$MU

Summary

DRAM shortage driven by AI CPU demand will extend through 2027, benefiting memory suppliers as DDR5 prices maintain premium over legacy DDR4.

Reasoning

AI CPUs requiring 300-400GB DRAM (4x typical), coupled with GPU HBM demand, creates sustained memory shortage. Samsung and SK hynix supply constrained. DDR5 for AI CPUs showing price resilience vs collapsing DDR4.

AI CPUs Are Devouring DRAM — Memory 'Shortage' to Last Another Year The memory industry has posted record-breaking results driven by commodity DRAM prices that have surged more than 100%, and with the proliferation of AI-purposed central processing units (CPUs) now coming into play, forecasts suggest the shortage will extend for another year. Intel's recently unveiled "AI CPU" is expected to carry up to four times more commodity DRAM than previous generations. Combined with surging demand from graphics processing units (GPUs) that require high-capacity DRAM, observers expect the memory supply capacity of Samsung Electronics and SK hynix to fall short of demand. According to industry sources on the 2nd, CPU manufacturers are pursuing the integration of 300–400GB of DRAM into AI CPUs. This is up to four times the overwhelming scale compared to typical CPU products (96–256GB). CPUs Emerge as the 'AI Orchestrator' The surge in high-capacity DRAM demand for AI CPUs is tied to the AI industry's pivot toward an inference-centric structure. Whereas AI inference was once limited to simple Q&A, it now serves as the "orchestrator" coordinating various agentic AIs. The key in this process is "context memory." For a CPU to coordinate the entire workflow by referencing the outputs of each agentic AI, it must remember the content. This makes scaling up memory — the space for retention — essential. Until now, AI data centers have built computing infrastructure centered on GPUs equipped with High Bandwidth Memory (HBM). Leveraging the GPU's strength in training AI on vast amounts of data simultaneously, the focus has been on "AI training." Server configurations accordingly followed an 8-GPU-to-1-CPU pattern. However, as the industry's center of gravity shifts toward inference, server configurations with substantially expanded CPU ratios are spreading. In a recent earnings call, Intel executives explained: "In AI inference infrastructure, the computing structure has shifted to a CPU-to-GPU ratio of 1 to 4, and the trend is moving further toward 1 to 1." After GPUs, CPUs Join the Memory Battle — Demand Snowballs The competition for memory capacity has expanded from GPUs to CPUs, snowballing in scale. NVIDIA's next-generation AI chip "Vera Rubin" carries 288GB through 8 HBM stacks, while AMD's next-generation GPU MI400 boasts an even larger 432GB mammoth-class capacity. Google's recently unveiled custom chip — the 8th-generation Tensor Processing Unit (TPU 8i) — is also slated to feature 288GB of HBM. Add to this Intel's AI CPU "Xeon" and AMD's "EPYC" beginning to use up to 400GB of high-capacity DDR5, and the memory shortage is expected to persist longer. The market temperature is already being proven in spot prices. According to Kiwoom Securities, while the price of legacy DDR4 (16GB basis) plunged 16% in a single month in April, the spot price of DDR5 (16GB basis) — used in AI CPUs — rose 2.8% over the same period, maintaining its price premium. An industry source said: "The current DRAM market is understood to be roughly 10 percentage points short of demand. With commodity DRAM demand surging on top of HBM, the supercycle is highly likely to extend from the previously expected 2026 into 2027." $MU $DRAM

Return by horizon