On August 27, 2026, NVIDIA made an official announcement regarding the launch of its NVHBM custom high-bandwidth memory technology. This innovative technology integrates the memory controller directly into the base die of the 3D HBM stack. Such an integration leads to a remarkable 30% boost in single-stack bandwidth and a 15% decrease in power consumption when compared to the standard HBM4e. Moreover, it liberates up to 30% of the main chip package area. As a pivotal component within the NVLink Fusion ecosystem, NVHBM is reserved exclusively for custom chip partners. Among the first batch of collaborators is Amazon's Annapurna Labs. Their next-generation Trainium 4 AI chip has already been equipped with reserved access interfaces for this technology. By optimizing the memory controller's position and customizing PHY interfaces, NVHBM significantly simplifies the complexity of multi-chip routing. This makes it an ideal solution for scenarios with high-bandwidth requirements, such as physical AI and agent AI. NVIDIA has plans to deeply integrate NVHBM with GPU architectures like Blackwell Ultra and Rubin. Additionally, it intends to make this technology accessible to third-party XPU customers through the NVLink Fusion technology, thereby creating a comprehensive full-stack AI computing power solution that encompasses GPUs, CPUs, and networking components.