The artificial intelligence boom is no longer just putting pressure on processors and data centers, but has begun to extend to memory, which is the least conspicuous component and the one most closely tied to users' devices.
With the rapid expansion in training and operating artificial intelligence models, the demand for high-bandwidth memory, server memory, and solid-state drives dedicated to data centers has increased.
Meanwhile, phones and personal computers still needed dynamic random access memory and NAND flash memory, but in a market where manufacturing capacity had become more crowded and more expensive.
This does not mean that data centers buy the memory of the phones themselves, but rather that the two parties compete indirectly over manufacturing capacity, materials and advanced capabilities.
As more of this energy is redirected to higher-priority AI products, the prices of some traditional types of memory have increased, and hardware companies have begun passing on some of the increase to the consumer or reducing specifications to maintain prices.
Artificial intelligence opens a new front in the memory market
Modern artificial intelligence systems need to transfer large amounts of data between processing units and memory at high speeds, which has made high-bandwidth memory a key component in AI accelerators.
High-bandwidth memory differs from the low-power RAM or dual RAM found in computers and phones, but in essence it is based on dynamic random access memory.
But this high performance comes at a manufacturing cost, as memory manufacturers use part of the DRAM manufacturing base to produce high-bandwidth memory, which requires more chip power per bit.
Producing each bit of high-bandwidth memory requires approximately 3 times more chip-level manufacturing power than manufacturing 5G double-band synchronous dynamic random-access memory, which makes the decision to allocate capacity directly impact the supply of other products.
High-bandwidth memory also occupies an increasing share of the DRAM manufacturing capacity of the largest suppliers.
TrendForce estimates that high-bandwidth memory could require about 22% of the total inputs of dynamic random access memory (DRAM) chips at the Big Three companies by the end of 2026, compared to 18% at the end of 2025, with the percentage rising to 30% in 2027.
Directing a portion of the power used to manufacture dynamic random-access memory (DRAM) to high-bandwidth memory reduces the amount of conventional memory that can be produced with the same power, which is why the memory industry talks about the offset effect.
As the demand for high-bandwidth memory expands, it becomes more difficult to increase the production of all types of dynamic random-access memory at the same rate.
Data centers are under pressure from two sides.
The memory story doesn't stop at dynamic random access memory (DRAM), as data centers also need increasing amounts of fast storage to power artificial intelligence applications, which supports the demand for enterprise-grade solid-state drives (SSDs).
Samsung itself indicates that it expects continued demand for AI-related NAND flash memory solutions, particularly high-performance storage solutions used in inference processes, in parallel with the growth of high-bandwidth memory.
Tags:
technology