Can AI answer the $3 trillion question?
Three years ago, Sequoia partner David Cahn calculated that $200 billion in AI revenue would be needed to justify Nvidia’s GPU revenue. Now, with hyperscaling, he estimates $1.5 trillion in infrastructure spending for 2026, requiring $3 trillion in revenue to pay back — and that may be an underestimate due to rising memory and chip costs. On the other side, OpenAI and Anthropic are earning billions but still far from closing the gap. Apollo‘s chief economist Torsten Slok warns that hyperscalers (Google, Meta, Microsoft, Amazon) are projecting massive free-cash-flow acceleration by 2028, but if that doesn’t materialize, the market reaction could be severe — potentially tipping the economy into recession. The tension is between the enormous upfront investment and the uncertain revenue from AI products, especially as token prices fall and organizations turn to cheaper open-weight models.
The concrete approach laid out is Cahn’s updated math: starting from Nvidia’s GPU revenue, adding data center operating costs and margins, he arrives at a required revenue of $3 trillion. He notes that recent bottleneck dynamics and rising construction costs have increased the required revenue per gigawatt of CapEx. Meanwhile, Slok highlights the risk that cheaper models (e.g., from China) and falling token prices could undercut the revenue needed for hyperscalers to hit their cash flow targets. OpenAI’s CEO Sam Altman claims their latest model is 54% more token efficient on coding tasks, which benefits users but may reduce overall token consumption from the infrastructure built.
The takeaway for builders and investors is that the AI infrastructure build-out is a massive bet on future demand, but the unit economics are shifting toward efficiency and lower costs. If users don’t dramatically increase token usage, the revenue gap may remain. The macroeconomic stakes are high: a slower payoff for hyperscalers could trigger a correction. For serious builders, this means focusing on products that drive real usage and value, not just commoditized token generation, and being aware that the current spending frenzy may not sustain itself without commensurate revenue growth.
