Unique Research · Insight · October 24, 2023 · Adapted from Sequoia Capital, “AI’s $200B Question”
The generative-AI wave that began last summer has entered a second, higher gear. The catalyst: Nvidia’s Q2 earnings beat. Markets read it as proof that demand for GPUs and model training capacity is nowhere near saturated.
Consumer products such as ChatGPT, Midjourney, and Stable Diffusion had already put AI in the public eye. Nvidia’s numbers then showed founders and investors something harder: AI can already produce billions of dollars of net revenue for the right suppliers. Investment pace and valuations followed.
One giant question remains: what are all these GPUs actually for? How do end customers use them to create value for their customers — and how much revenue must that stack generate before today’s CapEx pays off?
GPU opex and the AI P&L gap
Rule of thumb Sequoia uses: for every $1 spent on GPUs, expect roughly another $1 of data-center energy and operating cost. If Nvidia’s GPU run-rate hits about $50 billion by year-end (a conservative take on analyst forecasts), associated data-center spend is on the order of $100 billion.
End users — Starbucks, X, Tesla, GitHub Copilot, or a new startup — also need margin. Assume they need ~50% gross margin: then those GPUs must support roughly $200 billion of steady annual revenue just to clear the stack. That figure still excludes cloud-provider profit. The real bar is higher.
Public filings show much of the incremental data-center CapEx coming from Big Tech — Google, Microsoft, Meta — with ByteDance, Tencent, and Alibaba also named as large Nvidia customers. Looking ahead, Amazon, Oracle, Apple, Tesla, and CoreWeave should contribute meaningfully.
The $200B question: how much of this CapEx is tied to real, present end-customer demand — and how much is speculative build for demand that may arrive later?
Even generous AI revenue leaves a hole
Public reporting puts OpenAI around $1B annual revenue. Microsoft has spoken of ~$10B of AI revenue from products such as Copilot. Sequoia’s thought experiment gives Google (Duet/Bard), Meta, and Apple similar ~$10B AI revenue each, and ~$5B each to Oracle, ByteDance, Alibaba, Tencent, X, and Tesla. Even under those generous assumptions, more than $125 billion of annual CapEx still lacks a matching revenue story.
That gap, Sequoia argues, will largely be filled by the startup ecosystem.
From infra build to end-customer value
The point of the essay is the supply–demand gap in plain sight. Deep-learning breakthroughs since 2017 finally meet a capital boom. Overbuilding infrastructure is expensive in every tech cycle — and historically also collapses the marginal cost of building new products. Sequoia expects the same pattern in AI: today’s GPU scarcity becomes tomorrow’s abundance; inference costs fall; more builders enter.
For startups, the mandate is clear: shift attention from infrastructure to end-customer value. Infra is being built. Many foundation models are in flight. Usability has many paths. The $200B question reduces to: what will you do with all of this — and how will it change people’s lives?