Most important take away

The AI industry is shifting from generative AI (passive, prompt-and-response) to agentic AI (active, multi-step autonomous workflows), and this transition fundamentally changes the hardware bottleneck from GPUs to CPUs and memory. Morgan Stanley estimates up to $60 billion of incremental CPU total addressable market by 2030, making the supporting supply chain — memory, foundry, substrate, CPU and memory interface, capacitors, and CPU sockets — the key place for investors to position.

Summary

Sean Kim, head of Morgan Stanley’s Europe and Asia Technology team, frames a foundational shift in AI: from Gen AI that helps you think (chatbots that respond to prompts) to Agentic AI that helps you do (autonomous systems that remember, plan, act, and adapt across multi-step workflows). This shift restructures the underlying compute stack and creates a new investment thesis centered on CPUs and memory rather than GPUs alone.

Actionable insights and investment advice:

• The bottleneck is shifting. In Gen AI, GPUs and LLMs dominate the workload. In Agentic AI, CPUs become more critical because they orchestrate tasks and connect systems to broader digital infrastructure. Memory is arguably the most important layer because agents that retain user preferences, documents, tone, and task history create a “context flywheel” — a stickiness moat that grows with use.
• Memory is being redefined from storage to continuity. LLMs have fixed context windows that drop content once exceeded; serious agentic work (e.g., coding agents operating over days/weeks on large codebases) requires persistent memory, short-term orientation, and active retrieval. This drives sustained demand for memory content per system.
• Quantified opportunity: up to ~$60 billion of incremental CPU TAM by 2030, within a total CPU market exceeding $100 billion. Up to ~70% of incremental DRAM/memory shipment is tied to this agentic theme.
• Morgan Stanley is more positive on the agentic AI supply chain into 2027, citing content growth, pricing power, and capacity constraints. Specific areas highlighted:
  • Memory (DRAM and high-bandwidth memory makers)
  • Foundry
  • Substrate
  • CPU and memory interface
  • Capacitors
  • CPU sockets
• No specific tickers are named in the episode, but the thesis points investors toward memory manufacturers, foundries, substrate suppliers, memory interface chip makers (e.g., the kind that supply CXL/DDR controllers), and passive component makers serving server CPU platforms.
• Bottom line for investors: as AI moves from answering to acting, watch the infrastructure behind the shift. The next leg of AI returns may come less from prompt-layer/model companies and more from the processor and memory supply chain enabling agentic workloads.

The episode closes with the standard disclaimer that the content is informational only, not an offer, solicitation, or tax/legal advice.

Chapter Summaries

• Introduction — Setup of the shift: today’s chatbot interactions are user-driven (ask, refine, copy, check), but a new class of AI acts autonomously, remembers context, plans workflows, and adapts.
• Gen AI vs. Agentic AI — Gen AI is passive (prompt in, answer out); Agentic AI is active, more autopilot than co-pilot, executing multi-step workflows on the user’s behalf.
• The Three Stacks of Agentic AI — Brain (LLM), orchestration (CPU managing the doing), and knowledge (memory). Memory is the most important layer because it personalizes the agent and creates a context flywheel that increases switching costs.
• Rethinking Memory as Continuity — Memory is not just storage; it is long-term state, shared knowledge, and behavioral grounding. Fixed LLM context windows are a major limitation for serious agentic work like coding agents, requiring persistent memory and active retrieval.
• Investment Implications — Bottlenecks are moving from GPUs to CPUs and memory. Morgan Stanley quantifies up to $60B incremental CPU TAM by 2030 and up to 70% of incremental memory shipments tied to agentic AI, and is bullish on the memory, foundry, substrate, CPU/memory interface, capacitor, and CPU socket supply chains into 2027.
• Closing — The next big AI leap is less about the prompt and more about the processor; standard disclaimer about informational content.