HBM does have elements of customization and co-design with companies like Nvidia, but the level is not very strong. The real customization is limited mainly to the packaging and the base die. The many layers of $Roundhill Memory ETF(DRAM.US) DRAM dies stacked on top remain fully standardized according to JEDEC specifications, meaning they are essentially the same for all customers.

For example, when Samsung’s HBM3E failed to pass Nvidia’s qualification and its share dropped dramatically from around 60% to 20%, the company did not scrap the production. Instead, it redirected that capacity to supply Google’s TPU and AMD. Physically, the HBM3E chips provided to Nvidia are the same as those supplied to AMD and Google, which shows that capacity can still be partially transferred between customers relatively easily.

Looking ahead to HBM4 and future versions, customization is expected to increase. This includes integrating custom logic or cache directly on the base die. In more advanced designs, the HBM4E memory controller and custom die-to-die interfaces can even be built into a logic base die. According to SemiAnalysis, companies such as OpenAI, Nvidia, and AMD are working on their own custom HBM solutions, but these efforts primarily focus on the base die while the upper DRAM layers stay standardized.

Because of the need for close cooperation on packaging, customers are often required to sign long-term agreements. However, since the core DRAM components are standard and production capacity remains somewhat flexible and transferable, the overall customization barrier is only moderate. As a result, HBM is considered only partially customized and counts as roughly half a point in this context.

@Bridge Buzz SG