Chip Stack Is Broadening

Artificial intelligence spending is spreading beyond Nvidia-style graphics chips and into the rest of the server: central processors, memory and the packaging that links them together. (livemint.com) Morgan Stanley said on April 22 that “agentic” artificial intelligence systems — software that chains together multiple steps and tools — shift more work onto central processing units, which coordinate those steps, and onto dynamic random-access memory, which holds working data. The bank estimated that could add $32.5 billion to $60 billion to the central-processor market by 2030. (livemint.com) The same note said central-processor orchestration can account for 50% to 90% of workload latency in agentic systems, and that those workloads could drive 15 to 45 exabytes of extra dynamic random-access memory demand by 2030. Morgan Stanley said that range equals 26% to 77% of projected 2027 annual DRAM supply. (livemint.com) A central processor is the traffic cop in a server, while a graphics processor handles the heaviest parallel math. When artificial intelligence moves from generating one answer to managing a sequence of prompts, tools and memory lookups, the traffic-cop job gets bigger. (livemint.com) Memory is also becoming a performance component, not just a storage bin. Morgan Stanley said persistent context and continuous learning in agentic systems increase content demand across central processors, DRAM, foundry capacity, Ajinomoto build-up film substrates and interconnect layers. (livemint.com) That broader demand is already showing up in company results. SK Hynix said on April 23 that its first-quarter operating profit rose 158% from a year earlier to 7.4 trillion won, as high-bandwidth memory sales to artificial-intelligence customers stayed strong. (reuters.com) SK Hynix said high-bandwidth memory made up 44% of its DRAM revenue in the first quarter, up from 40% in the previous quarter, and it said 2026 high-bandwidth memory volumes were already sold out. The company also said it expected demand for artificial-intelligence server memory to remain solid despite macroeconomic uncertainty. (reuters.com) Taiwan Semiconductor Manufacturing Co. is expanding on the packaging side, which is the step that stacks and connects finished chips so they can move data and power fast enough for artificial-intelligence systems. A company executive said on April 22 that TSMC plans to open a chip-packaging plant in Arizona by 2029. (reuters.com) TSMC already builds chips in Arizona, and the packaging plan would add a later stage of production that the United States has pushed to localize along with wafer fabrication. Reuters reported the Arizona packaging plant would support the company’s advanced chip-on-wafer-on-substrate technology, which is widely used in artificial-intelligence accelerators. (reuters.com) The result is a wider semiconductor trade than the 2023 and 2024 rush into graphics processors alone. If agentic artificial intelligence keeps shifting bottlenecks from raw math to coordination, memory and packaging, more of the chip industry stands to capture the next round of spending. (livemint.com)