Nvidia is moving the memory controller into the HBM base die, with Amazon's Annapurna Labs first in line.

NVHBM, announced August 26, relocates NVIDIA's custom memory controller from the compute chip into the HBM base die, claiming up to 30% more bandwidth than standard HBM4E, 15% lower HBM power, and up to 25% more freed area on the XPU compute die (NVIDIA). Next-generation Trainium4 will support NVHBM alongside NVLink Fusion. Nvidia selling memory architecture into the custom silicon that competes with its GPUs is the strategic move, and it's the same week AWS and Nvidia announced 2 million additional GPUs deployed across AWS infrastructure over two years (GlobeNewswire), while Amazon AI chief Peter DeSantis has said AWS is in talks to sell Trainium externally as a direct Nvidia alternative.