After the rise of intelligent agents, the distribution of the entire AI value chain has changed.

The main narrative of AI investment is undergoing a structural shift. Morgan Stanley’s latest research indicates that as AI moves from “content generation” to “automated task execution,” the incremental logic of the next wave of AI infrastructure will expand from a “single-chip computing power race” to a “full-stack system engineering” — GPUs remain central, but no longer enjoy exclusive budget and premiums.

According to Chase Wind Trading Desk, Morgan Stanley research analyst Shawn Kim directly wrote in the report, “Intelligent agents mark a structural shift from computation to orchestration.” In intelligent agent workflows, orchestration time on the CPU side can account for 50% to 90% of total latency, leading to an estimated additional CPU market size of $32.5 billion to $60 billion by 2030, and pushing the total server CPU TAM to the $82.5 billion to $110 billion range.

Meanwhile, DRAM, ABF substrates, wafer foundries, storage, connectors, and passive components will all leap from supporting roles to new bottlenecks and profit pools. This will generate an extra 15 to 45 exabytes of DRAM demand by 2030, equivalent to 26% to 77% of the entire industry’s annual supply in 2027.

This assessment implies that the beneficiaries of AI capital expenditure will spread from a few chip giants to the entire global supply chain. The next round of excess returns may come more from those “enabling links” that first become bottlenecks in intelligent agent workflows and are hardest to rapidly expand production. As bottlenecks shift across different segments, the weight distribution along the AI value chain will also change.

From “Generation” to “Action”: Intelligent Agents Shift Bottlenecks from Computing Power to Orchestration

The typical workflow of generative AI is relatively simple: after a user request arrives, the CPU performs minimal preprocessing, the GPU handles token generation, and then the result is returned. Throughout this chain, the GPU is the absolute protagonist, with the CPU playing a supporting role.

The logic of intelligent agents is fundamentally different. Completing a task requires multiple steps: planning, retrieval, invoking external tools and APIs, execution, reflection, and iteration. It also involves multi-agent collaboration, permission management, state persistence, and scheduling—many “control plane” capabilities. Morgan Stanley’s core conclusion is: intelligent agents do not bring “heavier” single inference tasks, but rather more steps, more states, and more coordination, which are naturally better suited for CPU processing.

This leads to two direct consequences: first, the ratio of CPUs to GPUs at the cluster level will systematically increase; second, DRAM will rise from a “capacity configuration” to a “core system component for performance and throughput.” Bottlenecks in data centers will increasingly appear in memory bandwidth, data transfer, interconnect latency, and system-level coordination, rather than solely GPU computing power.

CPU Ratio Reassessment: Moving from “1:12” Toward “1:2” or Reversal

In the past, a typical architecture for AI servers was “one CPU serving about 12 GPUs.” However, the report notes that as intelligent agent workflows lengthen and tool invocation and context management become more complex, this ratio is rapidly narrowing.

For example, NVIDIA’s roadmap estimates that near the Rubin platform, the CPU-to-GPU ratio is already close to 1:2; with more aggressive evolutions like Rubin Ultra, a reversal with 2 CPUs per GPU may even occur. Even improving from 1:12 to 1:8, the absolute CPU demand for large-scale deployments will see a significant jump.

If this trend holds, CPU demand elasticity will shift from “following server shipments” to “tracking intelligent agent complexity,” meaning CPU growth will become more structural rather than just a continuation of traditional hardware upgrade cycles.

Recalculating CPU TAM: $82.5 billion to $110 billion by 2030, with incremental gains from orchestration

Morgan Stanley adopts a “system layering” approach, separating the CPU opportunities brought by intelligent agents from the traditional server upgrade logic, establishing three independent analysis perspectives:

• Head Node CPU
  Corresponds to the rack control layer close to GPU systems. Assuming about 5 million AI accelerators globally by 2030, each with 2 high-end CPUs, and an average CPU price of around $5,000, this yields approximately $50 billion TAM.

• Orchestration CPU
  Covers the additional needs of intelligent agent orchestration, including planning and scheduling, toolchains, RAG pipelines, KV caches, vector databases, policies, and observability. Estimating an additional 10 million to 15 million CPUs at an ASP of about $3,000, this results in $30 billion to $45 billion TAM.

• Other CPU
  Includes storage nodes, some network nodes, etc., accounting for about $79B to $15 billion.

• 

Total, the server CPU TAM in 2030 is approximately $82.5 billion to $110 billion, with the incremental contribution from intelligent agents around $32.5 billion to $60 billion. The underlying assumption for this estimate is a global AI data center infrastructure sales of about $1.2 trillion by 2030 (compared to approximately $242 billion in 2025).

The report also provides an “upward revision switch”: if AI infrastructure sales reach $3 trillion or $5 trillion by 2030 (per NVIDIA estimates), the CPU TAM range could be pushed to $20.8 billion to $2.5B, or even $34.4 billion to $12k by 2060. This is not a baseline forecast but highlights the systemic amplification effect of “AI factory” scale expansion on CPU demand.

Memory Shifting from Supporting Role to Mainline: Additional DRAM Demand of 15 to 45 EB by 2030

The real differentiation of intelligent agents lies not only in reasoning ability but also in “sustainable context and memory.” Continuous context, KV-cache, intermediate states of tool invocation, and concurrent agent worksets make DRAM on the CPU side essentially an extension of HBM.

The calculation model is straightforward: additional DRAM demand equals the number of new orchestration CPUs multiplied by the average DRAM per CPU. Two assumptions are considered: 10 million CPUs with about 1.5TB each, and a more optimistic scenario of 15 million CPUs with about 3TB each. From this, it is derived that intelligent agents could generate 15 to 45 EB of additional DRAM demand by 2030, accounting for 26% to 77% of the DRAM industry’s annual supply in 2027.

Regarding the cycle outlook, the report notes a market structure variable: most memory suppliers are discussing 3-5 year long-term agreements (LTAs) with major clients, which could slow down price declines and improve profit visibility before 2027. “Memory hierarchy is becoming the core monetization path for AI systems” — host DRAM, memory interface chips, CXL extension, and tiered storage like SSDs/HDDs will all become more sustainable value points.

Segments with tighter supply will have greater pricing power: ABF substrates, foundries, and enabling components

The segments with the most potential for excess profit are those “slow to expand capacity and with long validation cycles.” The report highlights the following chains:

ABF substrates: The current AI-driven upward cycle for ABF substrates may extend into the end of this decade, with supply-demand gaps around 2026-2027. Just the “CPU TAM expansion” alone could lead to a 5% to 10% upward revision of ABF demand by 2030; the server CPU ABF substrate market is estimated to reach about $4.7 billion by 2030, with incremental demand of about $1.2 billion from CPUs.

Wafer foundries (especially advanced nodes): The CPU foundry market is estimated at about $33 billion in 2026 and about $37 billion in 2028. TSMC’s share in CPU foundry is expected to increase from about 70% in 2026 to around 75% in 2028; Intel may start outsourcing server CPU manufacturing to TSMC in late 2027.

BMC and memory interfaces: Aspeed is highlighted as the core beneficiary in CPU server BMC, holding about 70% of the market share in this segment. The new AST2700 platform could bring a 40% to 50% ASP uplift; Montage is positioned in the “memory interconnect” value chain, with a global revenue share of about 36.8%.

CPU sockets and passive components: The report uses Lotes and FIT as direct mappings for CPU sockets. For every additional 1 million CPUs demanded, Lotes revenue could increase by about 0.6%, and FIT by about 0.2% (based solely on socket count). For passive components, assuming about $30 worth of MLCC content per general-purpose server, the additional demand in 2030 could be around $500 million, representing about 2% to 3% of the global MLCC market at that time.

CPUs are the clearest incremental opportunity, but “enabling links” are more favored

The report admits that the growth of intelligent agent workloads will structurally benefit AMD’s cloud share, but maintains an Equal-weight rating on both AMD and Intel. It leans toward tracking intelligent agent themes through targets like NVIDIA and Broadcom, where capital expenditure and token growth more directly translate into profits, while valuation constraints are considered important.

From a macro perspective, the core value of this report is elevating AI investment from a “single-point compute arms race” to a “system efficiency and bottleneck economics”: GPUs are the engine, CPUs are the transmission and control systems, memory and interconnects are the oil and chassis — single-point extremes remain important, but the scale of returns depends on vehicle-wide coordination.

For the industry chain, this means that excess AI investment returns will become more dispersed and long-term: not only from the “most powerful GPUs,” but also from those “enabling links” that first become bottlenecks in intelligent agent workflows and are hardest to rapidly expand. High-frequency, trackable indicators include: upward revisions in CPU count and memory configurations in new platform BOMs, long-term contract signing rhythms of cloud providers, and utilization trends of ABF substrates and advanced process capacities.