Sequoia Partner, David Cahn on Who Wins in AI, Defence & The New $0–$100M Playbook
Summary
Cahn’s “steel, servers and power” thesis landed: AI’s operative unit moved from dollars toward gigawatts, making power his “best trade of 2025” and construction execution a moat. Generators were sold out until 2030, electricians were being flown to Texas, and AI construction became a material contributor to US GDP. Yet the end demand remains unresolved: his original $600 billion revenue question has become roughly $840 billion.
AI can be a civilization-scale technology and still be a bubble whose compressed timeline “will incinerate capital.” Cahn expects AI to transform society over 50 years, but markets are financing that outcome as if it must arrive quickly on today’s specific chips. The investor’s task is therefore survivorship: find companies with customer love that can endure volatility, not businesses dependent on infinite cheap capital.
The cleanest bubble winners are consumers of compute, because excess capacity lowers their COGS and raises gross margins; producers inherit commodity economics. Harry challenged that framing with AWS, Azure and Google Cloud, but Cahn argued those monopolies were built before their opportunity was obvious. AI’s value is visible to everyone, inviting competition and making monopoly profits less likely—good for consumers, difficult for capacity owners.
The system’s clearest fragility is the transfer of risk from Microsoft and Amazon to smaller operators and then back to chip suppliers through circular financing. Oracle and CoreWeave cannot absorb hyperscaler-scale risk, while chip companies can fund projects cheaply because the spending returns as booked revenue. With one gigawatt costing about $40 billion—or $50-$60 billion on Vera Rubin—the announced 100-250 gigawatt ambitions become unfunded multi-trillion-dollar questions.
The most important overestimate is timing, not AI’s eventual importance. Andrej Karpathy’s “decade of agents,” Richard Sutton’s doubts about the current paradigm, Ilya Sutskever’s “pre-training is dead,” and Sam Altman’s “gentle singularity” all cut against status-driven lab chatter about AGI in 100-300 days. If the breakthrough arrives on Feynman-era chips in 2028 or a decade later, today’s H100 and B100 warehouses still bear the loss.
Neither premium venture brands nor abundant capital can manufacture product-market fit: “capital is fuel, but capital does not create the engine.” Cahn accepts that Sequoia can improve recruiting and marginally change probabilities, but “you can’t make a company succeed”; Profound was already ripping before Sequoia invested. The same discipline applies to metrics: margins can improve from 30% to 70%, while today’s strongest adoption signal is the “zero to 100 club”—not a rigid threshold, but evidence of unusually strong demand.
Defense may be “the next AI,” but it will produce a few national champions rather than a broad SaaS-like ecosystem. Harry challenged Sequoia’s absence from Helsing and Anduril; Cahn conceded Sequoia was late to defense, said the sector is roughly two years after the Transformer paper and before its ChatGPT moment, and estimated it is only about “1%” through a 50-year catch-up. He sees Anduril in the US, Kela from Israel and Stark in Europe as candidates in a market where deterrence—not celebrating “cost per kill”—is the objective.
Deep dive
1. Gigawatts replaced dollars as AI’s operative unit
Cahn’s 2024 call was that investors were treating AI as bits when its real bottlenecks were atoms: “steel, servers and power.” Operators were flying electricians to Texas, buying generator capacity sold out until 2030 and fighting for positions deep in constrained supply chains.
That physical framing became his “best trade of 2025”: power emerged as the constraint, and Sam Altman began discussing gigawatts rather than dollars. Construction, steel and other data-center activity also entered measured GDP, making AI one of the largest contributors to recent US GDP growth in Cahn’s telling.
The unresolved issue remains “is the customer’s customer healthy?” His 2024 arithmetic took $150 billion of Nvidia chips to roughly $300 billion of data-center investment, then required $600 billion of end revenue at a 50% gross margin. Recalculated in summer 2025, the question became approximately $840 billion.
2. Construction execution is becoming a moat—and Meta was a miss
The shovel is now hitting the ground, exposing the delays Cahn expected after declaring AI “shovel ready.” He rejects blanket predictions that every project will slip: outcomes will vary, and the companies able to coordinate scarce vendors and compoundingly complex supply chains will separate from the pack.
His largest unforeseen outcome was talent pricing. A perceived 25-year-old AI expert can receive a $50 million package, while a recognizable star might command $1 billion. The justification—raising the probability of a trillion-dollar outcome by 1%—is mathematically coherent, but humans probably confuse 1% with 0.01% or 0.001%.
Cahn also marked his bullish 12-month Meta prediction wrong. Vertical integration did not deliver the expected model performance, prompting extraordinary recruiting packages; still, Zuckerberg’s intense intervention illustrates why Cahn remains optimistic over a longer horizon and why founder-led companies can behave differently in a crisis.
3. The bubble can be real while selective application bets compound
“I do think we’re in an AI bubble,” Cahn said, noting that a contrarian view became consensus once Sam Altman, Vinod Khosla and Jeff Bezos acknowledged some form of excess. The useful questions now are which companies survive and what follows—not whether the label applies.
His tension is temporal: AI might be among the most important events in human history over 50 years, while a short market cycle destroys capital placed against a specific delivery date and chipset. Amazon surviving the dot-com collapse is the model for separating a durable company from its financing environment.
Eight years of investing lets Cahn avoid finding ten AI deals in one season. He backed Weights & Biases when deep learning was considered small, Runway before Stable Diffusion, and Hugging Face when “NLP” and a successor to BERT—not today’s expansive AI narrative—defined the opportunity.
Voice is one present application bet. Sesame’s conversational product reached one million users and five million minutes within weeks; it felt interruptible, memorable and non-robotic enough that Cahn decided within ten minutes to invest. His 10-year call: people will talk to—and form relationships with—their AI rather than remain trapped in phone screens.
4. Overbuilt compute enriches users and commoditizes owners
Cahn’s framework is deliberately simple: “Consumers of compute benefit from a bubble.” Overproduction pushes compute prices and application COGS down, lifting gross margins for companies that turn raw power into intelligence and products customers actually love.
Producers face the inverse. Even an excellent operator cannot control pricing when rivals manufacture the same commodity, which is why Cahn expects compute infrastructure to behave more cyclically and command lower multiples—closer to oil economics than software economics.
Harry’s pushback—worth keeping: AWS, Azure and Google Cloud are exceptional infrastructure businesses. Cahn’s answer was that their monopolistic positions were “hiding in plain sight”: neither cloud nor Google’s eventual scale was universally understood when those businesses began, allowing early leaders to establish durable market share.
AI offers no such concealment. Everyone knows an extraordinary company can reach $1 trillion, so everyone enters; Cahn thinks investors carry “too much monopoly and not enough commodity” in their mental models. Less economic rent is worse for suppliers but healthier for consumers, who want AI at low cost.
5. Capital deployment incentives still favor the wrong side of the trade
Harry observed that “consumers win” is now accepted venture wisdom. Cahn’s rejoinder: rhetoric changed, allocation barely did—he estimates more than 80% of AI dollars still flow to compute producers because infrastructure absorbs vastly more capital than applications.
That creates an institutional selection bias. Capital-heavy companies call investors continually, while efficient businesses may resist fundraising; yet the reluctant raisers can be the best investments. Sequoia’s Zoom investment was his example: the company was profitable and did not need the money.
There is no coordinating mechanism that makes the spending stop. Cahn sees roughly ten powerful players around a recursive chessboard, each reacting to the others through first-, second- and third-order incentives; the result looks orchestrated but is largely “uncoordinated and incentive driven.”
Therefore spending persists until incentives change. The bubble is game-theoretic without requiring a conspiracy or collective decision, and the companies consuming the resulting cheap capacity can benefit even if the capital providers eventually suffer.
6. Circular financing exposed the system’s “wobbly building”
Borrowing Nassim Taleb’s framing, Cahn said predicting the collapse is harder than identifying instability: “You can’t really predict when the wobbly building falls, but you can notice the fragility.” For AI, the most visible fragility is the growing circularity of infrastructure deals.
A year earlier, Microsoft and Amazon acted as risk absorbers, signing 20-year leases and buying out five years of generator capacity. Microsoft’s withdrawal from two data centers signaled that hyperscalers would no longer hold every “hot demand hot potato” for the ecosystem.
Oracle and CoreWeave then assumed much of the demand, but their smaller balance sheets cannot absorb equivalent risk. Chip companies consequently began financing the buildout; because funded purchases return as chip revenue, Cahn suggested their cost of capital in some deals “might even” be negative.
Announcements also obscure funding: projects may be only 10%-20% financed before sponsors raise the remainder. Cahn priced a gigawatt at about $40 billion, or Jensen Huang’s $50-$60 billion using Vera Rubin, and translated the advertised 100 and 250 gigawatts into “AI’s $8 trillion question” and “AI’s $20 trillion question.”
7. The dangerous mismatch is between permanent warehouses and temporary chips
Cahn sees nearly the entire capital machine pointing toward AI: private capital is concentrated there, while seven companies representing roughly 40% of the S&P 500 trade heavily on the same narrative. The risk is less the direction than the compressed period over which investors expect payoff.
Today’s construction is full of “B100s and H100s.” If the decisive capability instead requires Vera Rubin, Feynman—the 2028 chip—or ten years rather than two, an operator cannot upgrade a physical warehouse “with my fingers”; it owns a stranded building full of legacy equipment.
He therefore disputes the default 2008 analogy. Apart from exceptions such as Oracle, the buildout has largely been financed with cash and equity rather than credit, so an unwind is more likely to hit share prices and household portfolios than cascade first through bank balance sheets.
Alasdair Nairn supplied the concentration analogy: Japan once represented roughly 43% of the equity market and the US 41%, and avoiding Japan defined top performance after the reversal. The Magnificent Seven are stronger cash machines, Cahn stressed, but their shared sensitivity to one AI narrative is still concerning.
8. AI may transform GDP without producing monopoly-scale profits
Cahn broadly accepts Masa’s claim that AI could affect 5% or more of GDP over time. He rejects the next leap: assuming a 50% margin and roughly $4 trillion of economic profit treats today’s unusually monopolistic big-tech structure as the normal endpoint.
A McKinsey analysis he cited estimated that economic profit above the cost of capital equals only about 1% of global GDP. Most value flows to wages, workers and customers, making persistent excess returns difficult—and Cahn hopes AI’s benefits similarly accrue broadly rather than to a few companies.
Harry offered legal AI as evidence of inflated demand: essentially every law firm is shopping for a provider today because it has been told to find one, versus perhaps 5% in a more typical market. Cahn agreed that several variables are overstated but identified the timeline as the most consequential one.
The experienced voices are lengthening it: Karpathy proposed a “decade of agents” instead of AGI in 2027; Sutton questioned whether the current paradigm suffices; Sutskever said “pre-training is dead”; Altman described a “more gentle singularity.” Against that, junior-lab status games compress estimates to 100, 200 or 300 days.
9. Venture brands change probabilities but cannot manufacture winners
Harry argued that anointing a company with capital, distribution and brand can widen its talent and customer moat, citing Profound. Cahn’s answer was categorical: “I don’t believe in kingmaking,” because the stomach-punching venture lesson is “you can’t make a company succeed.”
He conceded a real flywheel. A Sequoia cap table can recruit five important engineers—especially candidates susceptible to social proof—and marginally improve outcomes. But Profound already had customers lining up and “the business was ripping” when it reached Sequoia’s investment committee.
The disagreement narrowed to magnitude: Harry sees brand and rapid follow-on capital as significant causal advantages; Cahn sees them as smaller than founders, product and existing market pull. Believing that a $20 million check creates “the Sequoia company” in a category is, in his view, how investment committees make mistakes.
The same caution applies after funding: “Capital is fuel, but capital does not create the engine.” Some founders behave as if cash is absent, but Cahn calls that exceptional; the engineer joining just after a billion-dollar raise and little revenue may internalize victory before customers have awarded it.
10. Margins can heal, but adoption speed is today’s cleanest signal
Gross margin matters directionally because it indicates how much proprietary product sits above foundation models, but Cahn refuses to treat it as destiny. He has watched a company move from 30% to 70%, while Snowflake’s initially criticized margins did not prevent it becoming a strong business.
AI provides an explicit repair mechanism: compute costs keep declining. Cahn can therefore imagine even some 0%-gross-margin businesses eventually working, although his own investments usually begin higher; obsessing over analytical purity can interfere with the actual job, “to make money” for LPs and founders.
He reframes “triple, triple, double, double” as the “zero to 100 club.” Harvey and OpenEvidence have demonstrated that trajectory, while Clay and Juicebox are among the companies he cited as on or having crossed it; reaching $100 million is not a formal requirement, but investors should believe exceptional companies can approach that velocity.
Because everyone is online and wants AI, a great product can spread far faster than an early internet company could. Harry said data he had seen supported focusing on the speed from $1 million to $50 million rather than time to the first $1 million; Cahn agreed, while preserving UiPath’s counterexample—nine years to $550,000 ARR.
11. Scar tissue and solvency outrank successive markups
Juicebox spent three years finding its market: its founders began with a college music application, then evolved toward AI recruiting. The CEO started around 22 and the CTO around 19; Cahn thinks enduring that painful search made them better founders once growth arrived.
Clay likewise spent three or four years “in the wilderness” after Sequoia’s Series A, which Cahn placed in 2019 with some uncertainty, changed substantially and later added Varun as a co-founder. Cahn subsequently led another investment slightly above a $1 billion valuation—evidence against the mythology that every great company races from seed through Series B within 12 months.
His survival rule is “anything multiplied by zero is zero.” Volatility is irrelevant to a durable business, but overextension followed by bankruptcy ends the compounding; “momentum has its own reality” until the distortion field disappears, when sober investors can counterbalance a founder’s necessary aggression.
Missing Datadog sharpened his concentration. An unconfirmed story about Dragoneer said it had cultivated Datadog for years as a top priority, so Cahn now spends roughly 80% of his time on his top five opportunities and another 20% on the next 15: selectivity is a response to scarce attention, not scarce deal flow.
12. Defense’s delayed ChatGPT moment will produce only a few champions
Challenged directly on Sequoia’s absence from Helsing and Anduril, Cahn conceded: “Sequoia was late to defense.” His counter-call is that “defense is the next AI”: the sector is roughly two years after the Transformer paper and has not yet reached its mass-recognition ChatGPT moment.
Harry questioned whether the thesis requires permanently rising conflict. Cahn instead centered deterrence: “You only go to war because you have to”; defense exists to prevent war as the world order changes. Modern companies have barely entered force structures, leaving perhaps 99% of a 50-year catch-up unfinished.
Buyer concentration is decisive rather than incidental. A single government customer rewards companies able to serve national objectives, so Cahn expects venture-funded R&D businesses to consolidate into a few national champions: Anduril in the US, Kela from Israel serving allied markets, and Stark as a prospective European champion.
Harry’s closing challenge was that defense is not really a category capable of supporting dozens of venture outcomes. Cahn agreed: unlike AI, where he hopes to make 20 more investments, defense may justify one every few years. Both rejected “cost per kill” framing in favor of safety and deterrence.