Agentic AI Shifts Bottlenecks From GPUs to CPUs and Memory

The AI Revolution is Shifting Gears: From Thinking to Doing, and the Infrastructure That Will Power It.

This conversation with Shawn Kim, Head of Morgan Stanley’s Europe and Asia Technology Team, reveals a critical, often overlooked, evolution in Artificial Intelligence: the transition from Generative AI (GenAI) that assists with thinking, to Agentic AI that actively does. The immediate implication for users is a move from passive interaction to autonomous workflows. For investors and technologists, however, the hidden consequence is a fundamental shift in computing bottlenecks, moving from GPUs to CPUs and, crucially, to memory. This analysis is essential for anyone building, investing in, or simply trying to understand the future of AI infrastructure, offering a distinct advantage in anticipating the next wave of technological demand.

The Unseen Shift: Why CPUs and Memory Are the New AI Bottlenecks

The current AI landscape is dominated by discussions around Large Language Models (LLMs) and the Graphics Processing Units (GPUs) that power them. This is the era of Generative AI (GenAI), where AI acts as a sophisticated assistant, capable of summarizing text, drafting content, or answering queries. While incredibly useful, GenAI remains largely passive; it responds to prompts, but the user still directs the workflow, refining outputs and executing the subsequent steps. This is akin to having a brilliant but uninitiated intern.

The true paradigm shift, as Shawn Kim articulates, is the emergence of Agentic AI. This is AI that doesn't just think but acts. Imagine an AI that remembers your preferences, understands your long-term goals, orchestrates multi-step processes across different digital tools, and adapts dynamically to changing circumstances. This is not just a co-pilot; it's an autopilot for complex, multi-stage workflows.

"That is the shift from GenAI to agentic AI: from AI that helps with thinking to AI that helps with doing. GenAI is mostly passive. It takes a prompt and produces an answer. Agentic AI is active -- less a copilot for one task but an autopilot for multi-step workflows."

This transition fundamentally alters the computing requirements. While GPUs remain vital for the "thinking" part--the core LLM processing--Agentic AI introduces new demands. The "doing" aspect, the orchestration of tasks, planning workflows, and managing interactions with other systems, places a much greater emphasis on Central Processing Units (CPUs). These are the conductors of the AI orchestra, coordinating the various components and connecting the AI to the broader digital infrastructure.

But the most profound, and perhaps least obvious, consequence lies in the "knowledge" stack: memory. Agentic AI necessitates a persistent, context-aware memory that goes far beyond simple data storage. This memory needs to retain user preferences, document history, communication tone, and task progression. This creates what Kim calls a "context flywheel": the more context an agent collects, the more personalized and effective it becomes, leading to a sticky user experience that is difficult to abandon.

This is a critical distinction from the limitations of current LLMs, which have fixed context windows. Once a conversation or a task exceeds this window, older information is lost. For a simple query, this is manageable. But for an agent tasked with, say, refactoring a large codebase over weeks, this limitation becomes a showstopper. True agentic capability requires persistent memory, short-term orientation, and active retrieval--the ability to recall past decisions, understand file changes, and locate relevant code without explicit user guidance for every dependency.

This shift from passive assistance to active execution, powered by advanced orchestration and persistent memory, redefines the computational bottlenecks. The implications for investors are stark: the demand for CPUs and memory components is poised for significant growth.

The Downstream Effects: Reshaping Tech Supply Chains

The move towards agentic AI isn't just an incremental improvement; it's a catalyst for a significant reshuffling of the technology supply chain. The analysis presented suggests a future where CPUs, not just GPUs, become a primary bottleneck. Kim estimates that agentic AI could drive as much as 60%, or $60 billion, of incremental CPU total addressable market by 2030. This is a substantial expansion within an already large market.

Furthermore, the demand for memory, particularly DRAM, is projected to surge. The need for persistent, context-rich memory for agentic systems means that up to 70% of incremental DRAM bit shipments could be tied to this theme. This isn't just about more storage; it's about more sophisticated, faster, and contextually aware memory that enables the AI agent to maintain state and continuity across complex, long-running tasks.

"We see CPUs as the new bottleneck, with memory seeing the highest content increase. We estimate as much as 60 percent, or $60 billion of incremental CPU total addressable market by 2030, within a total CPU market of more than $100 billion. We also estimate up to 70 percent of incremental DRAM bit shipment tied to this theme."

These shifts have direct consequences for specific segments of the technology supply chain. Companies involved in memory manufacturing, foundry services (which produce the chips), substrates, CPU and memory interfaces, and even CPU sockets are likely to benefit. These areas are positioned to capitalize on increased content per unit, potential pricing power due to high demand and capacity constraints, and the inherent stickiness of the agentic AI model once established.

The conventional wisdom has been to focus on the "brain" of AI--the LLMs and the GPUs that train them. However, the agentic era requires a broader view, encompassing the entire operational infrastructure. The next significant leap in AI capability may not be driven by a more powerful prompt or a larger model, but by the processors and memory systems that enable AI to execute complex actions autonomously and persistently. This requires a strategic pivot for investors and technologists alike, looking beyond the immediate generative capabilities to the underlying infrastructure that makes sustained action possible.

Key Action Items

• Immediate Action (Next 1-3 Months):
  • Educate Teams on Agentic AI: Conduct internal workshops to explain the difference between GenAI and Agentic AI and its implications for current and future projects.
  • Identify Potential Agentic Use Cases: Brainstorm specific workflows within your organization that could benefit from autonomous AI agents, focusing on multi-step processes.
  • Review Current Infrastructure: Assess existing CPU and memory capacity and performance against potential future agentic AI workloads.
• Short-Term Investment (Next 3-9 Months):
  • Prioritize CPU and Memory Component Suppliers: Begin evaluating and engaging with suppliers of CPUs, DRAM, and related interface components, considering their capacity and innovation roadmaps.
  • Explore Persistent Memory Solutions: Investigate emerging memory technologies and architectures that offer enhanced state retention and faster retrieval for AI agents.
  • Pilot Agentic AI Tools: Experiment with early-stage agentic AI platforms or develop small-scale internal agents to understand their practical application and limitations.
• Longer-Term Strategy (9-18+ Months):
  • Develop a "Context Flywheel" Strategy: Design systems and processes that actively collect and leverage user and task context to create increasingly personalized and valuable AI agents. This requires a commitment to data privacy and security.
  • Strategic Partnerships in Foundry and Substrates: For hardware-focused organizations, forge deeper relationships with foundries and substrate manufacturers to secure capacity and co-develop next-generation components optimized for agentic AI.
  • Build for Orchestration Complexity: Invest in robust orchestration layers that can reliably manage multi-agent workflows, handle task dependencies, and adapt to dynamic changes in real-time. This is where immediate discomfort in designing complex workflows pays off in durable competitive advantage.