OpenAI’s Compute Chief: We Can’t Build Fast Enough | Sachin Katti

The current AI infrastructure build-out represents one of the largest industrial undertakings in history, functioning as a massive factory system that converts electricity into intelligence. As demand for compute consistently outstrips supply, OpenAI’s "Stargate" strategy utilizes a portfolio approach, partnering with hyperscalers while simultaneously developing custom silicon like the Jalapeño chip to maximize tokens per watt. Because inference now consumes a significant portion of compute resources, efficiency is critical. The primary challenges to this expansion are not financial, but physical: securing sufficient power grid capacity, implementing advanced liquid cooling, and overcoming shortages of specialized trade labor. AI is increasingly integrated into the design process itself, enabling faster iteration cycles for both hardware and networking protocols like MRC, which ensure reliability across massive, 100,000-GPU clusters.