Michael Burry Shorts NVIDIA and Palantir & Has Defensibility Died in a World of AI?

Summary

• Jason actually priced Michael Burry’s $1.1B short on Nvidia and Palantir rather than opining on it: with NVDA at $188, December $180 puts cost ~$9 and need the stock at $160 in 47 days just to return 2x — and below $100 for 8x — while anything above $180 loses everything. Two-year LEAPS cost $50-55 per put and only break even below $150. His verdict: a capex correction is “very hard not to imagine,” but “going from that kind of armwavy podcast statement to actually being able to make money on it, that’s damn hard.”
• The “where’s the revenue” debate is over, per all three: Altman says OpenAI hits $20B ARR this year, Anthropic projects $70B by 2028, and companies are raising 2025 estimates in-year — “an enormously positive signal.” Rory: “cynics sound smart and optimists get rich”; the only real question left is whether next year’s capex is $40B or $80B, “both of them are still huge.”
• Jason’s core 2026 call: AI stops being a tool and becomes part of your team. Replit’s Agent V3.1 is the first to “jump the line” — essentially infinite context, remembers a month of shared work, ships to production in 15 minutes — while copilots “didn’t work. It was a ripoff.” Gamma ($100M revenue, 50 people, raised at $2.1B) is the same story in marketing: “when the AI is part of your team for real, not VC talk, the amount of revenue that is accessible is so high.”
• Seed-stage defensibility generally emerges at scale: one of Jason’s investments drew five clones in the first 30 days, and what took incumbents three years now takes 90 days. Harry’s view: “the defensibility theorem emerges at scale… You can’t be anointed the winner up front. Get over it everybody.” Lovable went from 200 post at $4-5M revenue to $2B at $80-100M — “valuation has expanded to fill the risk that was reduced operationally.”
• The Sequoia transition — Roelof Botha out as steward after three years, replaced by Pat Grady and Alfred Lin — reads as the best firm in the world feeling behind on AI. Jason widens it: “the most folks from the last decade or 15 years are not the right people for the next decade… we don’t even need half the VCs we have today.” “Take your Nvidia shares and buy a beach house.”
• It’s “the most binary fundraising environment in our lifetimes”: Harry watched a classic SaaS company that went 400K to $3M ARR take 120 meetings to get one term sheet ($10M on 40 post). You’re hot AI-native with top-quartile growth or you’re not getting funded — and the best processes “don’t feel like a process” and “don’t require a data room.”
• Public markets are sorting on the same logic: Datadog +23% by co-attaching to AI compute spend (“sell [bleeped] to the people who are making AI and if they grow, you’ll sell more [bleeped] too”), Duolingo -25% with “the wrong kind of AI.” Harry’s taxonomy: attach to compute, replace humans, or steal incumbent revenue — “you don’t get any kudos for sprinkling AI dust on your product.”

Deep dive

Note: despite the feed title, this episode is the 20VC roundtable — Harry Stebbings with Jason and Rory — not the Carl Rivera interview.

1. Sequoia’s shake-up: even the best firm feels behind

• Rory’s outside-in read on Roelof Botha stepping aside as steward after a three-year tenure, replaced by Pat Grady and Alfred Lin: “whenever you have a CEO change… it’s cuz something is wrong” — an internal perception they could do better, missed rounds (the articles cite Cursor), passed deals. And oddly reassuring: “I get up and I go, God, we need to do better… and then you realize the best firm in the world is having exactly the same feelings.”
• Jason widens it into a generational call: “the most folks from the last decade or 15 years are not the right people for the next decade… I struggle to even recommend a lot of the CROs and executives I know for roles today… we don’t even need half the VCs we have today for the AI world.” His advice to the incumbents: “take your Nvidia shares and buy a beach house.”
• Rory’s structural frame — worth keeping: venture’s winners will be Walmarts or Chanels — mega-platforms (Thrive, Lightspeed, General Catalyst) or boutiques (Benchmark, USV) — and Sequoia sits in the middle as “a manager of managers of venture firms,” a “precarious place” once you no longer “eat what you kill.” Rory still credits the ruthlessness: not “it’s someone’s turn,” but “it’s not working, let’s make a change. It’s very Sequoia.” Jason on partnerships generally: carry can’t track performance, so “the daggers are always out.”

2. Pricing the Burry short: betting against AI capex is a terrible business

• Jason did something unusual — he ran the actual numbers on replicating Burry’s $1.1B Nvidia/Palantir short. Nvidia at $188 on Monday; December $180 puts at ~$9 each give you 47 days: the stock must hit $160 just to make 2x pre-tax, $100 — nearly a halving — for 8x, and above $180 “you lose it all.” Would he put $1M on that? “H, no.”
• The longer-dated version is no kinder: two-year LEAPS at the $180 strike cost $50-55 per put, losing money unless Nvidia falls below $150 and needing sub-$100 for the same 2x. “It just brings home how hard a business it is to bet against AI capex. That was my big takeaway.” Jason’s allocator take: against other asset classes, the risk-adjusted 2x “is not a good decision to make.”
• The tell Rory flagged: Burry filed his disclosure early — every prior quarter he filed on the last legal day. With only 47 days to be right, “you have every incentive to pile out the bad news and [bleeped]-talk the stock” — a polite way of announcing the bet to move the market.
• Rory’s honest two-sidedness: Palantir “north of 110-120 times revenues,” growing 50-60% last quarter — “very hard not to imagine the next two years that doesn’t have a significant correction.” But options are zero-sum: “for every idiot like Rory, there’s a smart guy the other side of the table who prices it better.” The shorts keep the system honest — “that’s an expensive way to be a policeman.”

3. The revenue already showed up — the only question is 40 or 80

• Harry’s challenge to the room: the whole “will the revenue show up” anxiety is answered — Altman says OpenAI hits $20B ARR this year, Anthropic projects $70B by 2028. “Are we being overly British? Are we looking for a problem that’s not there?” Both guests concede immediately.
• Rory’s refinement on which signal matters: raising a 2027 estimate is noise, but upticking your 2025 estimate mid-year “is an enormously positive signal” — and Anthropic in particular has been doing it. Even CoreWeave’s stumble was “we couldn’t get the data center up and running,” not lack of demand. “Right now it’s very hard to make an intellectual case for anything other than there is massive demand for compute.”
• The framing that survives: “cynics sound smart and optimists get rich.” This is “maybe the biggest mega trend since the early days of the internet — maybe it’s even bigger… playing against it is dumb as rocks.” The second-order question is all that’s left: “are we going to sign 80 billion of capex for next year or 40 billion? Both of them are still huge… but one of them is 40 billion more than the other.”

4. Gamma: stealth TAM expansion, $2M revenue per head

• The proof point of the week: Gamma raising $100M at a $2.1B valuation, having hit $100M revenue with 50 people — $2M per employee. Jason’s usage case, as told: instead of a dated prospectus, Gamma pulls SaaStr’s Salesforce and marketing-automation data, knows a sponsor’s prior leads and ROI, and builds “a fully dynamic piece of collateral in about 10 minutes” — for $100 a month.
• The economics insight buried in that anecdote: SaaStr paid zero for Google Slides and doesn’t know where the Microsoft Office key is — yet now pays $1,200 a year. “It is a stealth TAM expansion.” At that trajectory, “20x revenue doesn’t sound expensive… 1 to 100 in 11 months at 20x revenue sounds cheap and profitable.”
• The caveat, also Jason’s: Canva now has a Gamma clone that’s “not bad” — and when Canva’s Cliff was last on the show, they weren’t even competitors. “Gamma’s got to keep swimming… It used to take Cliff 3 years. Now it takes Cliff 90 days.”

5. 2026: AI stops being a tool and joins the team

• Jason’s report from 126 days of vibe coding — 10 apps launched without an engineer: Replit’s Agent V3.1 is the first of the ~20 agents SaaStr uses that “jumped the line” from tool to teammate. It has “essentially infinite context… remembers everything I’ve done for the last month,” and turned a new idea into production in 15 minutes. “The lame thing about co-pilots is they were just tools… When the AI is part of your team for real, not VC talk, the amount of revenue that is accessible is so high.”
• His definition of the threshold, when Rory pressed for the distinction: “sufficiently autonomous, knowledgeable and powerful to complete material high-value tasks on its own with some daily discussions — just like our team.” The test case: “Gamma, go into my Google calendar, create prospectuses for all 20 sales calls this week, pull the data from Salesforce and HubSpot and Marketo, distribute to the team — and if you can, join the meeting. It’s not that far away.”
• The periodization: “the co-pilot was the 2024 story. It didn’t work. It was a ripoff” — this year “OpenAI and Claude finally actually got good,” which is why Lovable, Replit, Gamma and Vercel exploded from essentially zero revenue. “Next year is AI is part of your team… That’s where we should be investing.” Rory’s echo of why this is stressful: “if you decide what I knew 6 months ago is still useful, you’re probably going to be wrong very quickly.”

6. Five clones in 30 days: what the hell does seed investing mean?

• The clone-speed problem, in Jason’s telling: it used to take Salesforce-class incumbents three years to mount real competition; now “I can think of one investment I’ve made that has had five clones in the first 30 days, including one from a cloud leader.” Hence the existential question he asked twice: “What the hell does seed investing mean?”
• His fallback answer is the old one, held with less confidence: “you’ve got to still bet on the best founders. You just can’t take that early first-month explosion as seriously as you used to. It’s not as defensible.”
• Harry’s synthesis — the episode’s sharpest framework: “The defensibility theorem emerges at scale. Once you become the anointed winner, once a market coalesces… it’s yours to lose. The idea that at the seed stage you’re going to find a defensible way to do codegen is absurd… You can’t be anointed the winner up front. Get over it everybody. It’s a high-risk game.” His version of hope: “a plane of stability that is later than it used to be that is still fragile” — Replit’s competitors mostly can’t build the agent (Bolt, the early leader, is “now just number three” and outsourced its agent to Claude), so at $100-250M “these Gammas start to build a moat.”
• Harry’s counter: vertical specialization compounds — his AI-patent-law deal Solve Intelligence gets better with every patent ingested. Jason’s pushback, kept as said: “those patents are public, right?” — and he thinks horizontal pulls away too: “would you fund another company now from scratch to do exactly what Cursor does? I think not… but I could be wrong. I’m processing in real time here.”

7. Do you know the winner at the B? And are we overpaying?

• Jason’s money question after the defensibility abstraction: “is that okay at 50 post for a seed round? If the deals are at three post and I can spread 500k checks, I get it. If I’ve got to spread $5 million checks at 50 post, it’s tougher.” Rory reframes: “are you getting paid for the risk?” — the same wrestle he has at $100-200M pre.
• Rory’s claim: between the A and the B, odds narrow from one-in-ten to one-in-three — in codegen, Cursor, Windsurf and Cognition had emerged by the B. Harry’s pushback, worth keeping whole: “I don’t think we know at all. Codex is making incredible ground, Claude Code’s making incredible ground” — and on Windsurf in retrospect, Jason says “it wasn’t clear it had a sustainable moat of any sort.” Harry’s darker math: if one-in-three has become one-in-seven-to-ten, “surely we are dramatically overpaying.”
• Harry’s tell for graduation: “when you go to the first board meeting where you suddenly realize you’re scared of the big-company adjacent competition, it probably means you’ve graduated from the baby class” — fearing Salesforce and Canva is proof you won startup-land.
• The Lovable case study crystallizes it: done at 200 post on $4-5M revenue, then $2B pre at $80-100M. Rory: consensus says massive operational risk was reduced — “and now, as always happens in the bull market, valuation has expanded to fill the risk that was reduced.” What’s left is the late-stage question: is the TAM big enough to support $2B?

8. Fund construction: diversify more — and Hummingbird’s god-tier answer

• The math Jason ran: if risk is up and seed checks are $5M, you may need 40 first checks — 200M deployed, 200M reserves, fees on top: “I need at least $500 million for my little seed fund to make the math work.” Rory catches Harry’s implicit answer — take more shots at lower ownership — as “embracing a more diversification story in the face of risk, which is math 101.” Harry’s evidence: an LP position in a fund with 100-150 positions at $100-150k checks that’s a 7x fund — “1 million on 50 does work if it’s a hundred-billion-dollar company, not the three-to-five-billion outcomes we’ve been playing with.”
• Jason’s rejoinder: “works even better if you own 10% of that company” — and his confession: “I’m obsessed with ownership now… I don’t see how I’ll ever own 18% of something at IPO ever again.”
• Which makes Hummingbird the episode’s closing hero: their first bio deal, BillionToOne, IPOs with Hummingbird holding an $800M position out of roughly a $150M fund. Rory’s frame: you can keep your multiple or keep your ownership — Hummingbird kept the multiple (8-10x funds at $40-100M), a Lightspeed keeps ownership for a 5x on far more capital. “For the marginal dollar,” the small fund that accepts follow-on dilution “is the compelling product.” Jason calls maintaining ownership on a nine-figure AUM “the elite game today… god tier.”
• The capital-efficiency coda: Navan public at $4.5B and BillionToOne at $5B — “the power of capital efficiency and bluntly running lean… Yes, we’re Captain Obvious today.”

9. The non-process process, in the most binary market ever

• Harry opened up about a founder who rebuffed his at-price term sheet to “run a process on the 19th of November” — he found it abrasive. Rory sides with the founder on principle (“I see more failed financings because they didn’t run a process than because they did”) but calls the execution “a slight own goal”: the best founders nurture three or four primed investors so that when ready, “they just send an email… you’re just in.” His controversial corollary: “the best-run processes don’t require a data room” — only a diligence file. Rory’s distillation: “the best-run processes don’t feel like a process, but they are.”
• Jason’s warning on the accidental version: sharing data serially with one interested investor who isn’t ready to commit means “you’ve had a failed process without ever doing a process. Share with everyone or share with no one.”
• The backdrop that makes all this moot for most: “the most binary fundraising environment” — you’re YC/Neo/South Park Commons or hot AI-native with top-quartile growth, “or who the hell is going to find you in your pre-seed round… there’s not a lot of gray zone anymore.” Harry’s tangible example: 400K to $3M ARR, classic enterprise SaaS that once drew five term sheets — 120 meetings, one term sheet, $10M on 40 post. The old triple-triple-double-double? “Every month that goes by, those deals are harder to do.”
• A revealing sidebar on styles: Harry’s partnership does 20 in-person meetings per partner per week — 80 net new companies, 3,500-plus a year. Jason: “I would resign… I sold my companies because I didn’t want to spend my life in meetings. Don’t try to get the meeting with me — send me a great deck and a great email.” Jason, the self-described “meeting junkie”: “the person inside living it every day has a crucial kernel of knowledge that you just can’t access any other way.”

10. Datadog up 23%, Duolingo down 25%: three ways to earn AI revenue

• Datadog crushed — stock up 23% on what Harry read as 15 million-plus AI-native customers — and Harry reprises Jason’s earlier frame as the explanation: “sell [bleeped] to the people who are making AI, and if they grow, you’ll sell more [bleeped] too.” Jason adds the 2026 wrinkle: the hyperscalers are “buying like classic B2B companies now” — OpenAI purchases like an Adobe or Microsoft, recycling the same procurement people — so anything attached to AI budget has a great year coming. Same move by legal tech’s Clio (likely), a 2008-founded company that attached its way from $3B to a $5B valuation.
• Duolingo’s 25% plunge splits them. Rory: “I don’t think there’s a mega story here” — still up ~80% from its IPO, guided slightly down, “the stock was overvalued and went down. Life goes on.” Jason: it has “the wrong kind of AI” — “you don’t get any kudos for sprinkling AI dust on your product. This is not 2023.” Duolingo already disrupted Berlitz and the language schools: “where’s the next level of human disruption?”
• Harry’s resulting taxonomy for every AI story: attach to the compute budget, replace humans, or use AI to steal an incumbent’s revenue — and the third is weakest (“there’s like 400 AI CRM startups saying they’re going to eat Salesforce’s lunch — that’s not exciting to me”). Jason’s education pushback: LLM tutoring roughly equals one-on-one human tutoring, a genuine unlock — but “where does the budget come from? The public school district’s not going to come up with another 10 million,” leaving the niche adult-learner spend that once went to human coaches.
• The closing register, Jason’s: “this is the first time software has gotten better since the three of us met. It’s the same crap [until now]. If you’re not truly excited, retire — no shame in that — put the rest into NASDAQ and you’re going to make more than most VC funds anyway.”