Memory is crucial for AI CPUs? Korean media: DRAM shortage will continue for another year

```

AI inference architecture is reshaping the computing power landscape, and the surge in CPU memory demand is expected to drive the DRAM supply gap to persist until 2027.

On May 2, according to Korean media Sedaily, as the focus of the AI industry shifts from training to inference, CPUs are becoming major consumers of memory, further intensifying the global DRAM supply shortage. Korea's "Securities Daily" cited industry sources saying that current DRAM supply is about 10 percentage points lower than demand, and the supply-demand imbalance is difficult to ease in the short term.

The report says Intel's latest AI CPU is expected to be equipped with up to 400GB of general DRAM, four times the standard CPU. Meanwhile, NVIDIA, AMD, Google and other manufacturers’ next-generation AI chips are continuously pushing up demand for high bandwidth memory (HBM).

With multiple demand drivers, the industry predicts that Samsung Electronics and SK Hynix's memory supply capacity will struggle to keep pace with the growth in demand, and the memory supercycle previously expected to end in 2026 is likely to extend to 2027.

At the same time, DDR5 spot prices have already reflected market sentiment. According to Korea Mirae Asset Securities data, in April this year, DDR5 (16GB standard) spot prices for AI CPUs rose 2.8% in a single month, while old-generation DDR4 fell 16% during the same period; the trends are obviously diverging, and DDR5 price premiums are expanding.

CPU becomes the AI inference "scheduling core," memory demand spikes

According to reports, the AI industry's shift to the inference stage is the fundamental driver behind this surge in CPU memory demand.

In the past, AI data centers built their computing infrastructure around GPUs, using HBM to support large-scale parallel training tasks, with server configurations typically as one CPU paired with eight GPUs.

However, as AI inference demand grows rapidly, especially with the rise of "Agentic AI," CPUs are undergoing a fundamental role transformation—from peripheral helper to orchestrator, coordinating various AI agent workflows.

The report states that the core of this role transformation lies in "context memory" capability. CPUs need to continuously track and invoke the output results of various agents, requiring far greater memory capacity than before.

Intel executives said at a recent earnings conference, "In AI inference infrastructure, the ratio of CPU to GPU architecture has shifted from 1:4, and is further narrowing towards 1:1."

This means that the proportion of CPUs in AI servers has significantly increased, and the consumption of general DRAM has multiplied.

Memory battle between GPU and CPU, demand snowball effect emerges

Reports indicate that the memory capacity competition is extending from GPUs to CPUs, showing a "snowballing" pattern of demand expansion.

On the GPU side, NVIDIA’s next-generation AI chip "Vera Rubin" uses eight HBM chips to configure 288GB of memory, while AMD's next-generation GPU MI400 reaches a massive 432GB. Google's newly released eighth-generation tensor processing unit TPU 8i is also expected to be equipped with 288GB HBM.

On the CPU side, according to industry sources, CPU manufacturers are actively pushing for AI CPUs equipped with 300 to 400GB of DRAM. Both Intel's Xeon series and AMD's Epyc series have begun adopting high-capacity DDR5. In comparison, standard CPU products have DRAM configurations of only 96 to 256GB; memory demand for the new AI CPUs has increased up to fourfold.

The simultaneous memory demand expansion by GPUs and CPUs is further intensifying the supply pressure on the already tight DRAM market.

DDR5 supply-demand imbalance intensifies, supercycle may extend to 2027

General DRAM prices have doubled, delivering unprecedented performance for the memory industry, while the supply-demand imbalance remains difficult to reverse in the short term. According to reports, an industry insider said:

"Currently, DRAM market supply is about 10 percentage points lower than demand. As demand for general DRAM outside HBM surges, the supercycle is highly likely to extend from previously projected 2026 to 2027."

DDR5 spot market price trends confirm this judgment. According to Mirae Asset Securities data, in April this year DDR5 (16GB standard) spot prices rose 2.8% in a single month, while DDR4 plunged 16% during the same period. This price divergence directly reflects that AI CPU-driven demand for DDR5 is reshaping the memory market’s supply-demand structure.

The report points out that for Samsung Electronics and SK Hynix, with HBM production already highly strained, a further surge in general DRAM demand will further test their overall supply deployment capabilities. The market generally expects the memory industry boom cycle will be longer than previously anticipated.

Risk Alert and DisclaimerThe market has risks, investment should be cautious. This article does not constitute personal investment advice, nor does it take into account individual users' specific investment goals, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Invest accordingly, at your own risk. ```