Starting in late 2025, DDR5 and SSD The price has been skyrocketing, and I still have no intention of turning back. I originally thought that the situation would begin to ease at the end of 2026, but I didn’t expect that it might be too early. There will be a second wave of pressure coming soon:CPU。
according toKorean media SE Daily Quoting industry sources, it is pointed out that due to the agency-based AI It is becoming more and more popular, and CPU manufacturers are planning to equip the next generation AI CPU with 300~400GB of memory, which means 4 times the current 96~256GB. In other words, in the past, it was mainly GPUs that competed for memory. In the future, even CPUs will join in. Players who want to buy memory at full price may have to wait.
AI agents push CPU into a memory arms race: a single fortress has 400GB, and DRAM will be in short supply until 2027
The AI wave in the past two years has mainly focused on large-scale language model training and reasoning, so it all consumes the computing power of GPUs. However, as the models are strong enough, it has now gradually entered the era of agent AI. It no longer throws a question and gives an answer, but makes the model act like a real agent. It will automatically plan steps, call external tools, look up information, read files, change the program code, and then go back and correct itself.
This entire process relies more on the CPU for coordination, not the GPU. Every tool call, every API request, every intermediate state temporary storage and decision-making logic must be run by the CPU. The GPU only contributes when “model reasoning is really needed.”
Intel has said that the ratio of GPU to CPU in the AI data center has converged from the original 8:1 to 4:1, and may even approach 1:1 in the future. Arm estimates that traditional AI data centers require approximately 30 million CPU cores per GW, but in the era of agent AI, demand will jump to 120 million cores, a full 4 times.
The CPU is responsible not only for operations, but also for “context memory”.
When tasks are continuously executed, conversation records, tool return results, intermediate decisions, and states that may need to be retried are all temporarily stored. These must be placed in the main memory of the CPU, so the capacity requirements will naturally increase. Originally, a server CPU with 96 ~ 256GB DRAM was enough, but in the era of agent-based AI, it will obviously not be enough.
For this reason, “SE Daily” recently pointed out that CPU manufacturers are currently planning to equip their AI CPUs with 300 to 400 GB of memory.
If it does go to 400GB, that’s a bit of an exaggeration, more than the 288GB HBM3E of the NVIDIA GB300, Vera Rubin, and AMD MI350X.
However, SE Daily did not mention which form of memory 400GB refers to.Foreign media Wccftech Presumably, one possibility refers to high-density memory directly integrated into the CPU package, that is, HBM or new standards such as HBF and ZAM that are under development are directly placed next to the CPU, similar to the design of HBM on the GPU.
The other is simple violence, with a single DIMM memory capacity of 400GB. The current server CPU platform can reach 4~8TB by fully plugging in standard DIMMs. In the future, if a single DIMM has 400GB, the total memory capacity will be quite terrifying.
Source: KOCPC Chinese