David Cahn on AI Infrastructure, Capital Spending and Venture Conviction

David Cahn frames AI infrastructure around “servers, steel and power.” He argues that large data centers may need frequent redesign as models and hardware change. His claim that a frontier model will never be trained twice on the same data center is a provocative prediction, not an established engineering rule.

Cahn tells Harry Stebbings that growing up with a twin gave him both companionship and competitive drive. He describes his father’s family fleeing Nazi Germany and his mother’s family immigrating from Syria, and says their sacrifices give him a sense of responsibility to make use of his opportunities.

His father coached him to focus on what came next rather than celebrate a score, Cahn recalls. Stebbings turns to Cahn’s AI writing: can the technology change the world while capital expenditure becomes too high? Cahn says those are separate questions.

Cahn says he has invested in AI companies including Hugging Face and Runway and served on the Weights & Biases board. He believes AI will profoundly change society over his lifetime, but argues that hundreds of billions spent in the next two years may be difficult to recover. His investments are relevant to his optimistic view.

He interprets recent comments by major technology executives as recognition that investment is risky but strategically necessary. If the technology delivers, the spending may pay off; if not, it may be hard to recover. This is Cahn’s interpretation of company strategy, not proof that executives agree with his calculations.

Large cloud companies may spend to protect their existing positions even without knowing exactly how AI investments will pay back, he says. He cautions against inferring guaranteed near-term AI revenue from the fact that firms are spending heavily. His market-size comparisons are approximate and specific to the interview period.

Cahn thinks excess compute supply could benefit startups through lower prices and improved margins, while incumbents bear much of the infrastructure investment risk. Stebbings counters that the cost of entry can strengthen incumbent power; Cahn agrees that both effects are possible. Neither outcome is guaranteed.

Compute is not an abstraction, Cahn argues: it comes from physical data centers, GPUs, cooling and labor. He uses a hypothetical $2 billion data center built over two years to show how chip generations and cooling requirements can change before a facility is ready. These are illustrative assumptions, not a universal construction cost or timeline.

Even if models become more efficient, Cahn thinks progress may make some infrastructure outdated. He quotes the idea that a frontier model will not be trained twice in the same data center and points to larger GPU clusters. That is an expectation about future hardware and scale, not evidence that existing facilities cannot be reused.

As models grow, he believes the data center itself becomes a critical asset. He outlines the chain from a cloud company to a real-estate developer, general contractor, subcontractors and electricians. His account of announced projects and labor costs is illustrative and not a verified industry cost breakdown.

Stebbings asks whether model research will outpace physical construction. Cahn says he does not have a strong answer: some researchers emphasize scaling, others expect breakthroughs in reasoning, data or efficiency. He sees stronger evidence for scaling at the time but hopes both paths advance.

Stebbings worries that AI features add inference costs without allowing software companies to charge more. Cahn says usefulness creates a potential reason to pay, but accepts that competition can erase pricing power. He expects the result to depend on barriers to entry and data advantages, rather than value alone.

On vertical integration, Cahn praises NVIDIA and Jensen Huang while expecting more coordination between data-center operations and model development. He points to Meta and Elon Musk’s efforts and asks how separately owned cloud providers and model labs will coordinate. His views are competitive assessments and predictions.

He argues that competing in frontier models requires a profitable business outside AI to fund investment. He contrasts cloud providers defending existing revenue with Meta using advertising revenue and having more strategic flexibility. This is a strategic hypothesis, not proof that every independent model provider must fail.

Cahn considers partnerships between cloud companies and model labs potentially durable, partly because large acquisitions may face antitrust barriers. He says smaller technology companies with distribution have opportunities but may struggle to match the largest firms’ spending. These are his expectations, not confirmed merger plans.

He thinks public-market investors are sophisticated, even if there can be a gap between Wall Street and Silicon Valley’s information. Asked about compute, algorithms and data, he suggests major labs have increasingly similar access and redirects attention to physical infrastructure: servers, steel and power.

In chips, Cahn expects improving price-performance and continued NVIDIA innovation, alongside competition from AMD, Broadcom and startups. He also flags dependence on Taiwanese manufacturing and the uncertain political and supply-chain consequences. These are forecasts and risks, not guarantees about chip markets.

By “steel,” he means construction and industrial equipment including generators and batteries. He says cloud companies ask manufacturers to expand capacity, but suppliers must decide whether demand will persist long enough to justify new factories. Large customers can help by committing to future purchases.

Cahn says he is looking for opportunities in data-center supply chains and real estate. He mentions investments by other firms and a long-duration battery company he backed at Coatue. His enthusiasm is an investor’s thesis and involves a direct investment interest, not a promise of returns.

He believes power demand from AI could accelerate investment in generation, storage and energy technology, and cites solar and battery projects. He expects a gradual increase in renewable energy and grid capacity rather than necessarily a sudden crisis. These are his projections; the interview does not quantify future supply or establish that demand alone will solve grid constraints.

On open versus closed AI models, Cahn says he does not expect AGI immediately and is comfortable with having both kinds of models. He expects AI to improve productivity and is less concerned about catastrophic risk than some others. These are his personal risk judgments, not evidence that open or closed systems are safe.

He expects China to make progress and warns against underestimating competitors, while remaining optimistic about US advantages. Asked whether his “$600 billion question” could become larger, he says it could. The number is a model of revenue needed to support spending, not a measurement of actual AI revenue or confirmed expenditure.

He explains the rough calculation: take NVIDIA’s annualized revenue, multiply by two for the assumed total data-center cost, then multiply by two again for the implied downstream revenue requirement. The multipliers are simplifying assumptions, and the result is not an audited forecast or a bill already owed.

Cahn expects more third parties to build data centers and lease them to large technology companies, spreading cash payments over time. He likens the commitment to debt and wonders whether it may make spending less visible. Accounting treatment depends on contract terms; a lease is still a long-term obligation even if financed externally.

He says real-estate investors may evaluate such arrangements primarily against the tenant’s credit and accept a yield above its bond yield. He speculates that more capital could compress the spread and cites structured financing as an emerging opportunity. These are illustrative investment views, not risk-free returns or universal deal terms.

Cahn describes himself as a student of psychology and connects Adam Smith’s invisible hand to predictable human behavior. He references Isaac Asimov’s Foundation as a fictional exploration of large-scale behavior. He treats capitalism as a system that channels human incentives, a philosophical interpretation rather than a predictive law.

Turning to venture investing, Cahn clarifies that he was an associate, not the lead investor, on several major deals at Coatue. He built models and spoke with customers. His lesson is to observe what customers actually do as well as what they say.

He recalls customers criticizing products from Marqeta, UiPath, Snowflake and Databricks or predicting they would switch, yet continuing to use them later. His point is that payment and usage can reveal value despite objections. The examples are retrospective and involve companies he worked on; they do not prove every customer stayed.

Early in his career, Cahn says he learned by identifying what each colleague did exceptionally well and trying to acquire part of that skill. He cites sourcing, leadership and loyalty. He sees both value in learning from others and the need to develop a personal investment style.

He defines venture success as generating very large gains and calls investors who do so “sluggers.” That is Sequoia partner Cahn’s preferred benchmark, not the only possible way to assess an investor or fund. He says learning from experienced partners matters, but each investor must ultimately make decisions and develop conviction.

Cahn says conviction means putting one’s reputation behind a company in a partnership meeting. He describes Sequoia’s limited number of investments per partner as a constraint that makes each decision consequential, and recalls a PayPal story about scarcity forcing focus. These are his descriptions of the firm and its philosophy.

He says Sequoia encourages new partners to contribute but is also a demanding environment because of its history and high bar. He describes being reshaped by the firm’s focus on building important companies. His praise of colleagues’ sourcing and selection reflects his position within the partnership.

Cahn proposes a founder framework with four dimensions: analytical and intuitive approaches applied to technology and people. He associates them with engineering, self-improvement, product intuition and leadership. His suggestion that larger companies require all four is a heuristic, not a validated prediction of company value.

He cites Supabase founder Paul as someone he believes embodies those qualities and says founders can develop them over time. Cahn says younger investors can offer to grow alongside founders, while experienced investors bring different strengths. His positive evaluation concerns a company he has backed.

Cahn ranks founders ahead of market and product in his assessment and says selection is central to venture. He thinks the Sequoia brand can help win a deal, but says his own deal flow still depends on outbound work and earning founders’ attention. His ranking is a personal framework.

To reach founders early in his career, he sent personalized Cameo videos and recorded short Loom demonstrations of products. He says the tactics helped start conversations. He still sees himself early in his career and hopes to develop the conviction to back a truly exceptional company.

Cahn describes waking at 5 a.m., exercising and cycling to the office, where he likes to meet founders early. His memorable first sourced investment was Starburst Data: he met founder Justin Borgman during customer diligence, built a relationship over about a year and eventually participated with a $30 million check, he says.

In quick fire, Cahn shares his religious belief, recalls meeting Supabase’s founder and says a transformational hire may justify unusually generous equity. The hypothetical 3–5% range is advice, not a valuation of a specific hire. He also says impending parenthood has changed his perspective.

He praises investor Mike Volpi and predicts OpenAI will remain important, citing ChatGPT’s brand and a revenue figure. The interview does not establish whether that figure is annualized or realized revenue, and the company’s future is uncertain. His favorable views of investors and startups reflect professional relationships.

Cahn says Sequoia raises the bar for selection even if its brand helps win deals. He finishes by revealing that he is learning to drive; Stebbings says he does not drive either. They joke about self-driving cars and close the interview.