David Cahn on AI Infrastructure, Startup Growth and Defense Investing

Harry Stebbings interviews Sequoia Capital partner David Cahn about AI infrastructure, investment cycles, startup growth, hiring and defense. Cahn argues that AI is in a bubble but can still transform the world over time. He discusses businesses Sequoia and he have invested in, so his assessments carry financial and professional interests.

Cahn says his earlier focus on steel, servers and power anticipated the physical constraints of AI data centers. He describes competition for electricians, generators and energy, arguing that building capacity depends on physical supply chains as well as models and data. His characterization of the scale of AI's contribution to GDP is an assessment, not an independently verified calculation.

His earlier '$600 billion question' was a simplified model: $150 billion in Nvidia chips implied roughly $300 billion in data-center investment and $600 billion in revenue at a 50% gross margin. He says updating the exercise in summer 2025 yielded about $840 billion. These are model outputs based on assumptions, not measured revenue requirements for every project.

Stebbings asks whether construction delays will proliferate. Cahn expects outcomes to vary and argues that the ability to build data centers can be a moat, especially when everyone competes for the same suppliers. He recommends tracing the supply chain beyond the large companies commissioning facilities.

Cahn identifies two surprises from the previous year. First, he says he did not anticipate exceptionally large AI talent packages. A company might justify an enormous award if a researcher meaningfully increased its chance of winning a huge market, but he warns that people may overestimate the individual's effect on that probability.

Second, Cahn admits his prediction that Meta would perform well in AI was wrong over the preceding 12 months. He had expected its vertical integration and founder leadership to help and remains optimistic over a longer horizon. That is a forecast, not evidence that the prior prediction was correct.

He argues that leading model companies are moving down the stack into power, chips and infrastructure, and predicts that competitive pressure will encourage more providers to focus on vertical integration. Stebbings notes that OpenAI and Anthropic did not initially have the same integrated structure as large technology companies.

Cahn believes AI is experiencing a bubble, but distinguishes the risk of a painful market cycle from the long-term importance of the technology. He invokes Amazon after the dot-com crash to illustrate how a company can survive a bubble; neither survival nor AI's future impact is guaranteed by that analogy.

Having invested in AI for years, Cahn says he wants one or two opportunities a year that he believes can endure market volatility. He mentions earlier investments in Weights & Biases, Runway and Hugging Face, and more recent investments in Clay and Juicebox. His favorable outlook on these companies is an interested investor's judgment.

His test is whether a company can succeed without relying indefinitely on cheap, abundant capital. Strong customer demand, product-market fit and a capable team may help a business through a downturn, he argues, though they cannot guarantee survival.

Cahn's framework distinguishes consumers from producers of compute. If capacity becomes oversupplied and prices fall, consumers may see lower costs and higher margins; producers face the risks of a commodity cycle. This is a conditional economic argument, not a prediction that every application will win or every infrastructure provider will lose.

Stebbings challenges the framework by citing Google Cloud, AWS and Azure. Cahn responds that today's powerful incumbents benefited from opportunities that were not initially obvious, whereas the apparent size of AI is drawing intense competition. His description of large firms as monopolistic and prediction of limited monopoly profits in AI are interpretations, not legal findings.

Stebbings suggests that backing compute consumers has become a common view. Cahn replies that much AI capital still goes to compute producers because they require more financing, and argues that capital-deployment incentives can pull investors toward capital-intensive companies. His percentage estimate is not accompanied by a disclosed calculation.

Asked whether spending can stop through coordination, Cahn says companies act on their own incentives and their decisions interact. In his view, behavior will change when incentives change, rather than through a single coordinating mechanism. He describes fragility without claiming to know exactly when or how a correction will occur.

Cahn points to circular financing arrangements as a reason investors have become more concerned. He describes a shift in perceived infrastructure risk from Microsoft and Amazon toward smaller participants such as Oracle and CoreWeave, with chip companies also financing demand for their products. These are his interpretations of complex arrangements, not a transaction-by-transaction audit.

He says some announced infrastructure deals still need substantial financing. His illustrative estimate puts construction at roughly $40 billion to $60 billion per gigawatt; he extrapolates much larger revenue questions for 100 or 250 gigawatts. Those scenarios depend on cost, capacity and revenue assumptions, and do not establish that the projects are fully funded or will be built.

Cahn worries that capital is concentrated into a narrow period and particular generations of chips. If useful capabilities take longer to arrive, hardware already installed may age before demand catches up. He suggests spreading risk over more time and bets, while maintaining a long-term belief in AI.

Stebbings asks about Oracle's leverage. Cahn argues that much of the buildout has been financed with equity and cash and speculates that a downturn could show up more in equity prices than in a 2008-style credit unwind. He sets Oracle apart and does not provide a full financing breakdown; either type of risk could vary by company.

Both discuss the concentration of market value in large technology companies. Cahn cites a historical comparison with Japan and says a change in the AI narrative could affect a large share of the market, even if the companies have strong underlying businesses. The historical percentages he recalls are not independently established in the conversation.

On a proposed large AI contribution to GDP, Cahn agrees that the technology could affect a substantial share of economic activity but questions an assumption that firms will capture a 50% profit margin. He argues that competition may spread gains more widely. Both the projected GDP effect and the distribution of gains are uncertain.

Stebbings asks whether current demand is temporarily inflated, citing legal firms rushing to choose AI providers. Cahn stresses uncertainty around timing and says some researchers have described a longer, more gradual development path. His account of others' timelines is his interpretation rather than a definitive consensus or an established AGI schedule.

Stebbings argues that a prominent investor can 'make a king' by providing capital, brand and distribution, naming Harvey as an example. Cahn disputes that a venture firm can make a company successful: he credits founders, ideas and product-market fit, saying investors can help with difficult decisions but should remain humble.

Stebbings pushes back using Profound, arguing that Sequoia's involvement and a subsequent round can help attract talent, customers and financing. Cahn agrees that a well-known investor can improve recruiting and influence outcomes, but thinks the average effect is smaller than investors imagine. Sequoia's involvement in Profound is relevant to his account.

Asked whether AI application margins matter, Cahn says they are a useful directional indicator of how much product has been built on top of a model. He recalls an investment whose gross margin rose from 30% to 70%, and says falling compute costs could improve economics. This example does not show that every low-margin company will become profitable.

Cahn can imagine even a company with zero gross margin eventually working, although he says his own investments generally have higher margins. He cautions against rejecting a valuable business solely on current margin. Whether costs fall, customers stay and margins improve remains company-specific.

Stebbings asks whether the old 'triple, triple, double, double' growth playbook has been overtaken. Cahn describes a 'zero to $100 million' club: he looks for companies he believes can reach $100 million in revenue quickly, though he says they need not already be there. He cites Harvey, OpenEvidence, Clay and Juicebox as examples of companies on that trajectory.

Cahn treats rapid adoption as a sign of strong product-market fit in a market with widespread internet access and AI demand. He also says the ultimate aim is a lasting company, not speed alone. The examples include Sequoia investments, and revenue growth by itself does not prove durable customer value.

Stebbings notes that UiPath took years to reach modest annual recurring revenue. Cahn agrees that founders may spend a long time finding the right product before rapid growth begins. He recalls Juicebox's founders working for roughly three years on their recruiting product; he is an investor and his assessment of their progress is interested.

He gives Clay as another example of a long search for product-market fit. Cahn says Sequoia invested at the Series A and he later led a growth-stage investment at a valuation a little above $1 billion. He sees the difficult early years as valuable founder experience, but that is an investor's account rather than a guarantee of future performance.

Stebbings asks about companies raising successive rounds at quickly rising prices. Cahn says abundant capital can be useful but does not create the engine of a business. In his view, overcapitalization can create a premature sense of success and make disciplined hiring harder, even if some founders manage the trade-off well.

Cahn says a company that overextends and then goes bankrupt cannot benefit from a later market recovery. He describes momentum as a potential source of overconfidence and hopes investors can help founders think through survival and risk. His warning is a general framework, not a forecast about a specific business.

Asked what people underestimate, he chooses 23- and 24-year-olds. He says he meets hundreds of recent graduates to learn from them and recruit them into portfolio companies, spending substantial time at Juicebox. In his view, people who grew up with current AI tools may bring a valuable perspective.

Cahn suggests AI startups may benefit from hiring adaptable generalists rather than relying only on senior software engineers. Stebbings raises emotional maturity and lack of experience. Cahn accepts the trade-off, arguing that risks in a junior hire are visible while the risks associated with an experienced hire can be harder to notice; these are generalizations, not rules about individual candidates.

For young people choosing jobs, Cahn describes a 'mimetic algorithm': following admired people one year ahead. He thinks this can work but may miss a major change such as AI, because previous cohorts made choices with different information. He advises factoring in the technology while choosing work that fits personal priorities.

He distinguishes people seeking what a job can offer them from those asking what they can contribute, and says both approaches are legitimate. His rough percentages are impressions, not measured population shares. Stebbings says many UK students still pursue banking and consulting; Cahn says he has recently seen more young people seeking AI roles.

Stebbings challenges Sequoia for missing prominent defense companies Helsing and Anduril. Cahn acknowledges that Sequoia was late to defense and says the firm is working to catch up. He likens the sector to an early stage of AI's development, a comparison that reflects his investment thesis rather than a proven trajectory.

Cahn sees the war in Ukraine as a signal of change in military technology, while arguing that a broader turning point has yet to arrive. Stebbings questions whether the defense thesis assumes escalating conflict. Cahn emphasizes deterrence and describes his view that defense systems need modernization; the effects of particular technologies on conflict are not established by the interview.

Stebbings worries that defense suppliers face concentrated government buyers and difficult sales incentives. Cahn agrees that the customer structure may support relatively few successful companies and predicts national champions in individual countries. This concentration could constrain entrants; his predicted consolidation is not certain.

Cahn names Anduril as his perceived US leader and discusses Kela in Israel and Stark in Europe, both investments he says he has made through Sequoia. He believes those companies could become national champions. These are forward-looking claims from an investor with a direct interest, not independent assessments of the companies or their technologies.

Stebbings says defense may support only a small number of large companies rather than an ecosystem as broad as software or fintech. Cahn agrees that he expects to make only a small number of defense investments, and says his framework emphasizes defense and deterrence rather than reducing decisions to a cost-per-kill metric.

In the quick-fire round, Cahn says he finally learned to drive after waiting years for autonomous cars, prompted in part by becoming a father. He says parenthood has made his priorities more concrete and describes shared values as important in his marriage.

His investment miss was Datadog: he recalls losing an opportunity to Dragoneer, which he understood had pursued the company for years. He did not confirm that account with Dragoneer. The lesson he drew was to focus most of his time on a handful of highest-priority opportunities.

Cahn believes voice is an underappreciated interface for AI. He announces Sequoia's investment in Sesame and cites its launch usage figures of one million users and five million minutes in a few weeks. These figures and his assessment of product-market fit come from an interested investor and do not independently establish retention or financial performance.

He predicts people may develop ongoing relationships with voice AI over the next decade, while acknowledging that the idea currently sounds like science fiction. Asked what excites him most about the next ten years, he returns to the potential of AI to transform lives. The pace and consequences he anticipates remain uncertain.

Stebbings and Cahn close by joking about taking a drive together. Cahn says he is excited to experience the changes ahead; the interview ends with optimism despite the investment and infrastructure risks discussed throughout.