Three years ago, Sequoia partner David Cahn first flagged the massive gap between AI infrastructure spending and actual revenue. Back then, Nvidia’s $50 billion in GPU revenue implied $200 billion needed to pay off the investment. Today, with hyperscalers tripling down, Cahn’s new estimate for 2026 infrastructure spend is $1.5 trillion, requiring $3 trillion in revenue to justify. That number’s likely low as memory costs and specialized chips rise. Meanwhile, revenue lags: Anthropic hits $60 billion and OpenAI $13-20 billion, leaving a chasm. Apollo’s chief economist Torsten Slok warns hyperscalers promise huge free-cash flow by 2028, but cheaper models and falling token prices threaten returns. If they miss, it’s not just a sector problem—it could trigger a recession and market correction. Why it matters: The AI boom’s sustainability hinges on whether demand for token-thirsty apps can outpace the cost efficiencies that make them cheaper, a tension no one’s solving yet.