Samsung might have the one thing its memory rivals can’t copy: a foundry sitting next to its HBM lines. That matters more now that AI customers want memory built to their own specs.
According to a report out of South Korea, via SamMobile, Samsung may be developing custom high-bandwidth memory for the AI chips Meta is building in-house. The part would be based on Samsung’s seventh-generation HBM4E. Meta-specific functions would be integrated into the base die, the logic layer at the bottom of an HBM stack. That base die may be manufactured on Samsung Foundry’s 2nm process.
This is a report, not a confirmed contract, so treat the details as provisional. But the direction fits what the industry is doing. Samsung already supplies HBM to NVIDIA and AMD. With HBM4E, it will reportedly offer custom versions too, where a buyer’s requirements go straight into the base die. Samsung is said to be moving fast on custom base dies for both 2nm and 4nm, depending on what customers ask for.
Why does the base die matter? In older HBM generations it was a fairly generic piece of silicon. Now it’s a place to put custom logic close to the memory, which can help with bandwidth, power and latency. For a company like Meta, designing its own accelerators to cut reliance on NVIDIA, that’s a useful lever. Google, Amazon and Microsoft are all chasing the same goal with their own chips, so custom memory demand is likely to grow.
This is where Samsung’s structure helps. SK Hynix, the current HBM leader, and Micron don’t run their own foundries. They have to lean on outside manufacturers for advanced logic dies. Samsung can do memory, logic and packaging under one roof, and it is one of very few companies able to make chips on advanced 2nm nodes. That could mean tighter coordination and a simpler pitch to customers.
Still, a structural advantage isn’t a win. SK Hynix has had the stronger track record with NVIDIA in HBM, and Samsung has spent the last couple of years catching up on qualification and yields. Its 2nm foundry work also has to prove it can deliver at volume. Customers will care about shipped parts, not slides from a keynote.
If the report holds, though, it’s a meaningful signal. Samsung would be positioning itself as more than a commodity memory vendor. For developers and founders watching AI infrastructure costs, more competition in custom HBM is good news. It could ease supply constraints and give hyperscalers more room to push back on pricing. I’d watch for confirmation from Meta or Samsung, and for early sample timelines.



