why you should pay marquee prices when something 10x cheaper is available. And it would be horribly bad policy to ban that." — Rory O'Driscoll 00:12:54

"The Chinese administration are talking about preventing those companies from selling those models to the US... We think they're trying to sell it to us and we don't want to buy it and they think they shouldn't be selling it to us because it's so powerful. So we can, that's kind of just weird in and of itself." — Rory O'Driscoll 00:17:06

Founders Should NOT Sell at 3x Their Last Round — Even for $600 Million

The VC framing of "should we sell at 6 billion?" is completely wrong for founders. Jason argues the relevant threshold is whether the outcome is 10x, because the risk, sweat, and time of the next three years is not worth a 3x gain regardless of absolute dollar size.

"Going from 40,000 in the bank to 400 million versus 800 million. It's irrelevant if there's risk... 3x not good enough, man. Got to be 10x." — Jason Lemkin 00:26:56

"If you turn down a big ass offer, you better be sure it can be way bigger. You better have high certainty and high biggerness." — Rory O'Driscoll 00:28:02

The Application Layer in AI Is Not Here Yet — Infrastructure Is Everything

The prevailing narrative is that AI applications will be the big winners. The data says otherwise: the entire AI application layer (excluding Cursor) generates less revenue than what OpenAI and Anthropic spend just on training data.

"All the good investments sure seem to be in the infrastructure. Absolutely. Even the ones that look good in software, the numbers pale in comparison anyway, right? The absolute numbers pale. So I'm waiting for the era of the application layer in AI and making bets and seeing some good stuff, but I don't believe it's here yet." — Jason Lemkin 00:35:19

On-Prem Hosting Does NOT Adequately Resolve Chinese Model Security Risk

The common reassurance from vendors is that running open-weight Chinese models on-prem eliminates data exfiltration risk. Jason argues that no credible independent authority has certified this, and the history of Chinese products in sensitive contexts means CIOs will not and should not accept this reassurance.

"Who has said my data is not being exported through the most complicated borderline self-aware software of our lifetimes? Who can say that... I don't want to take this risk, CIO of some Fortune 500 Global 2000 company unless everyone... I don't know, man." — Jason Lemkin 00:14:52

Tranche Rounds Primarily Benefit Founders in Bragging Rights, Not Financially

The Sequoia-style tranche rounds (pricing early tranches cheaply and later tranches at big step-ups) appear founder-friendly but are actually a mechanism for VCs to extract returns that would otherwise go to the secondary market, while creating 409A complexity and second-class shareholder dynamics.

"Nature abhors a vacuum and Sequoia abhors leaving a dollar on the table. So what's happening is people are realizing everyone wants these growth rounds... now what we can do is do this tranche structure and effectively price the excess return away from Harry and back to us." — Rory O'Driscoll 00:06:25

3. Companies Identified

OpenRouter

A model routing layer that aggregates access to multiple LLMs; reportedly in acquisition talks at $5–6 billion. Mentioned as a company that was early, built real traction in a market now going mainstream, but faces commodification risk as routing gets embedded in adjacent platforms like Databricks. The hosts agree now is the optimal moment to sell.

"I think it's a great time for OpenRouter to sell. I think them leaking the story, it was very savvy... it's still a niche product that more and more people are going to build variants of themselves." — Jason Lemkin 00:20:43

"The NPV of the company on a standalone basis is, you know, a couple of billion, not huge. But the value to it right now to a hyperscaler, if they could shift 10% market share in the enterprise to them over the next half a decade by saying, dude, we are the cloud, we are the model agnostic people who will make it easy, could be interesting." — Rory O'Driscoll 00:23:46

Fireworks AI

US-based inference provider. Crossed $1 billion ARR in three and a half years. Raised $1.5 billion from Index, Gavin Baker, Lightspeed, and 20VC. Processing 40 trillion tokens per day, up from 15. Gross margins at mid-30s with a path to higher as they move into owning their own data center infrastructure. Expects to reach $2 billion ARR by end of year.

"Fireworks, a leading inference provider announced their latest round, which was a one and a half billion dollar round... Doing over a billion in ARR. Got there in three and a half years. And they announced around 40 trillion tokens a day up from 15." — Harry Stebbings 00:31:22

"Not only are they probably growing 5x to a billion, but they're probably growing 5x to a billion with expanding gross margins." — Rory O'Driscoll 00:33:52

Databricks

Data and AI platform raising $3 billion Series M at a $188 billion valuation. Cited as an example of a mature, cash-flow-positive company that should be public but remains private, and as a platform that will build its own model gateway/routing functionality, threatening OpenRouter's standalone value.

"Databricks raising $3 billion, a Series M, I love this, a Series M, at a $188 billion valuation." — Harry Stebbings 00:01:03

"If you're on adjacent platforms, if you're using Databricks Gateway, they'll have their own harness. They'll figure this out for you." — Jason Lemkin 00:21:10

Stripe

Private payments company valued at ~$150 billion, processing ~$1.9 trillion annually, ~$6 billion net revenue. Making a joint bid with Advent International to acquire PayPal at a 28% premium as a take-private. Praised as one of the best-run companies in tech with a pristine culture and reputation.

"They both process kind of $1.9, $1.8 trillion a year... Stripe is valued at like $150 billion... sub 10 times profits... a chance to really, you know, transform and double your footprint." — Rory O'Driscoll 00:50:10

PayPal

Legacy payments giant processing ~$1.9 trillion annually, ~~$30 billion gross revenue. Trading at a depressed valuation (~~$50 billion), described as having squandered opportunities since the PayPal mafia departed with a revolving door of executives. Subject to Stripe/Advent take-private bid.

"Ever since the PayPal mafia walked out has been just a revolving door of executives and is a real mess, and they've dissipated their opportunities." — Rory O'Driscoll 00:53:36

Harvey

Legal AI company. Cited as an example of a scaled application company that has built its own proprietary model — the benchmark for what serious AI application companies must do to escape generic model dependency.

"You want your big M model. As soon as you're at a certain amount of scale... You are going to want to have your own model, right? Like a Harvey or Cursor." — Jason Lemkin 00:39:00

"Adding two, three hundred million, right? It's amazing." — Rory O'Driscoll 00:36:23

Cursor

AI coding assistant, cited as having been acquired by Anysphere/OpenAI at a ~$60 billion valuation. Called out as the defining AI application success story and the benchmark all other app companies are measured against.

"Probably why Cursor wasn't dumb to sell at 60 billion." — Jason Lemkin 00:30:34

"You start with Cursor at four, because I think coding is an app." — Rory O'Driscoll 00:36:23

Scale AI (Surge)

Training data and data labeling company. Cited as generating ~$3 billion in revenue, making it larger than virtually any AI application company and rivaling the foundation models as a revenue center in the AI stack.

"Scale AI Surge, another billion. You get to four or five billion." — Rory O'Driscoll 00:37:16

Moonshot AI (Kimi)

Chinese AI company behind the Kimi K2 model — a 2.8 trillion parameter near-frontier open-weight model. Generated so much demand post-announcement that new consumer signups were blocked. Was being pitched to Harry at a $20 billion valuation via SPV.

"We can't even sign up new as consumers for Kimmy because it's blocked. They have so much demand since this happened." — Jason Lemkin 00:05:42

TSMC

Taiwan Semiconductor Manufacturing Company. Highlighted for its model of long-term supply chain cooperation with NVIDIA (no written contract, 30-year relationship) and ASML — contrasted sharply with the adversarial DRAM market dynamic.

"NVIDIA is now TSMC's largest customer... no one in that entire supply chain has ruthlessly gouged each other... it's a real, hey, we know we're going to be dealing with each other for 10, 20 more years, trusted relationships." — Rory O'Driscoll 00:15:04

ASML

Dutch semiconductor equipment company making the EUV lithography machines that enable TSMC's chip production. Praised alongside TSMC for rational, long-term pricing behavior contrasted with DRAM suppliers.

"ASML makes the machine that enables TSMC, and TSMC makes the wafers that makes NVIDIA." — Rory O'Driscoll 00:15:31

Valor Atomics

Private nuclear energy company. About to raise at a 3x step-up in valuation in just four to five months. Cited as an example of a frontier-tech company that should remain private (contrasted with nuclear SPAC companies that went public prematurely).

"Valor Atomics looking like they're about to raise at a 3x step-up in four or five months." — Rory O'Driscoll 01:12:19

Thinking Machines Lab

AI lab reportedly announcing a model called "Inkling." Cited as one of the few potential US competitors that could fill the open-weight model gap but has yet to do so aggressively.

"Thinking machines had an announcement last week. I think they announced a model... I think they made a comment on something that you can build upon. Inkling, I think it was called." — Rory O'Driscoll 00:17:54

Ramp

Corporate spend management company that launched its own model routing/LLM gateway product competitive with OpenRouter. Cited as an example of how routing functionality is getting embedded in adjacent platforms, threatening OpenRouter's standalone value.

"We have Ramp releasing an OpenRouter competitor, just as OpenRouter are supposedly about to get bought." — Harry Stebbings 00:01:03

Base10 Partners

US-based inference provider alongside Fireworks and Together. Cited as a beneficiary of the open-weight model trend.

"If I'm someone like who makes my money as a hyperscaler hosting... if you're Base10, if you're Fly.io, if you're Fireworks, if you're Together, this is your market and your moment." — Rory O'Driscoll 00:32:00

Decagon

AI customer support company. Used as the example of why banning Chinese models would be bad policy — a startup doing commodity inference work should not be forced to pay premium US model prices.

"If you're a Decagon and you're a startup doing inference on customer support queries for a very boring consumer product, there is no reason why you should pay marquee prices when something 10x cheaper is available." — Rory O'Driscoll 00:12:54

Advent International

PE firm co-bidding with Stripe on the PayPal take-private. Speculated to be used as a mechanism to keep parts of the deal off Stripe's balance sheet during the rationalization period.

"They're doing a joint deal with Advent, a PE provider... maybe they're using Advent to almost keep it slightly off balance sheet for a period of time while they rationalize it." — Rory O'Driscoll 00:54:05

Corweave

GPU cloud provider mentioned as having been depressed in stock price partly because memory/DRAM prices rising 2x has increased the cost of building AI compute products.

"One of the things no one ever says is the fact that memory prices, the cost of building the product you're trying to build has gone up by 2x because the suppliers are charging you more." — Rory O'Driscoll 01:17:15

CyrusOne

Data center company that IPO'd at a $3 billion market cap on ~$1 billion ARR growing at 16%. Cited as a cautionary example of a company using an "AI veneer" that markets saw through — no multiple uplift, no growth acceleration.

"If you kind of reach this slow growth and you put a veneer and a wrap around it, it's still growing at a billion in revenue... The lesson for me to CyrusOne is you got to deliver." — Jason Lemkin 01:13:51

4. People Identified

Lynn (CEO, Fireworks AI)

CEO of Fireworks AI. Built company from zero to $1 billion+ ARR in three and a half years. Articulated clear vision for vertical integration into data center infrastructure. Expects to double revenue to $2 billion by end of year. Margins at mid-30s expanding.

"Lynn said specifically that they were at mid-30s in margins, and that would move up as they eat more of the stack and they do plan to move into the data center layer themselves." — Harry Stebbings 00:34:14

"Lynn said in the show, the future would be every company having specialized models with their own data." — Harry Stebbings 00:37:57

Jesse Zhang

Cited for publishing real token consumption data showing ~2.5x growth in token use since January at highly regulated companies, driven by agentic architectures. Praised for providing actual data rather than opinion in the Chinese model debate.

"Jesse Zhang had a Twitter article today... it was pretty good. I think people might have missed it because it's real data, which is what I like. But he said, here's one of our most regulated companies... Just our token use here has gone up what looks to be about 2.5x since January." — Jason Lemkin 00:10:43

Dean Ball

Policy/communications director at OpenAI (started two weeks before this episode). Previously at the Trump administration on AI policy and the Hoover Institution. Triggered a firestorm by tweeting about Chinese models using the term "AI communism," which was criticized as both hysterical and nakedly self-interested given his employer's direct financial interest in restricting Chinese model access.

"He set off a firestorm with the tweet... He used the word AI communism and it was very over exaggerated... when you start even hinting about significant regulatory... changes that will massively benefit you, you got to expect that everyone's going to say, dude, of course you're going to say that." — Rory O'Driscoll 00:08:56

Emil Michael

Former Uber executive, currently at the Defense Department. Got publicly sideways with Anthropic. Came down against banning Chinese models — notable because he is a famous Bill Gurley antagonist, making his alignment with Gurley on this issue a strong signal of the correct position.

"The second one was Emil Michael... who was a guy at the Defense Department who got totally sideways with Anthropic... what I like about that guy is that man knows how to hate. And one of his biggest hates for the last decade and a half has, of course, been Bill Gurley... So if Bill and Emil are on the same side saying, don't ban these models, then you've got to know that there's got to be some truth in that." — Rory O'Driscoll 00:12:04

David Sachs

Former US AI Czar. Came out against banning Chinese open-weight models, calling it rubbish. Cited as an authoritative voice on the correct policy position.

"The first was David Sachs, the former AI Czar, who basically said this is rubbish, stop." — Rory O'Driscoll 00:12:04

Bill Gurley

Benchmark Capital partner. Publicly argued for free markets in Chinese model access, against banning. Praised by Jason as having ~30 IQ points on him and someone Jason always learns from even in his grouchiest moments.

"Rich, grouchy billionaire. Grouchy Bill Gurley's got probably got 30 IQ points on me. OK, and he's seen it all. Right. And even his grouchiest point, I learned something from." — Jason Lemkin 00:13:38

Gavin Baker

Investor (Atreides Management). One of the lead investors in Fireworks' $1.5 billion round. Noted for a framework he called "cross-sectional comparisons" — arguing that whatever valuation assumptions you make for NVIDIA should be consistently applied to all beneficiaries in the same supply chain.

"I think Gavin Baker had a very interesting term... He was basically saying, whatever assumptions you make to value NVIDIA about the future... you should make roughly the same assumptions in valuing the DRAM providers, in valuing all the other beneficiaries of that." — Rory O'Driscoll 01:18:55

Peter Thiel

Co-founder of PayPal, early investor in Stripe. Cited for the full-circle narrative: early PayPal mafia member who invested early in Stripe is now effectively "buying PayPal back" through Stripe's acquisition bid.

"One of the very early Stripe rounds, I know Peter Thiel was an investor, Sam Altman, a number of the folks who were involved or connected with the PayPal mafia back in 2000, 2001, before they sold to eBay... and stuck early money into Stripe. And now 15 years later are having the joy of buying PayPal back." — Rory O'Driscoll 00:59:04

Sam Altman

CEO of OpenAI, early Stripe investor. Part of the PayPal mafia-connected group that seed-funded Stripe and is now watching those investments compound through a potential PayPal acquisition. Mentioned also as having a ~2% stake referenced in the context.

"All the early PayPal guys that did the pre-seed along with Sam Altman's 2%, they're going to do pretty well in the end." — Jason Lemkin 00:58:59

Ben Affleck

Actor turned entrepreneur. Made $587 million from selling his production company to Netflix — more than 10x the combined box office pay from his three highest-grossing films. Used as a humorous contrast to the podcast's hand-wringing over a $6 billion acquisition offer.

"We're like, how much did Ben Affleck sold it for, 500 and some odd million to Netflix? Our jaws drop and we're arguing whether we should sell a portfolio company for 6 billion." — Jason Lemkin 00:48:24

Brandon McCall

Cited for publicly criticizing Sequoia's tranche round structure on Twitter. Praised by Harry as brilliant marketing against a competitor's tactics.

"Brandon McCall has mouthed off, and I say that nicely, but mouthed off on Twitter about Sequoia's tranche rounds. I think it's brilliant marketing for Sequoia." — Harry Stebbings 01:04:41

5. Operating Insights

The "10x or Don't Sell" Founder Decision Framework

When founders face acquisition offers, the VC framing of IRR multiples is the wrong lens. The correct question is whether the business can realistically be 10x larger, not 2-3x. The personal financial delta between a 3x and 10x outcome is not worth the additional years of grinding risk.

"If it's 10x, if you know in your heart and soul you are building a company 10x bigger than this, right or wrong? Like, I don't know. Then F and say no and go for it. Here's a few more shares, in fact, friends. Let me reload you. But they only vest at 10x." — Jason Lemkin 00:28:25

Data Labeling With Domain Experts Is a Step-Function Performance Multiplier

Building any serious AI application with workflow-specific labeling data — even just 20-30 questions answered by subject matter experts — produces dramatically superior outputs versus relying on generic models. This is actionable immediately for any team building AI products.

"To really get it great, it needed labeling to make its... Now, I'm going to put model in quotes, right? It uses Sonnet and Opus. But so there's different definitions of model. And it was good. But man, once I started labeling all of this, it got exponentially better." — Jason Lemkin 00:38:31

Sell Into Commodification Windows — Not After

The optimal M&A exit timing is precisely when a space begins commodifying but before the market has fully priced in the commodification. Acquirers are still motivated, valuations are still elevated, and the company's strategic value to a buyer peaks at this inflection point.

"The perfect outcome is to sell the moment it becomes commoditized, but before everyone fully realizes it. That's when they'll give you the money, but that's before the value decreases rather than it accretes." — Jason Lemkin 00:30:12

When a Board Rejects an Acquisition Offer, It Means They're Going to Accept It

For public companies, the initial rejection is a fiduciary and legal requirement, not a real signal of intent. The moment a board retains bankers after a rejection, Delaware business judgment law creates a process that almost always ends in acceptance at a modestly higher price.

"The board rejected it. And the fact that the board rejected it means to me that they're going to accept it. You reject it because no investment bank will tell you you're allowed to make your highest offer up front." — Jason Lemkin 00:54:40

Tranche Rounds Give Bragging Rights but Create Cap Table Complexity — Use Them Deliberately

Tranche rounds (pricing early tranches cheap, later tranches at big step-ups) are not harmful to founders if the second-tranche investors don't care about their 1x position. But they complicate 409A valuations, create second-class shareholder dynamics, and require a drag-along. Use only when the headline valuation serves a strategic competitive signaling purpose.

"It does give you bragging rights... if I can come out and say I've raised at 5 billion with Sequoia leading, it will create fear among other VCs to fund competitors." — Harry Stebbings 01:10:43

"In the litany of mistakes that you can make with your cap table, doing a two-tranche round that makes all your second-tranche people feel like second-class citizens, it's not the worst thing in the world, provided you don't give a damn that they're second-class citizens." — Rory O'Driscoll 01:11:31

6. Overlooked Insights

The DRAM Market Is in an Unstable Pricing Bubble That Will Violently Correct

Rory briefly mentioned the DRAM market dynamic in contrast to TSMC/ASML, but the real insight is buried: SK Hynix, Samsung, and Micron are raising prices 40% per quarter on memory chips because they have oligopoly pricing power in a demand spike. This is explicitly building toward a brutal correction — and when it breaks, it will cascade through every company in the AI supply chain that has baked in current memory cost assumptions.

"The DRAM guys are like, screw you. We're raising prices 40% this quarter. Oh, next quarter, you still need our stuff. Raising another 40%, right?... When that pricing breaks, it'll be brutal to the downside. But maybe that's a year, two years from now." — Rory O'Driscoll 01:16:00

This is directly impacting CoreWeave's stock depression today (memory costs doubled), and will impact any hyperscaler or inference company that hasn't locked in supply. No one in the AI conversation is focusing on DRAM as a systemic risk vector — but Rory flagged it and immediately moved on.

The US Has No Competitive Open-Weight Model and Nobody Has Explained Why

The panel explicitly noted that four or five US companies (Google, Meta, Thinking Machines, Reflection) have the capability to build competitive open-weight models, and none of them are doing so aggressively — while five Chinese companies are sprinting. The panel glossed over this as a question without an answer, but the likely explanation Rory floated is actually explosive: the US companies' comparative advantage depends on distillation from frontier models, which Chinese companies can do with US models but US companies cannot legally do with Chinese models. This creates a structural asymmetry that explains the entire gap.

"Maybe it's because the dirty little secret is a lot of their advantage is distillation, which you can't legally do if you're US-based." — Rory O'Driscoll 00:18:38

"There's this new category called LLM Intelligence, two companies in existence as premium products. Their combined market cap is $2 trillion... There are four or five other companies in the US that are capable and have proven their ability to build something roughly comparable. None of them are taking advantage of this. And there's five Chinese companies that have proven their ability to build something roughly comparable, and they're cranking night and day to take advantage of it." — Rory O'Driscoll 00:19:33