Morgan Stanley CPU thesis

Morgan Stanley said on April 20 that agentic artificial intelligence could widen data-center chip spending beyond graphics processors and add as much as $60 billion to the market for server central processing units by 2030. (reuters.com) Reuters reported the bank put the incremental opportunity at $32.5 billion to $60 billion on top of a data-center central processing unit market already above $100 billion. Morgan Stanley named Nvidia, Advanced Micro Devices, Intel and Arm as potential beneficiaries in central processors and accelerators. (reuters.com) The same note pointed to Micron, Samsung and SK hynix in memory, and Taiwan Semiconductor Manufacturing Co. and ASML in manufacturing and equipment. Reuters said Morgan Stanley also expects tighter parts of the supply chain to gain pricing power if demand keeps rising. (usnews.com) Agentic artificial intelligence refers to software systems that can pursue a goal with limited supervision by splitting work into smaller tasks and coordinating multiple models or tools. IBM says those systems rely on orchestration, real-time decision making and repeated interaction with data, applications and other agents. (ibm.com) That matters for servers because graphics processors excel at parallel math, while central processors still handle general-purpose computing, system control, data movement and many latency-sensitive steps around an artificial intelligence job. Nvidia says its Grace central processor is built for data centers running artificial intelligence, analytics and cloud workloads, and Intel said last week that scaling artificial intelligence requires “balanced systems” rather than accelerators alone. (nvidia.com, intel.com) Advanced Micro Devices makes the same case in its server pitch: many inference workloads still run on central processors and do not need a dedicated accelerator. Arm says its Neoverse designs target “AI-first infrastructure” in hyperscale and enterprise data centers, where central processors coordinate general compute with accelerator-heavy workloads. (amd.com, arm.com) The memory call in Morgan Stanley’s thesis reflects a basic bottleneck: artificial intelligence servers need to move huge volumes of data between chips without waiting. Micron says its HBM4, or high-bandwidth memory, exceeds 2.8 terabytes per second of bandwidth, and SK hynix says its 16-high HBM4 stack is designed to speed both training and inference. (micron.com, product.skhynix.com) The manufacturing names reflect another choke point: packaging and lithography. TSMC says its CoWoS packaging platform is built for high-performance computing and can combine large logic chips with multiple memory stacks, while ASML remains the dominant supplier of extreme ultraviolet lithography systems used to print the most advanced chips. (tsmc.com, asml.com) Morgan Stanley’s timing lines up with a broader spending wave around artificial intelligence infrastructure. In March, Morgan Stanley Research said nearly $3 trillion of artificial-intelligence-related infrastructure investment is expected to flow through the global economy by 2028, with more than 80% of that spending still ahead. (morganstanley.com) The bank’s new point is narrower than the first phase of the artificial intelligence trade: if software agents do more of the work inside data centers, the winners may not be limited to the graphics chips that trained the biggest models. The next test is whether cloud companies and model builders actually buy more central processors, memory and packaging capacity in the numbers Morgan Stanley is projecting. (reuters.com, morganstanley.com)