Korean outlet The Bell reported on November 13, 2026 that Samsung Electronics is developing custom memory based on seventh-generation high-bandwidth memory (HBM4E) for Meta's in-house AI chips. Citing industry sources, the report says Samsung plans to use Samsung Foundry's 2nm process for the base die. Samsung is said to be taking an approach of choosing the base die process according to each customer's requirements, and whether it can expand its own 2nm business depends on how many customers, like Meta, it can secure for that process. This picture is also consistent with NVIDIA's official announcements, Samsung's earnings materials, and reports about production plans for AMD.

AD

Meta's custom HBM4E to use Samsung's own 2nm process, differing from official announcements

HBM is memory in which multiple DRAM chips (core dies) are stacked, with a base die placed underneath. The base die contains the circuitry that connects the core dies to the processor. According to The Bell, in earlier HBM3 and HBM3E the base die was also made using a DRAM manufacturing process, which limited how much complex logic circuitry could be integrated. Starting with HBM4, Samsung adopted its own 4nm logic process for the base die so that more advanced functions could be built in.

According to the report, Samsung plans to offer standard and custom products separately from HBM4E onward. The custom products are being developed for NVIDIA, AMD and Meta, and the base die process will reportedly be split between 2nm and 4nm depending on each customer's requirements. Early custom base dies are expected to incorporate an HBM controller and die-to-die (D2D) interface circuitry, which could reduce power consumption during data transfer.

Samsung's official statements, on the other hand, are not as specific as this report. In its third-quarter 2026 earnings announced on October 29, the company said Foundry revenue and profit improved significantly. It cited increased sales of HBM base dies, growth in sales from advanced processes, and continued orders from HPC customers, including major hyperscalers, for its 2nm process. It also said a new fab in Taylor, Texas, would begin operating in 2027, with 4nm and 2nm driving growth. However, the increase in HBM base die sales and the 2nm customer wins were explained separately, so orders for HBM base dies using 2nm have not been confirmed.

Seoul Economic Daily reports that, for Samsung's standard products, both HBM4 and HBM4E, which is currently in mass-production qualification, will have base dies made on Samsung's own 4nm process. The specific plan to use 2nm for Meta's custom HBM4E is, for now, based on The Bell's report. According to that outlet, a Samsung Foundry official said that production of next-generation HBM base dies and other products is gradually filling the capacity of its 2nm lines.

NVIDIA reportedly uses TSMC, while Meta and AMD use Samsung 2nm

What stands out in Samsung's custom HBM business is that, according to reports, not only the base die process but also the foundry that makes it differs by customer. For Meta, Samsung Foundry's 2nm is reportedly used; for NVIDIA's NVHBM, TSMC's 3nm; and for AMD, 2nm at the Taylor plant.

Customer / product Base die maker / process Plan / reporting status
Meta custom HBM4E Samsung Foundry 2nm Development plan reported by The Bell
NVIDIA NVHBM TSMC 3nm Seoul Economic Daily reports NVIDIA specified it and Samsung accepted
AMD custom HBM Samsung Taylor Fab 2, 2nm Maeil Business Newspaper reports orders equal to the full planned capacity, with AMD accounting for most. Mass production planned for 2030

Note: The projects are at different HBM generations and stages of development and production, so this is not a like-for-like comparison.

This split in manufacturing processes is tied to changes in HBM design itself. In an official blog post on August 26, NVIDIA announced "NVHBM," which integrates the memory controller, previously placed on the GPU side, into the HBM base die. According to the company, compared with standard HBM4E this approach can increase memory bandwidth by up to 30%, cut HBM power consumption by 15%, and reduce processor die area by up to 25%.

Moving the memory controller into the base die means the base die needs more complex logic circuitry than before. NVIDIA has also said it will build a framework in which multiple memory makers can supply NVHBM to a common specification, reducing the qualification work required for each maker. In other words, the industry is shifting from a model in which memory makers design and manufacture their own base dies to one in which customers specify the design and manufacturing process.

According to The Bell, some customers, including NVIDIA, have asked for base dies to be made at TSMC, and Samsung's engineers have begun work to support TSMC's process. Micron and SK hynix are also reported to use TSMC for base dies, but their positions differ somewhat. Micron reportedly uses an external foundry even for standard HBM4E, and SK hynix is said to use TSMC's 12nm process for standard HBM4. Samsung, by contrast, is reported to be choosing between its own and external processes for each customer.

Samsung's in-house process does have advantages, however. The Bell notes that outsourcing manufacturing to an external foundry can slow the reflection of design changes and the adjustment of production schedules. If Samsung can make base dies on its own 2nm process, it becomes easier to coordinate everything from design to manufacturing internally. And if more customers such as Meta and AMD adopt its 2nm process, that would also help raise utilization at Samsung Foundry's advanced production lines.

AD

Yield and customer process choices are the challenges for expanding in-house 2nm

That said, customers will not necessarily choose Samsung's 2nm process. In advanced chip manufacturing, yield, production capacity and supply stability are important factors in the decision. Seoul Economic Daily cited industry estimates that TSMC's 3nm process has a yield above 90%, while Samsung Foundry's 2nm process, SF2, is only around 60%. These are estimates for processes of different generations, not definitive results published by either company.

Still, the fact that NVIDIA reportedly specified TSMC's 3nm process for the NVHBM base die and Samsung accepted it shows how much influence customers have over the choice of manufacturing process. According to Seoul Economic Daily, Samsung is also qualifying the NVHBM base die on its own 2nm process, but this is positioned as a fallback in case supply of TSMC's 3nm process becomes tight.

Meanwhile, the standardization NVIDIA is pursuing is, for now, an effort tied to a specific product specification, NVHBM. It does not mean Samsung has switched to using external foundries for all custom HBM. Seoul Economic Daily reports that Samsung will continue to offer customers such as AMD an end-to-end approach, handling everything from base die design to manufacturing.

One piece of evidence for this approach is the production plan for AMD. According to Maeil Business Newspaper, Samsung has secured orders equal to the full planned capacity of its 2nm base die production at Taylor Fab 2, with AMD accounting for most of it. However, mass production is slated to begin in 2030, so actual supply is still some way off. If the plan proceeds as intended, AMD, along with Meta, would become a major customer of Samsung's 2nm process.

What to watch next is these customers' choices of manufacturing process. If Meta or AMD switch from Samsung's own 2nm process to a process specified at an external foundry such as TSMC, the HBM base die orders Samsung can secure on its own 2nm would shrink. And if an NVHBM-style common specification spreads to other customers and TSMC manufacturing is specified, the room for Samsung's in-house process could narrow further.

Conversely, if Samsung discloses shipments of HBM base dies using the 2nm process in its earnings materials or official announcements, that would provide concrete support for this report. The company has said the new Taylor fab will begin operating in 2027, but has not explicitly stated how that fab relates to 2nm HBM base die production. Once it becomes clear which fab will produce base dies for which customers, it will be easier to judge how far Samsung's in-house 2nm process will be adopted in the custom HBM market.