Morgan Stanley sees agentic AI widening chip spending beyond graphics processors to CPUs

Morgan Stanley said agentic AI could add $32.5 billion to $60 billion to a data-center CPU market already above $100 billion by 2030, shifting demand beyond GPUs toward CPUs, memory, and chipmaking equipment. The brokerage expects stronger demand for CPUs and memory as AI systems move from generation to autonomous action, with potential beneficiaries including Nvidia, AMD, Intel, Arm, Micron, Samsung, SK hynix, TSMC, and ASML. The note is positive for selected semiconductor and infrastructure names, though it is still analyst commentary rather than a company-specific fundamental update.

Analysis

The market is still treating AI as a GPU-only capex cycle, but the next leg is likely a systems-integration trade: as workloads become more autonomous, the bottleneck shifts from raw inference throughput to orchestration, memory bandwidth, and control-plane compute. That is structurally better for CPUs, DDR/HBM-adjacent memory, and the foundry/equipment stack than for a narrow basket of accelerators, because each deployed agent tends to multiply “background” compute and memory traffic per useful task rather than just replace one prompt with one token stream.

The second-order winner is not necessarily the most obvious CPU vendor; it is the company with the best attach rate into the data-center motherboard, platform software, and memory ecosystem. Intel has the most torque if enterprise buyers want heterogeneous, general-purpose clusters that can absorb agentic workloads without full GPU dependency, while ARM benefits if hyperscalers keep pushing custom server silicon for power efficiency. That said, any sustained shift toward CPU-heavy AI should also tighten supply at memory and leading-edge manufacturing, which likely improves pricing power for Micron, TSMC, and ASML before it materially changes unit volumes.

The main risk is timing: “agentic AI” is a 12-36 month revenue bridge, not a next-quarter catalyst, so the setup works best on pullbacks or via optionality rather than chasing a one-day headline move. The consensus may be overestimating near-term monetization and underestimating how much of the demand reallocation can be internalized by existing hyperscaler designs, which would cap upside for merchant CPU vendors. The cleaner expression is relative-value: buy the ecosystem levered to broader AI capex, not just the pure GPU complex, and fade any knee-jerk overreaction if investors extrapolate a full GPU replacement cycle too quickly.

Key Decisions for Investors

• Long INTC vs short NVDA on a 3-6 month horizon: use the spread to express a shift from GPU dominance to control-plane/CPU rebalancing; target upside if enterprise AI buying broadens, but cut if GPU supply constraints remain the primary capex driver.
• Add TSM and ASML on any 5-8% pullback over the next 2-4 weeks: these are the highest-quality ways to own broader AI intensity with lower product-specific risk; risk/reward improves if the market rotates from chip demand to fabrication bottlenecks.
• Buy a basket long of MU + AMD on weakness for a 6-12 month view: memory and accelerator exposure captures the broader compute-intensity theme without paying full multiple for the headline winners.
• Use call spreads on ARM into the next 2 quarters: best expressed as a platform-weighted beneficiary of custom AI server silicon adoption; defined risk is appropriate because valuation already discounts a lot of future design wins.
• Avoid chasing NVDA strength after the headline; instead, look for relative underperformance versus INTC/ARM/TSM as an entry point. The risk/reward is poorer if this remains a broad AI capex expansion rather than a true architecture rotation.