HBM Now 63% of AI Chip Component Costs as Memory Spend Doubles

Epoch AI’s component-level breakdown of AI accelerators from Nvidia, AMD, Google, and Amazon shows high-bandwidth memory eating an ever-larger share of the bill of materials. Between Q1 2024 and Q4 2025, HBM rose from 52% to 63% of per-chip component spending, while advanced CoWoS packaging fell from 19% to 15% and auxiliary parts dropped from 15% to 9%. Logic dies held steady at roughly 13-14% despite being the part most associated with cutting-edge fabrication.

The absolute numbers explain the shift. Total component spending on AI chips jumped from about $22 billion in 2024 to $52 billion in 2025, and HBM alone accounted for roughly $20 billion of that $30 billion increase. Memory capacity and bandwidth, not compute, is where the marginal dollar in AI hardware is going.

The concentration has supply-chain consequences: HBM production is dominated by a handful of vendors, and packaging capacity at TSMC remains a separate bottleneck. As memory’s cost share climbs past two-thirds, accelerator economics and availability become increasingly tied to the HBM roadmaps of SK Hynix, Samsung, and Micron rather than to GPU design alone.