SK Hynix will start making 12-High HBM4 memory modules in October 2025 for computer companies. The Korean manufacturer timed this production to coincide with NVIDIA's launch of its Rubin graphics processor in early 2026. Each HBM4 stack contains twelve layers of 24 gigabit memory chips that create 36 gigabytes per package. These modules deliver data transfer speeds of up to 2 terabytes per second for artificial intelligence computers. Factory workers improved manufacturing success rates from 60 percent to over 70 percent after fixing bonding problems.
TrendForce research company expects HBM4 memory to cost 30 percent more than current HBM3E modules at first. Cloud computing companies and AI accelerator makers will pay higher prices to access faster bandwidth and better efficiency. The new memory technology is expected to become the dominant choice by the second half of 2026. Growing artificial intelligence workloads need faster data movement for training and running complex computer models. HBM4 provides the capacity and speed that these demanding applications require.
NVIDIA designed its Rubin graphics processor to leverage HBM4's higher data rates for larger AI models. SK Hynix invested money in advanced tools that detect bonding defects at a microscopic scale during manufacturing. The company monitors temperature changes during module assembly to prevent warping and ensure electrical connections work properly. These improvements help create memory modules that meet performance standards under heavy loads of AI computing. System builders can plan their designs around known memory release dates to reduce uncertainty.
TrendForce research company expects HBM4 memory to cost 30 percent more than current HBM3E modules at first. Cloud computing companies and AI accelerator makers will pay higher prices to access faster bandwidth and better efficiency. The new memory technology is expected to become the dominant choice by the second half of 2026. Growing artificial intelligence workloads need faster data movement for training and running complex computer models. HBM4 provides the capacity and speed that these demanding applications require.
NVIDIA designed its Rubin graphics processor to leverage HBM4's higher data rates for larger AI models. SK Hynix invested money in advanced tools that detect bonding defects at a microscopic scale during manufacturing. The company monitors temperature changes during module assembly to prevent warping and ensure electrical connections work properly. These improvements help create memory modules that meet performance standards under heavy loads of AI computing. System builders can plan their designs around known memory release dates to reduce uncertainty.
