Deepseek, Stargate and AI’s $600 billion question with Sequoia Capital’s David Cahn
Summary
David Cahn reads DeepSeek as evidence that China reached GPT-4 parity, not as an unexpected frontier breakthrough. The more consequential signal, he argues, was Ilya Sutskever’s suggestion that “pre-training is dead” or “scaling laws are dead”: competitors were always going to copy the existing generation, while the source of the next leap remains unknown. His ranking is explicit: “DeepSeek maybe is overhyped, and what Ilya said is maybe a bit underhyped.”
Cheaper models are bullish for AI applications even if they threaten scarcity-based infrastructure assumptions. Lukas Biewald notes that decades of falling compute costs have usually driven greater aggregate usage; Cahn likewise calls cheaper inference “great for the application layer” and says the market “sort of freaks out at the wrong moments,” reacting to commoditization that was visible six months earlier. Satya Nadella’s longstanding argument—that models would be distilled and become much cheaper, benefiting hosts such as Microsoft—now looks central.
Stargate’s $500 billion ambition and DeepSeek’s efficiency represent competing capital-allocation regimes. Stargate assumes larger data centers keep producing better models, while DeepSeek points toward smaller models and value migrating upward to applications. The revealing change for Cahn is Microsoft stopping at its existing $80 billion data-center commitment, which may mark a shift from hyperscaler balance-sheet funding to “credit-funded data centers or leveraged data centers.”
Cahn’s $600 billion question remains unresolved because one year of infrastructure investment requires an enormous downstream revenue base. His napkin math starts with roughly $150 billion of NVIDIA GPU run-rate revenue, doubles it to $300 billion for power and data-center infrastructure, then doubles it again so applications can earn 50% gross margins. Another equivalent investment year takes the implied obligation to $1.2 trillion—“almost this debt that we’ve sort of invested in that we now have to go pay back over time.”
Hyperscaler capex is stabilizing, but AI revenue has not caught up. Cahn estimates Microsoft at roughly $20 billion per quarter and Google at $13 billion, with Amazon likely stabilizing in the low $20 billions and Meta in the low teens; that keeps the question from immediately becoming a trillion-dollar question. Yet OpenAI still represents the lion’s share of ecosystem revenue, leaving something close to the earlier “$500 billion hole.”
The spending persists because the cloud oligopoly is trapped in a strategically rational prisoner’s dilemma. A roughly $500 billion cloud market is the “golden goose” for Microsoft, Amazon, and Google, while seven companies account for 33% of the S&P 500; existing profits therefore finance the race. Executives believe the opportunity is immense, but their defensive logic is equally important: “We have to spend, or we’re going to fall behind.”
Cahn’s clearest application-layer thesis is profession-specific AI search built around how users actually think. He uses Perplexity 10–20 times daily and sees parallel products for lawyers and doctors, with differentiation across intent extraction, proprietary data, answer formatting, and “cognitive architecture.” The opportunity is an “AI search pal” mapped to a profession’s problem-solving patterns, not merely a generic model with a new interface.
His venture framework favors founders and problems he can still believe in when fashion reverses, while demanding far more than technical superiority. A product that is “20% better” may be commercially irrelevant if a busy buyer cannot see why switching matters; conversely, patience paid off at Weights & Biases, and deep conviction enabled early bets on Runway and Form Energy. That same search for substance culminates in AI’s deepest uncertainty: whether it becomes merely superhumanly intelligent or also self-reflective and conscious—“fundamentally, that’s a religious question in many ways.”
Deep dive
1. DeepSeek validates catch-up, not the next scaling regime
Cahn describes DeepSeek as a smaller, distilled Chinese model trained partly on outputs from existing systems, with unusually low training cost. The market treated it as evidence that foundation models may be commoditizing, but he calls it an update to the prevailing mental model rather than an astonishing discontinuity.
His timeline begins with GPT-4 in mid-2023. Google, Meta, xAI, and others reached comparable quality during 2024; DeepSeek showed that China had now caught up as well. “This is always how tech works: you come out with a breakthrough, and people catch up.”
The unresolved question is what follows parity. Cahn tells people to watch Ilya Sutskever’s NeurIPS talk three times because its “pre-training is dead” implication challenges the scaling race itself: matching yesterday’s model matters less than discovering tomorrow’s training regime.
2. Stargate transfers the data-center wager from balance sheets to leverage
Stargate proposes spending $500 billion on data centers just as DeepSeek suggests comparable capabilities might require less compute. Cahn preserves both possibilities: bigger clusters might yield AGI, while cheaper models might redirect the industry’s renaissance toward builders and end-user applications. “None of us know where the world is going to go.”
The new information was Microsoft’s revealed preference. It had held a right of first refusal on new OpenAI data-center building, yet after Stargate, Nadella effectively said he was “good” for Microsoft’s existing $80 billion commitment and was not increasing it.
Drawing on Ben Thompson’s analysis, Cahn sees a transition from equity-funded infrastructure—paid for from Microsoft, Google, and Amazon cash flows—to “credit-funded” or leveraged data centers that must be funded from the cash flows they generate. He calls that logical for a maturing market, while noting its resemblance to the late stage of the dot-com buildout.
3. AI still owes the infrastructure stack $600 billion of revenue
Cahn’s original question was disarmingly simple: “Where’s all the revenue?” For an estimated $150 billion of NVIDIA GPU run-rate revenue in 2024, he assigns another dollar of power, generators, buildings, and related infrastructure to every GPU dollar, producing $300 billion of total data-center investment.
The second multiplier belongs to applications. If a startup spends one dollar accessing AI infrastructure and targets a 50% gross margin, it must earn two dollars from customers. Applied to the $300 billion stack, that creates $600 billion of required ecosystem revenue to support one investment year.
That obligation accumulates rather than disappearing: another $150 billion of GPUs in 2025 implies another $600 billion of downstream revenue, taking the two-year figure to $1.2 trillion. Cahn characterizes the installed investment as “almost this debt” that future AI revenue must repay.
Capex has stabilized rather than continuing to grow exponentially. Cahn puts Microsoft near $20 billion per quarter and Google around $13 billion, expecting Amazon to settle in the low $20 billions and Meta in the low teens. The $600 billion question therefore may not immediately become a trillion-dollar question—but the existing gap remains.
4. Cloud’s prisoner’s dilemma keeps financing the gap
On the revenue side, Cahn says little has fundamentally changed since July 2024. OpenAI and Anthropic are larger, but OpenAI remains the lion’s share of AI revenue; big technology companies have not fully monetized AI, and Google’s effort to distribute its product through Gmail illustrates that the commercial model is still being worked out.
His earlier accounting, deliberately generous to startups and incumbents, still left a roughly $500 billion hole. Cahn is nevertheless optimistic that revenue will catch up; the number is useful because it forces founders and investors to seek products important enough to justify the infrastructure already built.
The immediate funding source is the existing cloud business, itself roughly $500 billion and highly profitable. With seven companies representing 33% of the S&P 500, Microsoft, Amazon, and Google can divert oligopoly profits into AI rather than wait for the new AI revenue to arrive.
That produces Cahn’s “prisoner’s dilemma of AI”: each hyperscaler genuinely sees immense upside, but each also fears losing its golden goose if a rival spends more. Zuckerberg and Sundar Pichai voiced versions of the same logic—payback is uncertain, yet underinvesting could mean falling behind.
5. Profession-shaped search is the clearest path to application revenue
Cahn calls AI search the killer app and uses Perplexity 10–20 times daily. Traditional search navigated users toward an answer; AI search traverses the web and pretrained knowledge, synthesizes the material, and returns the answer directly—an “order-of-magnitude improvement” for his work as a researcher-investor.
His stronger claim is that Perplexity maps itself onto his “theory of mind.” Professions create different patterns of thought, so products should reflect how an investor, lawyer, engineer, or doctor frames a problem rather than merely exposing the same general-purpose model.
Harvey applies that idea to lawyers, while OpenEvidence searches medical literature around a doctor’s patient problem. Each could become an “AI search pal” used 10 or 20 times daily. Lukas pushes the boundary: if research is defined broadly, perhaps most white-collar jobs already fit this thesis; Cahn agrees many do, but retains uncertainty about how universal it is.
Cahn identifies four dimensions for differentiation: extracting domain-specific intent, obtaining proprietary, synthetic, or human-generated data, formatting answers for the task, and designing a cognitive architecture around the user. A consumer may want long prose about philosophy; someone asking for an auto-market size wants the number immediately.
6. Great products need patience—and buyers willing to switch
Cahn’s first thesis came from seeing himself as a “seven-out-of-ten software engineer” in a world that would create many more developers like him. Productivity infrastructure led him to Supabase, used by more than 30% of YC companies; Replit, with tens of millions of users; and open-source commercialization models such as Confluent, Databricks, and Starburst.
AI followed when he reframed Snowflake: he saw Snowflake as on a path to become a $100 million company because it makes charts that support business decisions, then asked whether software could instead make decisions and automate processes—a potentially trillion-dollar opportunity. Weights & Biases customers supplied concrete evidence, including John Deere using AI to automate fertilizer distribution.
Weights & Biases taught patience. The product and team were strong, and serious deep-learning practitioners already used it, but the market remained tiny until momentum arrived around early 2020. “When you have a great team and a great product and you believe in the market, you just kind of have to wait.”
Hugging Face taught competitive selling: amid roughly 13 term sheets, Cahn spent a week in a Miami hotel, built a bot based on the Transformers library explaining ten reasons to partner, and recruited ten founders to call Clem. Failures supplied the counter-lesson: if a product is only “20% better, forget about it”—a busy buyer needs a compelling reason to notice and switch.
7. Long-horizon conviction changes which impossible bets get funded
Cahn now asks whether, five years after investing—when others may have abandoned the company—he will remain a “true believer.” Lukas challenges the downside: early-stage investing means many companies will not succeed, so conviction could become attachment. Cahn’s answer is that the same test also enables affirmative bets that look unappealing in the short term.
Runway qualified before Stable Diffusion was part of its story. Its three founders cared so deeply about video and art that Cahn believed they would pursue the problem regardless; with enterprise video already expanding, he wanted to work with them even before the product that ultimately mattered had been invented.
Form Energy was more radical: grid-scale batteries, possibly billion-dollar installations, proposed by Tesla’s former energy leader and a co-founder Cahn describes as one of the world’s top three battery scientists. After months of diligence, the Harvard professor he hired asked to be paid in company stock; five years later, Cahn says Form had built a billion-dollar West Virginia factory.
Buffett’s imagined 20-punch lifetime investment card sharpens the filter: both the people and the problem must deserve a punch. Cahn’s unfilled frontier is the “Shenzhen problem”—rebuilding deep Western supply chains, advanced manufacturing, and automation—plus AI robotics that he says would surprise him not to see in homes and factories within 20 years.
8. AI reopens an ancient religious question
Cahn treats each one-on-one founder relationship as venture’s “atomic unit of work.” Public visibility is not necessary; private conviction, candid pressure, and support “through thick and thin” are. His mantra, “everybody knows everything,” reflects a belief that people are more transparent to one another than polished public performances suggest.
His father taught him to separate form from substance: teachers and university presidents are people, not omniscient roles. The venture version is the choice between looking smart and being right—joining a fashionable deal can confer two years of brilliance, while a substantively correct, unpopular bet may look foolish until year seven.
Religion extends that same inquiry beneath institutional form. Cahn views traditions as different packages around a shared confrontation “with consciousness, with God, with truth,” all attempting to answer what it means to be human. Their thousands of years of accumulated analysis make them relevant to AI rather than obsolete beside it.
Gödel, Escher, Bach gave him a nonreligious formulation: human distinctiveness may lie in recursive self-reflection. AI will probably exceed his IQ and may make better decisions, but whether it knows that it exists—and could have the kind of self-reflection tied, for humans, to awe and empathy—remains open. The key fork is between “an unconscious, really intelligent AI” and “a conscious, really intelligent AI.”