Most important take away

The real bottleneck and value capture in AI today is the physical industrial layer — servers, steel, and power — not the models themselves, because frontier training requires brand-new data centers every cycle and big tech is locked in a prisoner’s-dilemma capex race to defend a $250B cloud oligopoly. For builders and investors, the actionable corollary is that startups (consumers of compute) win as compute prices fall, while the biggest near-term money will be made in the unglamorous supply chain: data center real estate developers, general contractors, generator and battery makers, and power/utility companies.

Summary

David Cahn, partner at Sequoia and author of the widely-cited “AI’s $200B/$600B Question” essays, argues that believing in AI and believing current capex levels are rational are two separate questions. He thinks the hyperscalers (Microsoft, Google, Amazon, plus Meta) are spending hundreds of billions not because they have a paid-back ROI model, but because they are in a prisoner’s dilemma defending a $250B cloud oligopoly that represents ~10% of global market cap. This is good news for startups: overproduction of compute means falling prices, which translates directly into higher gross margins for consumers of compute.

Actionable insights and patterns for technologists and operators:

• Compute is physical, not abstract. A data center is a ~$2B Illinois facility full of GPUs, liquid cooling, and contractors. “No one trains a frontier model on the same data center twice” — by the time training finishes, chips and cooling are outdated. Architect for replacement, not permanence.
• Servers, steel, and power are the three pillars. Chips (Nvidia/AMD/Broadcom), construction/real estate (Cyrus One, QTS, DPR as general contractor), and energy generation (NextEra, long-duration batteries) are where the industrial value is flowing.
• Vertical integration matters. You can’t have separate teams running the data center and building the model — Meta and xAI vertically integrate; OpenAI/Anthropic effectively rely on their parent hyperscalers’ data centers. To play in frontier models, you need a non-AI cash machine (Instagram, AWS, Azure).
• AI features only command pricing power with real value + barriers to entry. Injecting AI to check a marketing box just adds cost. Pricing power survives only where there are data moats or structural barriers; in commoditized categories, AI features race to zero gross margin.
• Listen to what customers do, not what they say. Cahn’s lesson from diligence on Marqeta, UiPath, Snowflake, Databricks: customers routinely said they’d rip products out and build internally; almost none did. Watch behavior and willingness to pay, not stated intent.
• Off-balance-sheet financing of data centers is emerging. Hyperscalers are shifting toward 20-year leases on third-party-built data centers — effectively capitalized-lease debt — which compresses spreads but unlocks more supply. Real estate developers (KKR/Cyrus One, Blackstone/QTS) are major beneficiaries.
• Energy is the next bottleneck. AI capitalism will drive more energy/grid investment than the IRA ever could. Solar, batteries, and grid modernization quietly become massive markets.

Career advice from Cahn:

• Identify each strong person’s superpower and aim to be ~80% as good. If you can be 80% as good as 10 different people on their 10 different superpowers, you’ll be exceptional. Drink from the firehose early.
• Conviction is the one thing nobody else can give you. The hardest part of venture isn’t picking good companies — it’s deciding when to put your neck on the line. Partnerships are structured to test that conviction.
• Constraints create greatness. Sequoia limits partners to 1-2 investments/year, forcing the question “is this one of my one or two?” Roelof Botha’s PayPal lesson: they were great when they had to be great.
• Play the cards you’re dealt. As a young investor, the value proposition to founders is “we’ll grow together over 20 years” — don’t try to play the veteran’s game.
• Earn the right to win. Founders don’t owe you their time. Cahn’s tactics included sending cameo videos from a founder’s favorite actor and recording a daily Loom video for over a year demonstrating cool products to prospects.
• Founder framework (2x2): Science vs. Intuition × Technology vs. Humans. Science+Tech = engineer. Science+Human = hardcore mindset/self-optimization. Intuition+Tech = product visionary (Ivan at Notion, Chesky at Airbnb). Intuition+Human = leadership. One trait makes you good; two great; three a billion-dollar company; all four a $10B+ visionary.
• Pay transformational hires real equity. If someone can double the value of your business, give them 3-5% — don’t anchor to standard bands.
• Founder > market > product, and within venture skills, selection > winning > sourcing in importance (though sourcing must be mastered early in career).

Chapter Summaries

Childhood and formation. Cahn describes being a twin as the single most important fact of his life — it grants permission for non-conformism and breeds hyper-competitiveness. His family history (paternal side fled Nazi Germany, maternal side from Syria; both parents first-generation college) drives a what’s-next mentality.

The $600B Question and game theory of AI capex. Believing AI changes the world and believing current capex levels pay back in 24 months are different claims. Zuckerberg and Pichai have implicitly conceded the spend is speculative; they’re forced into it to defend a $250B cloud oligopoly. This is great for startups (cheaper compute, higher gross margins) but a real risk for the hyperscalers.

Data center as the most important asset. Compute is physical — Illinois data centers, electricians flown in, $2B per facility. Frontier models can’t be trained twice on the same data center. Scaling laws plus rapid GPU generations (H100 → B100) force constant rebuilds. Vertical integration between model team and data center team is becoming mandatory (Meta, xAI).

Pricing power and AI features. AI integration only earns margin where the product genuinely becomes more useful AND there are barriers to entry. In commoditized markets, AI features race to zero gross margin.

Servers, steel, and power. Cahn’s three pillars. Chip wars accelerating (Nvidia dominant but AMD/Broadcom chasing fat margins). Construction supply chain (DPR, Cyrus One, QTS) is where the real industrial action is — and where Cahn is spending his time. Power/energy revolution will be driven by AI capitalism more than by the IRA.

Open vs. closed source and AGI. Cahn isn’t worried about near-term AGI risk; happy both options exist. Notes Ethan Mollick’s point that VCs claiming to believe in AGI while still funding SaaS are being inconsistent.

China and US competitive position. Bullish on America’s culture, immigration, and capitalism but warns against underestimating China.

Off-balance-sheet financing. Hyperscalers shifting to 20-year leases of third-party-built data centers — essentially capitalized-lease debt — to make capex look smaller. Real estate investors and Blackstone-style structurers are making strong risk-adjusted returns underwriting hyperscaler credit at a spread.

Lessons from Coatue investments. Marqeta, UiPath, Snowflake, Databricks: customers all said they’d churn or build internally; they didn’t. Listen to behavior, not stated intent.

Joining Sequoia. “Sequoia hires investors, breaks them down, and rebuilds them in the Sequoia mold.” Constraint of 1-2 investments/year forces conviction. The bar is being one of the 50 most important companies in the world.

Founder assessment framework. The 2x2 of science/intuition × technology/humans. All four traits = visionary founder (Cahn cites Paul Copplestone at Supabase).

Personal practice. In office by 7am, gym at 5:30am, bikes through Stanford up Sand Hill Road. First lead deal was Starburst Data — found Justin Borgman during Databricks customer diligence, cultivated him for 12 months before he raised.

Quickfire. Believes in God (contrarian in SV). Pay transformational hires 3-5%. Having a baby in two months. Picks Mike Volpi as the VC he’d most swap portfolios with. OpenAI is the foundation model company most certain to exist in 10 years due to ChatGPT brand and consumer distribution. Doesn’t know how to drive (NYC native, currently taking lessons).