NVIDIA's NVHBM technology threatens Broadcom's high-bandwidth memory market share

NVIDIA is moving deeper into the custom AI chip market by offering companies a modular approach through its NVHBM (NVIDIA High-Bandwidth Memory) technology, which consolidates memory control and proprietary connectivity into a base die design. Rather than designing accelerator chips themselves, NVIDIA can now provide the memory subsystem, networking fabric, rack architecture and software while letting partners focus only on the accelerator logic. Amazon's Annapurna Labs is the first disclosed customer planning to use NVHBM with its Trainium4 AI chip. The strategy puts NVIDIA in direct competition with Broadcom, which currently dominates the high-bandwidth memory interconnect market and lacks a comparable base-die offering.

1. NVHBM moves the memory controller from the accelerator chip into the base die of a high-bandwidth memory stack
2. NVHBM replaces the standard JEDEC memory bus with NVIDIA's proprietary die-to-die connection
3. Amazon's Annapurna Labs is the first disclosed customer planning to use NVHBM with its Trainium4 AI chip
4. MediaTek also adopted NVIDIA's NVLink Fusion platform
5. Counterpoint Research analyst Neil Shah questioned some of NVIDIA's performance claims, noting HBM4E is not yet shipping
6. Broadcom lacks a comparable base-die offering comparable to NVHBM, according to analyst Shah
7. The approach gives NVIDIA control over the memory controller, interface and link protocol