Agentic AI rewrites AI infrastructure demand, with CPU and memory becoming the second growth pillar

Overview

The report discusses how the spread of agentic AI across the global technology sector is changing AI infrastructure demand. Morgan Stanley believes agents are no longer just an early concept, but are increasingly being deployed in enterprise productivity, advertising, commerce, and cloud computing use cases. Because agent workflows include planning, retrieval, tool calls, execution, and iteration, AI data centers need stronger CPU orchestration, memory, networking, and storage capabilities in addition to GPUs.

Core views

The core view is that the bottleneck in AI infrastructure is shifting from pure model compute to system architecture. GPUs remain important, but agentic AI will increase the CPU-to-GPU ratio and lift demand for CPUs, DRAM, NAND/eSSD, HDD, advanced packaging, and the related semiconductor supply chain. The report keeps the bear-case CPU TAM at USD 77 billion, raises the 2030 base case to USD 125 billion, and lifts the bull case to USD 283 billion. On memory, the report believes the CPU-driven framework will generate about 74 EB to 221 EB of additional DRAM demand.

Analysis framework

The report uses both bottom-up and top-down frameworks to estimate orchestration CPU TAM, and then maps CPU demand into DRAM demand. The bottom-up approach focuses on the size of knowledge workers, AI adoption rates, concurrent sessions, the number of agents per session, the number of CPU cores per token, and changes in pricing. The top-down approach derives market size from installed AI data center GW, rack counts, and the CPU-to-GPU ratio.

Methodology notes

• Market sizingBottom-up orchestration CPU TAM model

  Knowledge workers, agent sessions, concurrency rates, agents per session, CPU core counts, and ASPs together determine orchestration CPU demand.

  The report assumes roughly 1 billion knowledge workers globally by 2032, AI adoption approaching 99% by 2030, concurrent session rates rising from 4% in fiscal 2026 to 19% in 2030, and a significant long-term increase in agents per session, which leads to a 2030 orchestration CPU TAM of about USD 79 billion.

• Market sizingTop-down AI data center GW model

  Potential CPU demand is estimated from AI data center installed capacity, rack power, and the CPU-to-GPU ratio.

  Starting from an assumption that AI data center installed capacity rises from about 24 GW to about 35 GW, the report uses NVIDIA racks as a proxy and assumes the CPU-to-GPU ratio improves from 1:2 to 2:1, resulting in a 2030 bull-case orchestration CPU market of about USD 238 billion.

• Demand transmission frameworkCPU-led memory demand model

  Agentic AI increases effective context capacity per request, CPU-side memory provisioning, and rack SSD demand.

  The report argues that marginal memory demand is no longer driven only by GPU-side HBM, but shifts toward DRAM serving the CPU orchestration layer, rack SSDs, context storage, intermediate states, and warm KV caches.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• INTEL CORP (US.INTC)

  A beneficiary exposed to server CPUs, host CPUs, wafer capacity, and advanced packaging

  Strengths

  The report cites Intel's 1Q26 comments saying the shift from foundation models to inference and then to agentic AI is driving demand for Intel CPUs, wafers, and advanced packaging. DCAI revenue grew 22% YoY, and Xeon 6 is being used as the host CPU in NVIDIA DGX Rubin NVL8.

  Weaknesses

  The report does not provide a specific Intel rating, target price, or share-gain estimate, and agentic AI CPU competition also includes Arm, AMD, and cloud vendors' in-house chips.

  Comparison

  Compared with GPU-centric beneficiaries, Intel is more exposed to the CPU host, orchestration, and packaging layers; compared with the Arm ecosystem, its advantage comes more from its installed server CPU base and manufacturing/packaging capabilities.

  Risks

  If agentic AI adoption is slower than expected, cloud vendors increasingly use in-house Arm CPUs, or the CPU-to-GPU ratio improves less than assumed, incremental demand could come in below the model.

• ADVANCED MICRO DEVICES INC (US.AMD)

  A beneficiary exposed to server CPUs and AI infrastructure CPUs

  Strengths

  The report says AMD materially expands the server CPU market, and rising demand from agentic AI workloads and high-core-count CPUs is favorable for its server CPU business.

  Weaknesses

  The provided content does not disclose AMD-specific financial data, ratings, or target prices, and AMD faces competition from Intel, Arm, and cloud vendors' in-house CPUs.

  Comparison

  AMD is a direct CPU-side beneficiary alongside Intel; relative to Arm, its benefit path is more tied to x86 server CPUs and cloud/data center deployments.

  Risks

  Server CPU competition, customer chip self-sufficiency, changes in the pace of AI infrastructure spending, and valuations that already reflect some optimism.

• DRAM supply chain

  Agentic AI lifts CPU-side memory and context storage demand

  Strengths

  The report expects incremental DRAM demand of about 74 EB in the base case and about 221 EB in the bull case, and notes that Samsung and SK Hynix are advancing 3- to 5-year long-term agreements with large tech customers.

  Weaknesses

  DRAM demand estimates depend on assumptions about CPU TAM, concurrency rates, agents per session, and memory configuration, so parameter changes can materially alter the outcome.

  Comparison

  Compared with the early AI cycle driven by HBM, the incremental demand from agentic AI places more emphasis on CPU-side DRAM, rack SSDs, and context-state storage.

  Risks

  Long-term contract pricing, supply expansion, the pace of customer demand realization, and changes in AI workload architecture could all affect earnings leverage.

• NAND/eSSD, HDD, and storage

  Supports context, intermediate states, and warm KV caches for agent workflows

  Strengths

  The report lists NAND/eSSD and HDD as full-stack beneficiaries, arguing that agentic AI requires more persistent memory, external tools/APIs, and data-path support.

  Weaknesses

  The report provides less specific disclosure on storage capacity and revenue elasticity than it does for CPUs and DRAM.

  Comparison

  Relative to CPUs and DRAM, storage benefits depend more on whether agentic applications create large amounts of persistent context and intermediate-state reads and writes.

  Risks

  If agent architectures optimize away storage usage, or if storage supply expansion pressures prices, the upside may be limited.

• Foundry, advanced packaging, substrates, and semiconductor equipment

  Upstream manufacturing and materials beneficiaries of AI infrastructure expansion

  Strengths

  The report's global beneficiary list includes TSMC, ASML, AMAT, KLAC, Tokyo Electron, Ulvac, ABF substrates, and PCB/material companies, showing that the benefit chain spans manufacturing and equipment.

  Weaknesses

  Different segments will see different timing and earnings leverage, and the report does not provide quantitative estimates for each one.

  Comparison

  These assets are not direct sales carriers of agentic CPU demand, but they benefit from AI chip growth, advanced packaging, and system-level hardware expansion.

  Risks

  Export controls, capex cycles, customer concentration, and mismatches between advanced-node and packaging supply and demand.

Key data

• 2030 base-case server CPU TAMUSD 125 billionRaised from the prior framework, including about USD 79 billion in the orchestration CPU layer as well as the host/cloud CPU market.
• 2030 bull-case server CPU TAMUSD 283 billionIncludes about USD 238 billion of orchestration CPU market demand plus host and cloud CPU demand.
• Bear-case CPU TAMUSD 77 billionThe report says the prior bearish scenario assumptions are maintained.
• Incremental DRAM demand in the base caseapproximately 74 EBDriven by agent scheduling workloads and higher CPU-side memory demand.
• Incremental DRAM demand in the bull caseapproximately 221 EBAbout 4.9x the 45 EB DRAM market in 2026.
• Host CPU market assumptionabout USD 45 billion in 2030The report uses Mercury data to estimate the Grace, Vera, Graviton, and other host CPU markets.
• CPU server salesUSD 31 billion in FY2026 to USD 45 billion in FY2030Used to model the combined main-node CPU and general cloud CPU server mix.
• Arm data center demand signalCustomer demand for new Arm AGI CPUs exceeds USD 2 billion in FY27 to FY28The report views this as evidence that AGI/agentic CPU opportunities are accelerating.
• Intel DCAI revenueUSD 5.1 billion, up 22% YoYThe report cites Intel's 1Q26 disclosure, highlighting Xeon 6 as the host CPU for NVIDIA DGX Rubin NVL8 and Google deployments.
• Meta-AWS Graviton agreementtens of millions of Graviton coresThe report says this is one of the clearest real-world examples of agentic CPU deployment.

Impact & implications

If the report's view proves correct, the beneficiaries of AI infrastructure capex would expand from GPUs to a much broader system-level supply chain. CPU vendors, the Arm ecosystem, DRAM and storage suppliers, foundries and advanced packaging, substrates, interface chips, BMCs, connectors, ODMs, and semiconductor equipment companies could all see incremental demand. From an investment standpoint, the market needs to reassess the long-term demand elasticity of agentic AI for CPUs and memory, as well as the support that long-term supply agreements provide for pricing and supply visibility.

Risks

• Agentic AI is still in the price-discovery stage, and the eventual commercial breakthrough and revenue scale remain uncertain.
• CPU TAM and DRAM demand estimates are highly sensitive to assumptions about AI adoption, concurrency, agents per session, the CPU-to-GPU ratio, and ASPs.
• Some CPU and memory stocks have already reached new highs since mid-April, and valuations may already reflect part of the bullish outlook.
• Competition among cloud vendors' in-house chips, the Arm ecosystem, and x86 CPU vendors could change individual market-share outcomes.
• Export controls, U.S. Executive Order 14105, and related entity restrictions may affect certain investment and trade activities.
• If long-term storage agreements fall short on price floors, supply commitments, or prepayment structures, the structural improvement in the storage industry could be less than expected.

What to watch

• How AMD, Arm, and Intel discuss server CPUs, AGI CPUs, host CPUs, and packaging demand in future quarters.
• Whether large-scale agentic CPU deployment cases such as Meta-AWS Graviton expand to more hyperscale cloud customers.
• The pace of deployment, enterprise adoption, and revenue conversion for agent products at Microsoft, Google Cloud, and Meta.
• Whether the CPU-to-GPU ratio moves from 1:2 toward 2:1, and whether AI data center installed GW approaches the report's assumptions.
• The terms and execution progress of 3- to 5-year long-term agreements between storage vendors such as Samsung and SK Hynix and large tech customers.
• Whether DRAM, NAND/eSSD, and HDD prices, inventories, and supply-demand tightness validate the incremental demand.