Summary

Nvidia’s $100 billion commitment gives OpenAI permission to test the scaling thesis until returns—not financing—force a stop. Rory O’Driscoll called it “not an infinite money machine, ’cause it will end,” but said the market has effectively told Sam Altman, “Have a go.” Jason Lemkin emphasized Altman’s claim that OpenAI needs “three orders of magnitude more compute,” while Rory remained unsure the marginal $300 billion can earn an adequate return.
The AI trade rests on extreme concentration and a historic gap between spending and realized revenue. Roughly six customers reportedly generate 83% of Nvidia’s revenue, yet those buyers—OpenAI, Google, Meta, Oracle and peers—are determined to “spend themselves into oblivion.” Harry Stebbings distinguished the genuine application boom from an AI CapEx boom running near $600 billion annually against only $30 billion–$40 billion of current revenue.
The moment rhymes with 1999, but today’s giants possess enough cash and mutual guarantees to prolong the cycle. Jason contrasted Web 1.0 companies that simply ran out of money with an ecosystem where suppliers, customers and infrastructure providers backstop one another; Nvidia itself went from $3.8 billion of free cash flow in fiscal 2023 to $60 billion in fiscal 2025. The warning sign is capital allocation: Nvidia repurchased $9 billion of shares last quarter and authorized $60 billion while the stock and broader market look “very frothy.”
Venture’s apparent concentration is largely a new private-public market layered above a relatively stable early-stage business. Seventy-five percent of 2025 VC dollars reportedly went to 19 companies, but Rory argued that the remaining 25% still resembles historical venture, while an extra roughly $50 billion a year, or whatever the exact figure is, funds ultra-late-stage winners. “Triple, triple, double, double” is not dead at meaningful scale; it simply no longer guarantees an effortless raise, particularly outside the most fashionable categories.
A celebrated company and a celebrated return are not the same thing: OpenAI and Netskope could each deliver roughly a 7X to certain investors. Rory estimated that OpenAI’s earliest 2019 money may be up 7X–8X on a blended basis after follow-ons, despite Harry citing a claim that 10% of the world’s adult population uses it weekly; an early Netskope position could produce a comparable multiple. Whether to sell depends first on fair value, then on concentration, taxes and marginal utility—not on the emotional certainty that “there’s another card to play.”
Navan’s IPO filing is both a recovery story and a bid to reach public markets before investors choose adjacent comparables. The company disclosed $613 million of revenue, 32% growth, 10,000 customers and 110% NDR after travel revenue probably went to zero in March 2020. Although Navan is chiefly a travel-booking business rather than a Brex- or Ramp-style card company, the panel judged going first strategically smart—even before profitability—because “you definitely don’t wanna be last.”
Notion’s $500 million ARR and reacceleration show that scaled SaaS can earn an IPO without becoming an entirely different company. Growth around 30%–40% could support perhaps a $4 billion–$5 billion valuation, potentially more, but nowhere near every 2021 watermark. The broader reset was blunt: “flush” old decacorn valuations, price Airtable and peers on growth and free cash flow, and recognize that investors appear determined to repeat the same diligence mistakes in 2025.
The $100,000 H-1B fee may have a modest immediate startup effect, but it is directionally harmful to the talent engine—and emblematic of crude policymaking. Jason’s first startup employed two H-1B transfers among its first 10 people and, he said, could not have achieved its exit or life-saving work without them; big companies will pay while founders seek O-1 alternatives. In the same overheated market, “founder-friendly” has become table stakes and theater: the real test is who writes the check, recruits the executive and stays through the bad board meeting.

Deep dive

1. Nvidia’s $100 billion lets OpenAI run the scaling experiment to its limit

Rory rejected Harry’s “infinite money printing machine” framing because every aggressive financing structure eventually meets the underlying business. If OpenAI’s projections reach $100 billion-plus of revenue, everyone books gains and looks brilliant; if not, “it all comes back and bites you in the ass.”
The immediate consequence is more important than the circularity: nobody is likely to call timeout for another year or two. Capital and chip access have been supplied, participants can mentally mark up their positions, and OpenAI gets to discover whether continued scaling produces adequate economic returns.
Harry’s pushback was that GPT-5 emphasized efficiency and delivered less improvement than expected, suggesting scaling laws had already weakened. Jason answered with Altman’s own words: this commitment is “just a start,” OpenAI needs “three orders of magnitude more compute,” and believes the first $100 billion might help cure cancer and educate every student.
Rory’s hedge remained intact: he was not saying Altman is right, and he doubts the marginal $300 billion necessarily earns a return. His narrower call was that six years of OpenAI being “astonishingly right” made continued backing humanly inevitable: doubling down continues until one incremental double-down fails.

2. OpenAI gains momentum, but Anthropic is not obviously capital-constrained

Asked whether Dario Amodei should feel structurally disadvantaged, Rory separated capital from utility. Anthropic reportedly had investors “beating people off with a stick” in its recent round; if it wanted another $10 billion, he believed that money would arrive the next morning.
The possible disadvantage is preferential GPU access and the ability to contemplate custom chips or hundreds of gigawatts of compute. Jason saw OpenAI’s scale as qualitatively different, while Rory questioned what concrete constraint—beyond “momentum and bigness”—the Nvidia commitment actually removes.
Jason argued that OpenAI is also carefully managing monopoly optics. ChatGPT approaches Chrome-like consumer dominance—“the Standard Oil of tech”—yet Altman avoids denigrating competitors because OpenAI benefits from a viable number two and has already endured conflict with Microsoft and its own governance crisis.
Rory’s counterweight was consumer economics: ChatGPT is a remarkably poor extortionist monopoly because it currently subsidizes enormous consumer surplus. Nvidia looks closer to a conventional monopoly, making one undisclosed detail especially consequential: how OpenAI’s effort to build competing chips is treated when Nvidia is simultaneously becoming an equity holder.

3. Six buyers support a $4.5 trillion supplier and a $600 billion CapEx boom

Nvidia’s concentration is startling: Rory cited roughly six customers producing 83% of quarterly revenue for a company worth around $4 trillion–$4.5 trillion. Apple has approximately two billion customers and Microsoft hundreds of thousands of meaningful enterprises; Nvidia’s valuation is “single-threaded” to the decisions of six or seven people.
The offset is that none of those buyers appears ready to blink. OpenAI is escalating, Google has repeatedly signaled it will keep spending, Meta is willing to “tear up the book,” and Oracle is committed—leaving all six determined to “spend themselves into oblivion to win the prize.”
Harry extended the concentration downstream: two customers represented 55% of one data-labeling provider’s revenue and recur across competing vendors. These buyers are promiscuous because only a handful need labeling at scale, and speed matters more to them than squeezing suppliers on price.
Harry’s essential distinction was between the AI revolution and the AI CapEx boom. Applications are genuinely gaining adoption, but the extraordinary trade is roughly $600 billion of annual CapEx racing ahead of a market producing perhaps $30 billion–$40 billion of revenue; Mercor, Surge and anything attached to that spending have been “stuffing” dollars into their buckets.

4. This rhymes with 1999, except the ecosystem can finance itself for longer

Jason found the present closer to 1999 than anything in the past 20 years: unlimited possibility, vendor financing and the memory of Nortel and Lucent funding bandwidth customers to sell equipment. Nvidia’s use of equity rather than debt changes the instrument, not the resemblance—and 2000 showed how abruptly limitless belief can collapse.
Jason’s key difference was survival capacity. Amazon nearly ran out of money after its IPO, whereas today’s leaders guarantee one another’s commitments; his exaggerated but memorable formulation was that Nvidia has agreed to buy “300 years” of CoreWeave capacity, making an immediate cash exhaustion far less likely.
Nvidia’s free cash flow illustrates the new scale: Jason cited $3.8 billion in fiscal 2023, $27 billion in 2024, $60 billion in 2025 and perhaps $100 billion or something in the following fiscal year. That internally generated capital can perpetuate the cycle even if historical analogies fail.
Rory nevertheless questioned Nvidia’s $9 billion quarterly repurchase and $60 billion authorization, equal to its prior-year free cash flow. Jason suggested buybacks may offset RSU dilution; Rory called mechanically linking repurchases to dilution “as dumb as rocks”—buy when shares are cheap, preserve cash when they are dear.

5. Expensive markets punish long-term returns before they punish momentum

Harry’s 28-stock book had 27 positions green, prompting the honest conclusion that he was not suddenly that good. Jason’s answer was darker comedy: he was 0% cash, just as in 2008 when a collapsing market forced him to sell stock down 60%–70% merely to repair his roof.
Rory distinguished horizons. Valuation has little power to predict one-year returns, particularly while the Fed is cutting and momentum persists, but it correlates meaningfully with 10-year outcomes; buying at today’s levels implies a substantially lower-than-average long-term public-market return.
His practical response is asset allocation, not top-calling: holding cash imposes underperformance during a bull market, but that is “the cost of sleeping at night.” Investors should choose a medium-term mix consistent with their risk tolerance rather than optimize for the final stretch of a rally.
Jason’s froth indicator was LPs publicly bragging about returns on LinkedIn—behavior he associated with 2021. He joked that all their 401(k)s were in Nvidia, while corporate history suggests companies will again buy back aggressively near peaks and wish they had liquidity when income-statement optimism gives way to balance-sheet scrutiny.

6. Venture concentration is a second market, not the disappearance of the first

Seventy-five percent of 2025 venture dollars reportedly went to 19 companies, but Rory reframed the statistic. The remaining 25% resembles the long-standing seed-to-Series-B market, fluctuating perhaps 10%–20%; layered above it is an additional roughly $50 billion-a-year private-public business operating one or two valuation orders higher.
Concentration naturally rises by stage because companies fall away at each round. If businesses remain private longer, the endpoint may be “two foundation models and Databricks and Stripe raising a Series N or G”—still reported as VC, though economically distinct from conventional early-stage investing.
Jason said the true S-tier remains easy to identify. The murky zone is immediately below it, where one investor preempts at an outlier price while another sees churn, margins or weak differentiation; predicting the next financing has become much harder even when portfolio companies are growing well.
The panel rejected the categorical death of triple, triple, double, double. At roughly $10 million–$20 million of revenue it now requires far more work and meetings, and an unloved category can hurt; at $50 million–$100 million growing triple digits, investors will still show up because so few companies reach that combination of scale and velocity.

7. Fund economics make ordinary-looking 7X outcomes matter

Harry used ICONIQ’s exits—Netskope around $8.5 billion and Atlassian’s roughly $1 billion purchase of DX—to ask whether such wins still “count” beside Anthropic. Rory’s answer was categorical: if a 10% Netskope stake becomes roughly $700 million of liquid public equity, “it goes in your bank account,” which remains the mission.
OpenAI makes the distinction between corporate importance and investment multiple vivid. Rory estimated the earliest 2019 investors might be up roughly 7X–8X after follow-on rounds; first-round money could be 25X–30X while later capital earns about 3X, blending to 7X on $100 million–$150 million invested.
Harry objected that OpenAI has 10% of the world’s adults as weekly active users, yet Rory held the line: stature does not change fungible money. A Series A investor in Netskope could also earn 7X–8X, while a $5 billion-plus IPO can still produce an excellent 10X if ownership and entry price are right.
Fund size changes which returns are sufficient. A $10 billion vehicle must concentrate huge sums in perhaps five to seven deals; smaller funds have more ways to compound meaningful equity. Harry’s concern was social rather than mathematical: respectable exits increasingly become “an asterisk at the bottom of the DPI table.”

8. Selling OpenAI is a valuation decision overlaid with human utility

At the quoted $500 million valuation, Rory said an investor must at least consider selling. The first step is to form a fundamental view of fair value and upside—“it’s gonna double ’cause it’s always doubled” is not analysis—then overlay personal wealth, fund construction, taxes and liquidity needs.
Jason explained why almost nobody will volunteer to exit: ICONIQ reportedly received unprecedented calls from its own LPs seeking Anthropic allocation, and a move from $500 billion to $1 trillion requires no new sourcing or board work. Taking $50 million today when waiting six months might produce $100 million feels emotionally impossible.
Harry’s counterexample was himself: without $50 million already, a first large realization changes life and should be treated differently. Rory agreed that marginal utility matters; a liquid first $5 million should usually be secured rather than exposed to a concentrated continuation bet with merely equal expected value.
Drawing from The Missing Billionaires, Rory said investors commonly have risk aversion around two, while figures such as Elon Musk—and, in another context, SBF—behave closer to pure expected-value maximizers. The Vanderbilt descendants illustrate the failure mode: stock selection matters less than bet sizing, diversification and preserving accumulated wealth.

9. Early wins create both deal flow and the courage to keep rolling

Asked whether wealthy venture firms take more upside risk, Rory said “at the margin it’s almost certainly true.” An emerging manager may need half a fund of DPI; Sequoia can believe another opportunity arrives tomorrow and refuse a merely adequate offer without threatening the institution.
The mechanism extends beyond referrals. Venture success correlates weakly with many visible traits, but strongly with an early win: successful investors see better opportunities and “just have the stomach to roll the dice, and you get braver.” That can become overconfidence, yet risk-taking remains necessary for outlier returns.
Oren Zeev’s concentrated Navan position supplied the live example. Harry described approximately 20% exposure across multiple funds; Jason cautioned that SPVs and opportunity vehicles might make the core-fund concentration lower than it appears, while Rory acknowledged he personally would find a true 20% single-company bet difficult.
Brian Singerman’s maxim that concentration limits are “the enemy of great venture returns” met Rory’s quantification: ask how certain one company is to outperform the rest of the portfolio, not whether the investor feels brave. More concentration can increase outperformance, but only by accepting significantly more risk.

10. Navan is racing public before adjacent companies define the category

Navan filed with $613 million of revenue, 32% year-over-year growth, 10,000 customers and 110% NDR. Rory first celebrated the survival story: a concentrated investor watched a travel company’s revenue probably go to zero in March 2020 and now has a credible IPO within reach.
Jason’s strategic read was that Navan is going before Brex and Ramp. Brex had announced roughly $700 million of revenue growing 50%, while Ramp appeared larger and at least comparably fast; if public investors treat all three as peers, the perceived number three benefits from establishing itself before the leaders list.
Rory’s pushback was product-level: Navan predominantly earns from travel booking, Brex and Ramp from cards and payments, and Bill from accounts payable. Yet Navan’s S-1 and rebrand from TripActions claim a broader horizontal suite, so management cannot reject those comparisons while simultaneously seeking their valuation halo.
Profitability argues for waiting, but timing argues against it. Navan held OpEx flat or slightly down despite 4% inflation while growing near 30%, suggesting a determined push that might still require one or two years; this market offers liquidity now, and “you definitely don’t wanna be last” after two better-known adjacent companies are tradable.

11. IPO liquidity arrives through a long sequence, not on listing day

Rory corrected headline wealth calculations: an SEC filing may attribute the GP’s, partners’ and LPs’ shares to one named investor, even though that person owns only a fraction of the displayed billions. A conventional IPO usually imposes a six-month lock-up; direct listings can avoid it, and performance triggers sometimes release shares earlier.
If shares price at $14, trade to $18 or $19 and the company delivers its first quarter, insiders may complete a registered secondary during the lock-up. If the stock falls to $12, $10 or $9, that route becomes extremely difficult—one reason a 10%–15% IPO pop is not necessarily wasteful.
After lock-up, funds can sell or distribute shares to LPs, but board reporting obligations and quiet periods make disposal slow. Rory’s typical timeline was 12–18 months to leave the board and roughly 18–24 months to exit a substantial position, rather than immediate liquidity at the opening print.
Jason noted that holding while possessing inside information is lawful; Rory added that negative information can be sold, so the rule is not one-way. Rory recalled remaining silent during M&A talks that eventually produced a 30%–40% premium; the trade-off is restricted windows versus privileged strategic visibility. Nvidia’s early venture directors stayed from the 1997 IPO, and Mark Stevens reportedly may never have sold.

12. Skilled-immigration damage may be modest immediately but negative structurally

The discussion cited a new $100,000 payment for H-1B visas, 440,000 applications, roughly 70,000–75,000 acceptances and an estimated $19 billion–$120 billion GDP contribution. Rory’s firm conclusion was directionally negative: skilled immigration has been extremely good for US technology, though the word “material” is harder to establish.
Jason made it personal: two of his first startup’s initial 10 employees were H-1B transfers, essential to its material-science work, first exit and, he said, saving hundreds of lives. “Every single talented person” strengthening the US and its companies is desirable to him ethically, personally and economically.
His tactical forecast was modest disruption if the policy does not expand. Large technology companies will pay, founders increasingly navigate O-1 visas despite their stress and drawbacks, and startups “find ways”; that practical adaptation does not make the policy good.
Rory saw $100,000 as a crude proxy for a skills-based system and suggested STEM credentials or a points framework would target national benefit more rationally. Public anger focuses on alleged abuse—such as firms employing H-1B workers while laying off Americans—while obscuring the founders and technical specialists the ecosystem could not otherwise recruit.

13. Scaled SaaS can recover, but 2021 prices no longer anchor value

Notion reaching $500 million ARR while accelerating impressed the panel. Harry highlighted that “double, double, double, double at hundreds of millions” is exceptional. Rory agreed that a mature SaaS company need not reinvent itself: it can lean into AI, preserve its core product and regain 30%–40% growth within striking distance of an IPO.
With roughly 1,200 employees and assumed 30%–40% growth, Rory estimated perhaps seven to nine times next-twelve-month revenue—around $4 billion–$5 billion, potentially higher if forward numbers surprise. Jason’s stock came from acquisitions around a $10 billion mark; Rory’s dry answer was, “That’s your problem, not theirs.”
High-priced preferred shares should not hold the rest of a company hostage while it waits to regain $20 billion privately. The preference can convert or remain outstanding until the public value grows into it; a real $5 billion–$8 billion outcome still matters even when it disappoints a 2021 buyer.
Airtable and similar productivity companies will now trade on fundamentals: Rory suggested five to six times revenue for $200 million growing 20%, or seven to eight times for $300 million–$400 million growing 30%–40%. Microsoft’s grind and the shift toward AI ended the unbounded narrative; “God forbid,” free cash flow matters again.

14. The market has reset old marks while recreating old diligence failures

Jason’s deadline was January 1, 2026: after four years, investors should stop complaining about 2021 decacorns, mark them down and “flush them down the toilet.” Klarna’s former $45 billion peak was history; capital and attention should move to present fundamentals.
Rory’s sting was that investors appear determined to make “exactly the same mistakes in 2025.” He is seeing less diligence than in 2021, while Jason described hot AI deals decided on Saturday with neither supplied materials nor substantive review: “Why would you do diligence? All you can lose is one extra money.”
Jason had watched a tier-one firm issue a term sheet to a sizable company, use the 30-day closing period for deeper work, then withdraw. He called this a worsening competitive tactic: secure exclusivity first, investigate later, and leave the founder carrying the broken deal.
Rory now has more sympathy for investors asked to decide in one hour while founders share no data and present paid pilots as contracts. If immediate review reveals “10 things that aren’t true,” investors must retain the right to rescind; Harry’s cleaner policy is simply declining processes whose timetable prevents conviction.

15. Founder-friendly is theater until the company enters distress

Jason’s verdict was that “founder-friendly has become bullshit,” even while remaining table stakes for winning allocations. Praise, performative enthusiasm and routing a candidate request to a talent team prove nothing in a competitive bull market.
His operational definition is harder: write the check when nobody else will, remain at the board meeting after others disappear, help produce the eventual win and personally recruit the needed executive. Those actions—not “great job” regardless of performance—earn the label.
Rory preferred “founder honest”: tell the CEO exactly what you believe, because false reassurance is less useful than uncomfortable truth. Whether an investor is genuinely supportive can only be learned in a difficult deal, just as the quality of another VC becomes visible only after sharing adversity.
Harry supplied the specimen Jason would not: RevenueCat’s founders remembered him wiring personal money during the SVB weekend. Rory recalled partners holding a Sunday call to divide responsibility for companies until the US government intervened—“you only know what people are like in a tough deal.”

16. The quick-fire favored pragmatism over claims of technological leadership

On TikTok, Rory jokingly chose “never” because dangling a deal creates endless political leverage, while conceding it would probably happen eventually. Jason predicted the next 60 days, viewing it alongside tariff negotiations with China and India and hoping the harsher H-1B effects would similarly dissipate.
Jason assigned Meta’s smart glasses a “0% chance” of success despite owning roughly eight earlier pairs. His product thesis was simple: consumers do not need “a seventh screen” or to “play Tron in our eyes”; device paradigms are difficult to change, and many venture-backed wearables ultimately remain on shelves.
Atlassian’s acquisitions should help existing customers enter an AI-enabled engineering-management world, Rory said, but will not make it the dominant coding-agent company. Defending roughly $4 billion of revenue, a $40 billion market cap and 20% growth would be enough; Jason called DX and the broader buying spree “baby steps,” not a transformation.

No Priors Ep. 133 | With Alpha School Principal Joe Liemandt Eric Schmidt on AI, the Battle with China, and the Future of America