Samsung plans to outsource conventional memory module production to external partners while prioritising high-bandwidth memory for AI applications, signalling a strategic shift driven by industry capacity pressures and the rise of AI computing needs.
Samsung’s latest manufacturing shift is a telling sign of how the AI boom is reshaping memory rather than simply increasing demand for it. According to reports cited by DIGITIMES and echoed by specialist technology outlets, the company plans to move all additional production of conventional memory modules to external partners, including firms in India, Vietnam and the Philippines, while reserving more of its own constrained facilities for high-bandwidth memory, or HBM, and advanced packaging. The change is less about stepping away from DRAM than about deciding which parts of the business now merit the limited space Samsung can control most tightly.
The immediate constraint is physical capacity. Samsung’s back-end plants in South Korea are understood to be under pressure from the need to expand advanced packaging lines, particularly those used for HBM, which is increasingly central to AI accelerators. DDR5 module assembly is far easier to shift to contract manufacturers because it involves mounting packaged chips on circuit boards, whereas HBM depends on more specialised stacking and interconnect processes. By outsourcing more routine work, Samsung can keep its own sites focused on the manufacturing steps that are hardest to relocate.
That reallocation sits inside a wider investment push. Samsung disclosed in June plans to expand HBM production in Onyang and Cheonan as part of a much larger semiconductor programme running through 2040. The latest outsourcing move therefore looks less like a cost-cutting exercise and more like a structural adjustment to the company’s production footprint. Industry analysts say the pressure is being driven by AI, which is not only consuming more memory chips but also drawing in wafer capacity, packaging equipment and factory space that would otherwise support conventional markets.
The imbalance is particularly visible in HBM. TrendForce estimates that HBM will account for roughly 22% of total DRAM wafer input among the three largest suppliers by the end of 2026, rising to around 30% by the end of 2027, even though it is expected to represent only about 9% and 13% of DRAM bit supply in those years. That gap helps explain why overall memory supply can tighten even when chipmakers are increasing output. SK Hynix chief executive Kwak Noh-jung warned in July that the industry could face its worst shortage in 2027, while TrendForce expects DRAM supply to remain tight through that period as AI server demand continues to absorb capacity.
The strain is already extending beyond data centres. Reuters reported this week that smaller smartphone and laptop makers are preparing for memory scarcity through at least 2027, while Counterpoint expects smartphone shipments to fall 13.9% this year. Memory can account for as much as 60% of the component cost of some handsets priced at about $400, making the issue commercially significant even outside premium devices. Larger companies have more leverage, but they are not immune: Apple has already warned that rising memory costs are beginning to affect profitability, and both Samsung and SK Hynix have acknowledged the disruption caused by the shift towards AI-related products.
Samsung’s choice is notable because it is not simply chasing the most obviously profitable segment in the short term. TrendForce says rising prices for conventional DRAM have recently pushed per-wafer revenue for DDR5 64GB RDIMM above HBM, and that DDR5 profitability also briefly moved ahead of HBM because contract pricing for the latter has not adjusted as quickly. Even so, Samsung appears to be prioritising HBM capacity for future demand from AI systems, including next-generation accelerator platforms and custom chips being developed by cloud providers. In practical terms, the company is protecting the production steps most likely to matter in the next phase of the market, even if conventional memory remains valuable today.
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.
