AI memory demand is getting worse, and the pressure may soon move beyond GPUs. A new report says future AI focused CPUs could be equipped with 300GB to 400GB of memory, a major jump from today’s common 96GB to 256GB range per chip.
The reason is agentic AI. These systems do more than answer one prompt. They plan, call tools, manage longer context, and work through multi step tasks. That kind of workload needs more memory close to the processor, not only more raw compute.
CPUs are becoming more important in AI datacenters
GPUs still matter for AI, but CPUs are taking on a larger role. The report says the GPU to CPU ratio in datacenters has moved from around 8 to 1 toward 4 to 1, and could eventually approach 1 to 1 as agentic AI workloads grow.
That shift means CPUs may need much larger memory pools. More memory helps with context handling, data movement, and keeping AI workloads running efficiently. Even with compression techniques, memory demand is still rising.
| AI hardware trend | What is changing |
|---|---|
| CPU memory capacity | Could rise to 300GB to 400GB per AI CPU |
| GPU to CPU ratio | Moving from 8 to 1 toward 4 to 1, possibly closer to 1 to 1 later |
| DRAM demand | Expanding beyond GPUs into CPUs |
| Supply pressure | Expected to continue into 2027 |
| Consumer impact | Higher prices and fewer low cost memory options |
The report does not clearly say what type of memory will be used. It could involve CPUs with on package memory, HBM, newer memory standards, or much denser DIMMs. AMD has already made EPYC chips with HBM before, so the idea is not unrealistic. Newer MRDIMM designs could also help raise bandwidth and capacity, though those remain separate memory modules rather than memory built directly into the CPU package.
This growing demand adds more pressure to a memory market that is already tight. DRAM makers are expanding capacity, but those new fabs and lines take time to come online. Samsung has also warned that 2027 could be worse for DRAM supply than 2026.
The effect may not stay inside datacenters. When memory makers focus on high profit AI memory, lower end products can become harder to find or more expensive. Samsung has already moved away from LPDDR4 production, choosing newer and more profitable LPDDR5 instead. If the same pattern continues across other memory types, PC RAM, mobile memory, embedded systems, and cheaper devices could all feel the pressure.
For normal PC buyers, this means RAM prices may stay high for longer. If AI CPUs start using far more memory, the shortage will not be only about GPUs and HBM anymore. It will spread across the whole DRAM supply chain.
The main takeaway is that AI is changing what “enough memory” means. CPUs in AI servers may soon need several hundred gigabytes each, and that could keep the memory market tight well into 2027.
Discussion (0)
Be the first to comment.