The restructuring of AI inference architecture is reshaping the computing landscape, with surging CPU memory demand expected to extend the DRAM supply gap until 2027. According to a report from Korean media outlet Sedaily on May 2nd, as the AI industry's focus shifts from training to inference, CPUs are emerging as major consumers of memory, intensifying the global DRAM supply crunch. The Korea Securities Daily cited industry sources indicating that current DRAM supply falls approximately 10% short of demand, with the imbalance unlikely to ease in the near term.
The report highlighted that Intel's newly launched AI CPUs are projected to incorporate up to 400GB of general-purpose DRAM, quadruple the capacity of standard CPU products. Concurrently, next-generation AI chips from NVIDIA, AMD, and Google continue to drive demand for high-bandwidth memory (HBM).
With multiple demand streams converging, industry forecasts suggest that Samsung Electronics and SK Hynix will struggle to keep pace with accelerating demand. The memory super-cycle, previously anticipated to conclude in 2026, is now expected to extend into 2027.
Market conditions are already reflected in DDR5 spot prices. Data from Mirae Asset Securities shows that in April, spot prices for DDR5 (16GB standard) used in AI CPUs rose 2.8% month-over-month, while older-generation DDR4 prices fell 16% during the same period. This divergence underscores expanding price premiums for DDR5.
The fundamental driver behind the surge in CPU memory requirements is the AI industry's transition to the inference phase. Previously, AI data centers were built around GPUs as the core computing infrastructure, utilizing HBM to support massive parallel training tasks, with typical server configurations pairing one CPU with eight GPUs.
However, rapid expansion in AI inference demand, particularly with the rise of Agentic AI, is fundamentally transforming the CPU's role—from a peripheral support function to an orchestrator coordinating workflows across various AI agents. This shift centers on "contextual memory" capabilities, requiring CPUs to continuously track and invoke outputs from multiple agents, necessitating significantly larger memory capacities than before.
An Intel executive recently stated during an earnings call that "in AI inference infrastructure, the computing architecture has shifted from a 1:4 CPU-to-GPU ratio and is further narrowing toward a 1:1 balance." This indicates a substantial increase in CPU representation within AI servers, amplifying consumption of general-purpose DRAM.
The memory capacity race is expanding from GPUs to CPUs, creating a snowball effect in demand. On the GPU side, NVIDIA's next-generation "Vera Rubin" AI chip configuration incorporates eight HBM chips for 288GB of memory, while AMD's upcoming MI400 GPU reaches 432GB. Google's latest eighth-generation Tensor Processing Unit, TPU 8i, is also expected to feature 288GB of HBM.
On the CPU side, industry sources reveal that manufacturers are actively promoting AI CPUs equipped with 300-400GB of DRAM. Both Intel's Xeon series and AMD's Epyc series have begun adopting high-capacity DDR5. By comparison, standard CPU products typically configure 96-256GB of DRAM, meaning the latest AI CPUs represent a fourfold increase in memory requirements.
The simultaneous expansion of memory demand from both GPUs and CPUs is exacerbating supply pressures in an already tight DRAM market. General-purpose DRAM prices have surged over 100%, delivering unprecedented performance for the memory sector, yet the supply-demand imbalance shows no signs of near-term reversal.
One industry insider commented: "Current DRAM supply trails demand by approximately 10%. With exploding demand for general-purpose DRAM beyond HBM, there is high probability the super-cycle will extend from the previously projected 2026 into 2027."
DDR5 spot market price trends validate this outlook. Mirae Asset Securities data confirms the 2.8% monthly increase for DDR5 (16GB standard) in April, contrasting sharply with DDR4's 16% decline. This price divergence directly reflects robust DDR5 demand from AI CPUs reshaping memory market dynamics.
For Samsung Electronics and SK Hynix, the additional surge in general-purpose DRAM demand presents further challenges to their overall supply allocation capabilities, particularly amid already strained HBM production capacity. Market consensus now anticipates the memory industry's growth cycle will prove more prolonged than earlier predictions.
Comments