Samsung Electronics is reportedly developing customised high-bandwidth memory for Meta’s home-grown AI processors, hoping to secure another lucrative customer as the memory industry’s AI feeding frenzy continues.
According to TrendForce. citing industry sources, Samsung is developing specialised HBM4E memory with functions designed specifically for Meta’s processors. The company reportedly plans to manufacture the base die using its own 2nm foundry process.
The base die is the logic component beneath the stacked memory chips and manages communication with the processor. Samsung apparently wants to integrate customer-specific memory controllers and die-to-die interfaces into this component, potentially reducing power consumption and improving performance.
Meta has been developing its own AI accelerators to reduce dependence on Nvidia’s expensive hardware, although designing your own processor does not eliminate the need for advanced memory. Samsung would be delighted to supply it, particularly if the arrangement gives its struggling foundry division more work.
The Korean chipmaker reportedly intends to offer standard and customised HBM4E products, using either 2nm or 4nm manufacturing processes depending on customer requirements. This would allow companies developing specialised AI processors to optimise memory interfaces for their own architectures.
Samsung has an advantage here because it manufactures both DRAM and advanced logic chips, something its principal memory rivals cannot reproduce internally. Its foundry operation could produce the customised base dies while its memory division supplies the stacked DRAM, keeping more of the manufacturing business within the company.
Unfortunately, Samsung’s customers do not necessarily share its enthusiasm for keeping everything in-house. Nvidia has reportedly asked Samsung, SK Hynix and Micron to use TSMC’s 3nm process for certain customised memory base dies, hoping to standardise designs and reduce validation problems between suppliers.
That creates the entertaining possibility of Samsung winning lucrative memory contracts, only to discover that its customers want a Taiwanese competitor to manufacture some of the most important components. Samsung’s foundry executives will presumably be delighted with that arrangement.
Meanwhile, SK Hynix is developing its own customised HBM products and has confirmed that it is working closely with major customers. The company completed HBM4E sample shipments during the second quarter and has been progressing through customer evaluations.
Samsung is reportedly preparing to expand production of advanced HBM base dies at its semiconductor facilities in Texas, with AMD expected to take much of the planned output. The company’s proposed second Taylor fabrication plant could eventually manufacture these components using its 2nm process.
The development reflects a change in the high-bandwidth memory business, where customers increasingly want memory designed around specific processors and workloads. As AI accelerators become more specialised, manufacturers are being pushed to add more sophisticated logic to memory packages.
For Samsung, the attraction extends beyond selling more DRAM. Customised HBM could provide additional demand for its advanced foundries, which have struggled to challenge TSMC’s dominance despite years of investment and some spectacularly ambitious promises.
Industry deep throats say Samsung is developing Meta-specific HBM4E base dies using its 2nm technology, while other reports suggest the company is preparing additional US manufacturing capacity for advanced memory components.







