Future reasoning will consume 70% of computing power, leaving 30% for training | Silicon Valley investor Zhang Lu @AIGC2026
Fusion Fund's Lucy Zhang predicts a significant shift in AI infrastructure, with inference computing demands set to surpass training by a 70/30 split. She highlights that communication within data centers consumes vastly more energy than computation itself, suggesting a critical need for advancements in optical communication. Zhang also emphasizes that the primary bottleneck for physical AI is the lack of high-quality, real-world data, rather than model size or compute power, pointing to sectors like healthcare as rich sources for this data. AI
IMPACT Shifts focus to inference and data quality, potentially altering infrastructure investment and R&D priorities.