Sachin Katti describes the AI compute buildout as one of the largest infrastructure projects in history, with OpenAI planning to spend about $50 billion on compute this year. He explains that AI data centers are fundamentally different from traditional ones, acting as factories that turn electrons into tokens, requiring innovations like liquid cooling and behind-the-meter power. Katti discusses Project Jalapeno, OpenAI's custom chip designed to maximize tokens per watt, and notes that inference now dominates compute usage. He argues that the bigger risk is not building fast enough, as demand far outstrips supply. The conversation covers site selection, the Stargate strategy, and the use of offtake agreements for financing. Katti also addresses bottlenecks in the supply chain, including a shortage of skilled tradespeople, and the potential for AI to eventually design its own chips.