A new analysis from Epoch reveals that HBM memory now accounts for nearly two-thirds of AI chip component costs, marking a significant shift in semiconductor economics. Their breakdown of chips from Nvidia, AMD, Google, and Amazon shows memory's share climbing from 52% to 63% between Q1 2024 and Q4 2025, fundamentally altering how manufacturers allocate budgets for AI accelerators.

The research tracked four component categories: memory (HBM), logic dies, advanced packaging (CoWoS), and auxiliary components. While memory costs rose sharply, advanced packaging fell from 19% to 15% and auxiliary components dropped from 15% to 9%. Logic die costs remained steady at 13-14% throughout the period, suggesting this layer of the stack has reached relative maturity in cost structure.

Total AI chip component spending nearly doubled from $22 billion in 2024 to $52 billion in 2025. HBM alone contributed roughly $20 billion of that increase, reflecting the massive bandwidth demands of modern AI workloads. This cost explosion stems from increasingly dense chip designs requiring more high-bandwidth memory per device.

The data signals a critical bottleneck forming in the AI supply chain. As memory costs dominate budgets, companies face pressure to optimize HBM utilization or develop alternative architectures. This shift may accelerate research into compute-in-memory technologies and influence future chip design priorities across the industry.