Overview

NVIDIA has moved into a part of the AI chip business it doesn't normally touch: the memory that sits right next to the processor. Its new technology, called NVHBM, lets NVIDIA design a key piece of the high-bandwidth memory that companies use in AI chips, even when those chips aren't made by NVIDIA. This report examines NVIDIA's NVHBM, the significance of its arrival among existing players, and how much of the AI hardware stack it's starting to control.

What the Full Report Covers

The report walks through how NVHBM works, how its claims stack up, and how it compares to what Marvell and Broadcom are already doing in this space. It also looks at NVIDIA's new partnership with MediaTek and what it could mean for the competitive landscape, before placing NVHBM alongside other industry attempts to solve the same underlying memory bottleneck.

Key Questions the Report Answers

• What is NVHBM, and how does moving the memory controller into the base die change the HBM interface?
• What does the MediaTek NVLink Fusion partnership (and $3.5bn investment) mean for NVIDIA's distribution and dollar content per custom accelerator programme?
• How does NVHBM's timeline and risk profile compare to more radical memory-wall approaches (Intel XBM/ZAM, Qualcomm HBC, DRAM-on-logic)?
• What does NVIDIA acting as a de facto standards body for a proprietary HBM interface mean for memory vendor margins and second-sourcing?