Morgan Stanley: agentic AI to add $32.5-60B to data-center CPU and memory TAM by 2030

Morgan Stanley research published April 21, 2026 projects that agentic AI workloads will add $32.5 to $60 billion to the data-center CPU and memory TAM by 2030. The report reframes the AI infrastructure spend story from “GPU supercycle” to “balanced silicon stack.”

The thesis

Three workload categories the report argues will drive CPU and memory demand:

1. Agent orchestration. Tool-calling agents run CPU-heavy control flow between LLM call takes milliseconds; the plan-and-dispatch loop takes seconds of CPU work.
2. Inference serving at scale. Memory bandwidth bottlenecks attention layers more than compute does. Memory capacity and HBM pricing become the gating factor, not raw TFLOPs.
3. Long-context inference. 1M-token context windows shift the KV-cache problem into a memory-capacity problem. More memory per serving node directly translates into tokens-per-second.

Named beneficiaries

• AMD: EPYC CPUs in AI-serving racks. EPYC Turin and next-gen Genoa-X positioned for orchestration workloads.
• Intel: Xeon recovery narrative contingent on capturing inference-adjacent CPU share.
• Arm: data-center CPU design wins (NVIDIA Grace, AWS Graviton, Ampere) keep expanding.
• Micron and SK hynix: HBM3E and HBM4 volume. See SK hynix SOCAMM2 192GB mass production.
• TSMC: 2nm and below for both the CPU and memory-controller sides.

Why now

The report crystallizes a shift that’s been building since Q4 2025. Hyperscaler capex has stayed flat or grown, but the composition inside the capex envelope has shifted:

• 2023-2024: GPU-heavy (training runs dominate).
• 2025: GPU-plus-inference-silicon mix (Ironwood-class chips, Trainium, MTIA).
• 2026 onward: inference silicon plus CPU plus HBM-heavy memory for agent serving.

Google’s Ironwood TPU announcement two days later is the same structural story on the accelerator side.

Caveats

• Morgan Stanley’s $32.5-60B spread is wide. The low end reflects modest agent-adoption curves; the high end assumes enterprise agent roll-outs compound through 2028.
• The forecast is TAM expansion, not total data-center spend; CPU share of a growing pie rises only if GPU share plateaus.
• Published equity-research call is not an Apollo-style systematic prediction; treat the figures as a framing, not a ground truth.

Editorial read

-serving unit economics get cheaper, which lets frontier-model API pricing fall further (per-MTok prices for Gemini Flash and Haiku 4.5 already reflect some of this). Tool buyers benefit in two ways: cheaper base pricing and more generous free tiers funded by lower serving cost.

Related

• Google unveils Ironwood TPU at Cloud Next 2026
• SK hynix SOCAMM2 192GB mass production
• Anthropic locks in 3.5GW Google TPU capacity