Professor Kati highlights Intel’s strategic advantages in AI infrastructure amid growing compute demands driven by evolving AI workloads, emphasizing the importance of diverse, scalable hardware and robust supply chains to support rapid AI growth. She underscores the critical need for investment in foundational infrastructure and innovation in AI-driven chip and system design to sustain the AI supercycle and enable transformative advancements.
The discussion opens with Professor Kati, who brings extensive experience from networking startups to leadership roles at Intel and OpenAI, highlighting Intel’s recent emergence as a significant AI company. She emphasizes two key factors benefiting Intel: the global supply constraints favoring companies with manufacturing capabilities, and the resurgence of CPUs driven by evolving AI workloads, particularly with agentic AI. Despite market skepticism, she expresses optimism about Intel’s future under strong leadership.
A central theme is OpenAI’s rapid compute capacity growth, which has tripled annually over the past three years, closely correlating with revenue increases. Professor Kati explains that compute demand spans both training and inference, with inference now constituting the majority due to expanding AI applications like Codex and agentic workloads. OpenAI aims to maximize compute availability to empower unconstrained research and product development, targeting an ambitious 30-gigawatt scale by decade’s end.
The conversation delves into the complexities of sourcing and managing AI compute infrastructure, which encompasses chips, memory, networking, power, cooling, and land. Scaling to gigawatt levels involves orchestrating vast supply chains and ensuring operational stability, as AI hardware is sensitive to environmental factors. Energy consumption is a significant concern, with large AI data centers potentially impacting regional power grids, prompting exploration of alternative energy sources and infrastructure innovations that could benefit broader society.
Professor Kati describes the evolution of AI workloads from simple inference tasks to sophisticated agentic workloads that involve reasoning, tool use, and iterative problem-solving. This shift demands heterogeneous compute infrastructure combining GPUs, CPUs, and specialized accelerators optimized for different tasks to deliver low-latency, high-performance user experiences. She highlights the importance of matching workload components to appropriate hardware to optimize efficiency and responsiveness, noting ongoing efforts to reduce latency across the entire compute stack.
Looking ahead, she underscores the critical role of foundational infrastructure layers—such as chip fabrication, memory, and power systems—in sustaining AI’s growth, advocating for renewed investment and talent development in these areas. While acknowledging the dominance of companies like Nvidia in the accelerator market, she stresses the need for a resilient, diversified supply chain. Finally, she cautions against overreliance on superficial model wrappers and emphasizes the transformative potential of AI infrastructure innovation, including AI-driven chip and system design, to accelerate progress in the AI supercycle.