HBM Memory Emerges as Key Driver of AI Chip Growth

As artificial intelligence shifts from model training to inference, high-performance, high-bandwidth memory (HBM) is reshaping the semiconductor industry. Data-transfer bottlenecks leave GPUs idle for as much as 90% of the time in AI computing, with actual computation accounting for just 10%. High-bandwidth, high-capacity memory has therefore become crucial to breaking computing bottlenecks and determining whether chip giants such as NVIDIA can sustain growth.

Kim Jung-ho, a professor at the Korea Advanced Institute of Science and Technology (KAIST) known as the “father of HBM,” said in early July 2026 that future chip architectures would be memory-centric. He predicted that HBM and high-bandwidth flash memory (HBF) would be vertically integrated into 100-layer 3D semiconductor towers to overcome physical limits. Institutional analysts estimate that global HBM output value will approach $60 billion in 2026, allowing memory manufacturers to regain pricing power across the supply chain.