Morgan Stanley sees agentic AI lifting CPU demand

Morgan Stanley said that more autonomous AI systems could lift CPU demand by 2030, change data center buildouts, and spread AI spending beyond the GPUs that have led the market.

The bank said agentic AI could add US$32.5 billion to US$60 billion to a data center CPU market it estimates will top US$100 billion by 2030, while also raising demand for memory.

Morgan Stanley said companies including Nvidia, AMD, Intel, Arm, Micron, Samsung, SK hynix, TSMC, and ASML could benefit from the shift.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Agentic AI gives CPUs a larger role

As AI systems handle harder jobs with less human input, spending may spread beyond graphics processing units (GPUs), which have led AI hardware so far 1.
Agentic AI leans on central processing units (CPUs) for general work like code compilation, software tools, and database queries, since GPUs are not built for those jobs 1.
Nvidia pitches its Vera CPU as built for agentic AI and reinforcement learning, a way to train AI systems through trial and error 2.
Research firm SemiAnalysis cites Microsoft’s Fairwater data centers for OpenAI, where a 48-megawatt CPU and storage building backs a 295-megawatt GPU cluster, or about a 1-to-6 power ratio 1.
Nvidia also gives CPUs a larger role in agentic AI systems, including in materials for the Vera launch 2.

Agentic AI is squeezing memory

Memory is becoming the main limit for AI systems 3.
Agentic AI can depend on long, persistent contexts, so memory needs can snowball and outrun the capacity of one accelerator, such as an AI chip 4.
SemiAnalysis estimates memory will take about 30% of hyperscaler capital spending in calendar year 2026, up from roughly 8% in 2023 and 2024 5.
SemiAnalysis expects dynamic random-access memory (DRAM) prices to more than double in 2026, while High Bandwidth Memory (HBM), a faster memory used in AI servers, may stay undersupplied through 2027 5.
That could make suppliers like Micron and SK hynix more central as memory limits shape system costs and rollout schedules 3.