Micron has announced its collaboration with Nvidia to develop industry-first customised HBM4E memory, marking a strategic move towards integrated, performance-optimised AI accelerators and signalling a paradigm shift in memory design and industry competition.
High-bandwidth memory is moving beyond raw speed and capacity into a more tailored phase, as chipmakers seek to align memory more closely with the design of their AI accelerators. According to Newis, Micron has now joined Samsung Electronics and SK Hynix in developing custom HBM for Nvidia, signalling that memory suppliers are no longer competing only on bandwidth, but also on how precisely they can fit into a customer’s chip architecture.
Micron chief executive Sanjay Mehrotra said on the company’s latest earnings call that it is working with Nvidia on NVHBM, which he described as an industry-first customised HBM4E implementation. The arrangement is intended for Nvidia’s next-generation graphics processors and its NVLink Fusion platform, which allows third-party AI chips and CPUs to use Nvidia’s interconnect technology in larger systems. Tom’s Hardware reports that Nvidia has said the design can deliver up to 30% more bandwidth and reduce power consumption by 15% compared with standard HBM4E, while also freeing more room on the accelerator for compute logic.
The technical change lies in the base die at the bottom of the HBM stack. Instead of keeping all memory-control functions on the accelerator itself, part of that logic is moved into the memory package. That reduces the silicon area taken up by controller circuitry on the AI chip and can improve the efficiency of data movement between memory and compute blocks. Industry reporting suggests that this kind of customisation is becoming more important as AI systems are designed from the outset around specific memory and interconnect requirements rather than generic component choices.
The broader industry context is clear. Micron’s HBM4E is being positioned as a more adaptable memory generation, with a standard 2048-bit interface and mass production expected in 2026, according to Tom’s Hardware. Reports also say the company has worked with TSMC on the complex base logic dies needed for customised versions. For Nvidia, the move strengthens its NVLink Fusion ecosystem, where partners can build more tightly integrated AI hardware. For memory makers, it marks a shift from selling largely standard products to co-designing components that influence performance, efficiency and usable compute area across the whole accelerator.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
