ykevinzhang

This post was originally posted my Substack. I can be reached on LinkedIn and X. The US economy is all-in on AI, with data center capex responsible for 92% of US GDP growth in the first half of 2025. OpenAI alone is expected to spend $1.4T on infrastructure over the next few years, and caused a controversy on X when asked about its revenue-capex mismatch. So, can AI deliver the required ROI to justify the trillions in capex being put into the ground, or are we doomed to see a repeat of the Telecom bubble? The industry is operating in a fog of war. On one hand, AI labs like OpenAI and hyperscalers all cite compute shortages as a limiter on growth, and Nvidia continues to crush earnings. But on the other, actual AI revenues in 2025 are an order of magnitude smaller than capex. This post attempts to dissect this fog of war and covers the following: * A state of the union on current AI capex * How the quality and trajectory of AI application revenues impact cloud vendors * Why there is a $1.5T AI revenue shortfall relative to capex * How AI clouds are different from traditional data centers and the fundamental risks in the AI data center business model * How internal workloads from the Magnificent Seven ultimately decide whether we avoid an “AI winter” * A framework for investors to navigate the current market turbulence, across both public and private markets Let’s dive in. The dominant narrative is the mismatch between AI capex and AI application revenues To understand the current state of AI euphoria, it’s helpful to go all the way back to 2022, when the software industry was facing significant headwinds on the back of rising interest rates. The SaaS crash was short-lived, as the launch of ChatGPT in late 2022 rapidly reoriented investor interest around AI. The dominant narrative was that AI would replace headcount, not software budgets, unlocking trillions in net new revenues. There were also technical trends that drove demand for compute: the rise of