Summary

NVIDIA is paying $6B to license Poolside’s “model factory,” investing another $1B at a $12B pre-money, and moving 109 engineers into Nemotron. Harry framed the deal as Poolside having no choice but to fail up after it could not raise $2B for 40,000 GPUs and would not have had compute into next year. Rory’s double lesson: it’s now “almost impossible to compete at the frontier,” yet bets that fail a standalone DCF still fetch compelling exits in a hyper-growth market — “if you get 15X on your failures in venture, you’ll die a rich man.” When first prize is a trillion, fifth prize is still nine billion.
Jason Lemkin’s counter is the sharpest line of the episode: “9 billion doesn’t clear the bar for seed investing in 2026.” With epic dilution the effective seed entry was ~$600M, so, absent a hyper-concentrated portfolio, best deals still need 50X — Poolside would have had to exit near $63B to make classic seed math work; Harry’s 15X seed return, real as it is, doesn’t return a fund.
NVIDIA’s playbook, in Jason’s framing, is to deploy a roughly $70–75B free-cash-flow budget one-for-one across its ecosystem, though Rory estimates free cash flow after CapEx at roughly $50B. Poolside is TAM expansion via a US open-source complement, Mercor’s General Catalyst-led $20B round is “straight investing,” and Perplexity’s $30B round and OpenAI checks are vendor financing. Rory’s caveat from the ‘02 telecom crash: vendor financing means you must be right every step, not just in the end — “that stuff comes back to you, and you look like an idiot in two years.”
OpenAI’s 2027 IPO announcement was forced, in Rory’s telling: Q1 and Q2 GAAP figures were roughly $5.xB and $6.xB, with 18% QoQ growth tracking under $30B this year, against Anthropic’s $60B mid-year run rate. That made the Q3-reacceleration leak “existential.” Ranking has flipped — “they’ll go out whatever price they get in ‘27, ‘cause they can’t wait any longer”; if Anthropic trades at $2T, maybe OpenAI gets $1.5T. The Junger line hangs over it: “if danger can be described as the absence of choice, they were now in danger.”
“It’s all about code.” That’s the only sentence that matters — Rory’s Yahoo-and-search analogy for why Anthropic lapped OpenAI: consumer was never the highest-ROI use of limited compute. Jason piles on: selling $10,000 of tokens for $200 is “one of the worst business models of our lifetimes,” and being number two against open-weight ankle biters (Vercel: 68% open weights) is brutal. Meanwhile Anthropic’s claimed $30 trillion TAM draws Rory’s scorn: “Your TAM is, you know, the entire U.S. GDP. Thanks a bunch, Dario.”
On the cycle, Jason says “you have to believe we’re less than a third of the way through” — cloud lasted nine years — while Rory argues supply won’t blink and capital has one more turn of the crank via the two mega-IPOs. The real rate limiter is end demand: whether corporate America can absorb $200–400B of spend fast enough to validate Anthropic’s $200B 2028 GAAP target.
The 2027 enterprise fight is token addiction meeting CFO math: Stripe’s letter frames intelligence “like capital — fungible… it has to be managed and allocated,” which means pricing it, allocating it, and cutting headcount elsewhere. The CFO’s twin terrors: automation that lowers EPS (“Wall Street is going to say, ‘You’re a fucking moron’”) and retention — “if 30% of my company leaves to go work for Harvey, I’m dead in the water.” Stripe itself accelerated to 41% growth with billings up 71%.
The trio’s overinflated categories: Harry names customer support (one or two winners, sophisticated buyers build their own) and humanoid robotics; Rory names defense, where Anduril-type consolidators will “hoover up the rest”; Jason names VC-funded law/accounting rollups that “make too much sense on a spreadsheet.” Rory’s disciplined coda: hold the bias but keep it overcomeable — “what if it works?”

Deep dive

1. Poolside “fails up” for $12B — and the frontier wall gets higher

The deal: NVIDIA licenses Poolside’s model factory for $6B, adds $1B at a $12B pre-money, and moves 109 engineers to Nemotron. The leaked investor letter that Harry found “almost depressing” tells the real story — Poolside couldn’t raise the $2B needed for 40,000 GPUs, would not have had compute into next year, and Harry said it “had no choice but to fail up.”
Rory reread the letter twice and keeps its best phrase: “we have found ourselves on the right side of prediction in a market that has scaled exponentially in terms of capital intensity.” Translation: they were right three years ago that a US open-source model had a market, built it, and the next turn of the model crank was simply beyond them.
His two-sided lesson: negatively, “the next smartest people who are really going for it just hit the capital wall,” so everyone behind them will too. Positively, in an exploding market, bets that can’t clear a standalone DCF still carry serious value to an acquirer — NVIDIA has capital, makes the GPUs, and “we’ll take it from here.”
Why NVIDIA wants it: an open-source model is a complement in the economic sense. Chinese open models are taking token volume; a viable US open model on NVIDIA chips shifts more spend toward chips relative to foundation-model builders. “They’re like, ‘Yay, team.’”

2. Is 15X on a failure a win? The seed-math argument

Harry, a Poolside seed investor, calls it “like a 15X for us… pretty great,” with founders making a billion each — while noting neo-labs are universally out of favor among investors he talks to. Jason won’t let it stand: “9 billion doesn’t clear the bar for seed investing in 2026.” A 15X off a $9B exit implies a ~$600M effective entry after epic dilution; unless hyper-concentrated, best deals still need 50X — roughly a $63B exit.
Rory’s rebuttal — worth keeping: this was the failure case, and “if you get 15X on your failures in venture, you’ll die a rich man.” The bet was rational because closed-source winners are worth a trillion and Chinese open-weight players $50–100B; the 100:1 outcome was genuinely on the table.
His broader point on who funds frontier models: “the VC money ran out on Anthropic and OpenAI long ago, which is why no VC owns more than 1% or 2% of either of them.” Only Microsoft, Google, Amazon — and now NVIDIA — can finance state-of-the-art, “except for Apple, who sticks their money in their pocket and just has the stock go up.” The golf analogy lands it: finish seventh at the US Open, collect $5 million, on to next week.

3. Mercor at $20B and NVIDIA’s ecosystem budget

The round: General Catalyst leading at $20B, NVIDIA reportedly joining in size, with Mercor at roughly $2–2.5B in revenue. Harry, an investor, admits he “never thought this would be as big as it has got as quickly.” Rory’s honest puzzlement: neoclouds and Poolside expand chip TAM directly; funding a data-labeling company doesn’t obviously sell more chips — “maybe it is as simple as we think it’s a good business at 20 billion and stop thinking about it, Rory.”
Jason floated free cash flow of roughly $70–75B as NVIDIA’s ecosystem budget, while acknowledging uncertainty. Rory later estimated free cash flow after CapEx at something like $50B, with gross profitability well over $100B. Jason’s model is that the strategy team and top VPs go around a room with best ideas, “and there’s some guy that thinks data labeling is important.” Adding cash to the balance sheet, other than defensively, does nothing for a profitable company’s CEO — “if Wall Street lets you get away with spending it, I would spend 100% of my cash too.”
A Kroll report backs the valuation logic: above 30% gross margins there is no additional M&A benefit, while deals below 30% face a penalty. So, conditionally, if Mercor can be valued at 8× with 80% gross margins and growth, it “isn’t expensive.”
Rory’s taxonomy of the NVIDIA moves: Poolside is TAM expansion via adjacency, Mercor is straight investing, and Perplexity’s $30B round and OpenAI are vendor financing — not nefarious, but the ‘02 telecom unwind is the memory: “if you overextend credit and underwrite projections that aren’t realistic, that stuff comes back to you, and you look like an idiot in two years.”

4. How big is data labeling really — and who escapes the margin trap

Harry’s bull case: the largest data providers become $200B companies — if OpenAI and Anthropic are $2–5T, is 10% of their market cap crazy for their core data supplier? Rory’s negative math: think revenue, not market cap — combined frontier revenue ~$100B, and if training is 5–10% of that, it’s a $5–10B market. Harry fires back with actuals: Surge at $3–3.5B, Mercor $2.5B, Handshake $1B, Micro1 $0.5B. Rory’s reconciliation: if frontier spend hits $400–500B, 5% would be $25B, divided four or five ways; under another framing, a healthy ~$20B pool divided three ways. The multiple attached is what really swings it.
The Cursor precedent, per Jason: negative gross margins, subsidized unlimited use, then caps, own models — “the kids are figuring it out, man” — to $60B. Rory as self-appointed Debbie Downer: survivorship bias is real (“I can think of plenty of deals, including some we’ve done, where you started with shitty gross margins and you ended with shitty gross margins”), and training companies’ three-to-five big customers may give them less room to improve margins than Cursor’s hundreds of thousands. Anthropic went from -91% to +30% margins in a year, so underwriting improvement can be rational.
Jason’s now-signature partner-meeting challenge: “What if it all goes right? Entry price for any of these winners doesn’t matter.” Harry’s gloss: “optimists make money and pessimists are right.”

5. OpenAI’s IPO isn’t a choice — it’s the absence of one

Sarah Friar told employees OpenAI will be public in 2027, and Rory argues the statement was compulsory. Q1 and Q2 GAAP figures were roughly $5.xB and $6.xB, representing 18% quarter-on-quarter growth and tracking under $30B this year off $12.5B last year — while Anthropic sits at a $60B mid-year run rate, bigger and growing faster. Two more years of that gap “and you’re in irrelevance,” and Broadcom and NVIDIA start doubting the $200B chip orders.
Hence the concerted leak of a Q3-reacceleration narrative: “there was simply no way the Q2 trend could stand unchallenged and still leave OpenAI as a credible close number two… That’s why it leaked, because it’s existential for them.” Rory’s caveat stays: “until you see it in GAAP numbers, it’s hard to be sure.”
On price, Rory refuses a number but insists on the ranking: “it will be lower than the other guys now.” And on timing: “they’ll go out whatever price they get in ‘27, ‘cause they can’t wait any longer” — if Anthropic trades at $2T they might get $1.5T; at $1T, maybe $700B. His touchstone is Sebastian Junger: “If danger can be described as the absence of choice, they were now in danger.”

6. Number two in the industry you invented — and the sentence that matters

Jason’s structural worry: at the start of the year there were two choices; now there are “seven choices for number two,” with open-weight models nearly equivalent on performance and routers letting you use 78 models. Harry cites Vercel’s data — 68% and rising open weights. Rory recants his initial pushback and agrees: number one says “you gotta just buy us ‘cause we’re number one”; number two must plead “please buy us as well and don’t buy the cheaper guys” — with “a whole bunch of ankle biters on top.”
Jason’s deeper question: “what is the differentiated mission of OpenAI today?” These were the most mission-based organizations of our lifetimes; now he can’t name it — while calling Dario “nutso a lot of the time” (reportedly asking candidates if they’d be happy joining Anthropic if it all went to zero — “cool question, actually, but nutso”) and conceding Sam is now “the more likable guy.”
One panelist’s awe at ChatGPT’s consumer penetration — “for large majorities of the general population around the world, AI is ChatGPT” — meets Jason’s brutal answer: “That was the plan. He just got lapped… it just was not the highest-ROI use for limited compute.” Selling $10,000 of tokens for $200 is “one of the worst business models of our lifetimes.”
Rory’s crystallization: sometimes only one sentence matters — for Yahoo it was “it’s all about search”; today “it’s all about code,” the fastest-adopting, highest-propensity-to-pay market. He still sees a Google-scale ad business possible over a decade — then spots Anthropic’s WSJ-reported $30 trillion TAM claim mid-recording: “Your TAM is, you know, the entire U.S. GDP. Thanks a bunch, Dario.”

7. Hugging Face at $13B: sell at the peak of the open-weights wave

Jason’s refusal, verbatim: “I’m not smart enough to understand why anyone would pay 13 billion for it.” Rory half-defends the strategic logic — revenue is only ~$150M relative to a potential $15B outcome, but every IT incumbent watching two labs claim GDP-sized TAMs is thinking “shit, I better get me something,” and Hugging Face is the access point for enterprise open-weight strategies. For a Microsoft or IBM, “absolutely” a super interesting asset — though he “can’t make head nor tail of the price.”
Jason’s timing call: “if you have an AI product that’s benefiting from the transition to open weights, there can’t be a better time to sell than plus or minus 90 days from today” — citing Elad Gil’s “sell if you have an AI asset” and OpenRouter growing 15% at roughly $150–200M when Stripe bought it. And any acquirer must not touch it: like TBPN becoming an OpenAI commercial, “even if you put a little ad at the top, you destroy it.”

8. Publics week: Griffin’s free money, leverage discipline, KOSPI whiplash

Ken Griffin’s Citadel already unwound 80% of Leopold Aschenbrenner’s 4× short book. Rory’s read: buy at 10% below market from weak hands, watch it jump, take the money — “every three years some idiot gives you three or four billion of free money, and you politely take it and put it into real estate in Miami. It’s good to be Ken.” The transferable lesson circles back to NVIDIA: without leverage you only need to be right in the end; with 4:1 leverage — or vendor financing — “you have to be right every step along the way.”
Jason on KOSPI as “AI on steroids”: still up 56.46% for the year despite the dagger from a peak around 3,000 in June to 2,600 on July 29 — Korea is now effectively “two memory providers with a bunch of other stuff attached,” and Samsung engineers are, per the Journal, the nation’s most eligible bachelors for the first time ever.
The wealth ripple: 50% of NVIDIA employees reportedly worth over $25M; mediocre Dogpatch one-bedrooms at $10K/month, meaning “$240K pre-tax to pay the rent… probably 480 to feel good about yourself.” Harry from London: “the money’s not here… that dispersion of wealth is just nil.”
Rory’s geography of the boom: California took three-quarters of total venture dollars (Anthropic and OpenAI alone probably ~60%), poured into a peninsula of ~780,000 people. There will be a correction — he’s seen ‘99 and ‘07 — “but it’s not going back to where it was, ‘cause it never does. It ratchets up.”

9. Does the floor fall out? Supply won’t blink; demand is the rate limiter

Jason’s cycle call: “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.” The shape is concentration — fewer people generating more revenue, concentrating exits, wealth, and the seemingly crazy salaries that “normalize” when a third of the headcount does the work.
Rory’s mental model — “the only question that counts”: on supply, nobody blinks; NVIDIA is not likely to get on an earnings call to say “compute demand has slowed down.” Only two things stop the train: capital or demand. On capital, “until the public markets get in on the game, you haven’t exhausted all the money that’s there” — crashes come when you run out of marginal buyers, so the two mega-IPOs are one more turn of the crank.
The real question is whether corporate America can spend $200–400B fast enough to feed the beast — Anthropic’s talked-about $200B of 2028 GAAP revenue is itself “a significant slowdown” from 10X-ing. Rory has no opinion yet on when demand caps out; today it’s all there in coding.

10. Token addiction meets the CFO: intelligence as capital

Jason’s arc from this year into next year: token-maxing and “performative AI” gave way to $20,000-per-employee bills and budget caps, and next year brings the backlash of addiction: “Take away my agents, I quit… I won’t edit your goddamn podcast, I won’t write your code, I won’t fix your endless bugs if I can’t have 10 agents running 24/7.”
Rory’s favorite document of the week is Stripe’s letter: “we’ve internalized that intelligence is like capital. It’s fungible, there’s demand for it, and it has to be managed and allocated” — unlike seat-based SaaS, uncapped intelligence spend must be governed like money, not like software licensing. A mid-size US bank making $1B a year cannot run up a $100M token bill and cut EPS 10%.
The nasty bit Rory won’t let Jason skip: if the best employees get 3X token budgets and revenue doesn’t double, “who are the other three people we need to let go?” You cannot tell Wall Street automation reduced profits — “Wall Street is going to say, ‘You’re a fucking moron.’”
Jason’s counterweight from CFO conversations: the empowered ones “talk about nothing but retention” — starve people of AI and “you’ll be stuck with the folks that are still AI skeptics… the moldy oldies of SaaS. Jesus Christ, if 30% of my company leaves to go work for Harvey, I’m dead in the water.” On diffusion pace, one panelist’s example was a girlfriend who went from calling Legora a joke to saying “I just verify documents” in six months; that panelist thinks legal has hit the tipping point. Jason adds Higgsfield, which went from not working a year ago to $700M revenue today.

11. Stripe at 41% growth resets the public-software leaderboard

The print: Stripe accelerating to 41% at massive scale, billings up 71%. Jason’s frame: it’s become a derivative of AI, like a chip manufacturer — “you literally have to argue with an agent to get it not to use Stripe.”
Rory on why it’s the golden asset: a diversified core business plus an AI growth lift, with downside protection — “if the AI stuff peters out, they’re still gonna kick off cash like crazy.” And the stunner: share count is down versus three or four years ago — buybacks while private, “doing everything a public company can do while private.” Jason: “I’m gonna email that quote to some portfolio companies.”
Jason’s implication for public SaaS: when OpenAI and Databricks go public at roughly 80% growth and Stripe compounds like this, “nothing except Palantir approaches these — even Cloudflare isn’t this good,” and everything below the line becomes “a distant memory of the past.”

12. Instinct’s incident underscores it: you still can’t trust agents

The news beat: GroqBot and Instinct drew attention over information and data-security exposure, including concerns about giving Instinct access to passwords, bank accounts, and other sensitive material. Jason’s point is that this class of problem hasn’t been solved in a year: not by OpenClaw, whose Mac mini drama he lived through, and not by the new entrants. In his Moltbook experiment, the system once “attempted to buy six AP watches for the team, for $360,000.” Goal-seeking probabilistic LLMs err like a junior employee — “but they could do it a thousand times more.” Solvable in theory; “as of today, you can’t trust these agents.”
Harry’s inevitability case — “you guys remember ‘we’ll never put our credit cards online’” — gets partial agreement, but Rory questions whether idiosyncratic personal workflows are the right place to spend versus enterprise loan processing with real budgets. His borrowed Ben Thompson line: “Silicon Valley forgets every three years that the average American is not trying to be efficient.” Superhuman, Calendly, Evernote — the category is “real, but it’s mid… always just a little bit out of reach.”

13. The dumbest categories of the AI era — with a “what if it works” escape hatch

Jason’s closing question: where are we throwing cash we’ll later regret? Harry names customer support — one or two players take most of the market, and every sophisticated large technology company is building its own. Jason goes further: classic CS/CX “won’t even exist in 24 months,” merging into marketing and sales as “everything’s becoming one agent.” Rory, with investments in the space, concedes only that “even if not, it won’t be the biggest mess.”
Rory’s negative pick is defense — not because the products aren’t needed, but because account control with the Pentagon demands a portfolio of products, so “two or three companies like Anduril… will hoover up the rest of us.” Consolidation, not losses.
Harry adds humanoid robotics — visionary TAM versus the reality of dexterity and touch. Rory, on the Locus Robotics board with 15,000 purpose-built robots in the field, agrees via the viral robot-running-faster-than-Usain-Bolt video: “if I want a machine to do 100 meters really quickly, I’ll get a fricking Tesla.”
Jason’s own: venture-funded rollups of accounting and law firms — “it makes too much sense on a spreadsheet,” and he’s still waiting for the $20B outcome. Rory’s honest coda: he holds versions of all these biases but keeps them “absolutely overcomeable” by facts, a great entrepreneur, cynical comments, and portfolio construction — because, as Jason says, “we’ve never seen the type of creativity from founders like we’ve seen today… now is the moment, man.”

Rethinking Legacy Data Infrastructure with Eon Co-Founders Ofir Ehrlich and Gonen Stein MARKET UPDATE: Bitcoin Rally, Zcash ETF, Bessent Humbled, NVDA Earnings