Artificial intelligence depends on more than the companies designing well-known AI chips. It depends on an interconnected semiconductor chain that begins with specialized equipment and materials, moves through wafer fabrication and advanced packaging, and ends with the GPUs and application-specific integrated circuits (ASICs) that power modern AI systems. This is Post 4 in our eight-part AI investment series. In the broader value chain, semiconductors sit between physical infrastructure and computing: AI Materials → AI Power → AI Physical Infrastructure → AI Semiconductors → AI Compute & Networking The semiconductor layer is a chain within that chain. Its links include equipment, wafers and materials, foundries, advanced packaging, memory, and GPU or ASIC design. Each link depends on the others. A portfolio that owns one chip designer is not equivalent to a portfolio that is exposed across the semiconductor ecosystem. Why the Semiconductor Layer Is Necessary AI models require substantial computing power during both training and deployment. Training involves processing enormous datasets to develop a model. Inference: the process of generating a response or prediction: requires the model to run repeatedly for users, businesses, and machines. General-purpose central processing units (CPUs) remain important, but AI workloads increasingly rely on specialized processors. Graphics processing units (GPUs) and custom ASICs can perform many mathematical operations in parallel, making them better suited to the matrix calculations used by modern AI models. However, the processor itself is only one part of the system. An AI accelerator also requires: Leading-edge logic transistors High-bandwidth memory (HBM) Advanced substrates and interconnects Sophisticated packaging Testing and quality-control equipment Reliable fabrication capacity Without these supporting components, there is no functioning AI infrastructure. The economic value of a GPU is therefore tied to the wider supply chain that makes it possible to produce, package, connect, and operate the chip. That interdependence is the central investment point: AI is a massive economic value chain, and every layer depends on the layers underneath it. The Demand Drivers: AI Capital Spending Moves Upstream The semiconductor industry is cyclical, but AI-related capital spending is creating a powerful near-term demand impulse. Hyperscale cloud providers and large enterprises are investing heavily in data centers, servers, networking, and specialized AI hardware. The primary demand drivers include: Hyperscaler and enterprise AI investment Cloud providers are expanding computing capacity to support model training, inference, software development, and enterprise applications. Corporations are also building private AI infrastructure or purchasing capacity from cloud platforms. These investments flow backward through the chain: Data center construction supports demand for servers. Servers require accelerators and networking silicon. Accelerators require advanced packaging and HBM. Packaging and memory require specialized equipment and materials. Foundries must expand leading-edge wafer capacity. GPU and ASIC demand GPUs remain central to AI training and many inference workloads. ASICs are custom-designed processors optimized for specific applications or workloads. Hyperscalers increasingly use ASICs to improve performance, control costs, and reduce dependence on third-party processors. This creates demand for both merchant chip designers and the foundries that manufacture their designs. High-bandwidth memory shortages HBM is a specialized form of dynamic random-access memory (DRAM) designed to move data rapidly between memory and the processor. Rather than relying only on conventional memory modules, HBM stacks multiple memory dies vertically and places them close to the accelerator. AI processors require rapid access to large volumes of data. HBM therefore acts as a critical performance component, not merely a supporting accessory. The limited number of HBM suppliers and the complexity of producing and packaging stacked memory have contributed to tight supply. Advanced packaging bottlenecks Traditional packaging connects a finished chip to a circuit board. Advanced packaging integrates multiple dies, memory stacks, and interconnects into a compact, high-performance package. TSMC’s CoWoS technology is one example. CoWoS: short for Chip-on-Wafer-on-Substrate: allows logic chips and HBM stacks to work together in a single package. As AI accelerators have become more complex, advanced packaging capacity has become a significant production constraint. Fabrication capacity buildouts Foundries are expanding capacity for advanced process nodes, including 3-nanometer and future 2-nanometer technologies. The expansion is occurring alongside investments in U.S., Asian, and European semiconductor manufacturing. Capacity expansion increases the long-term supply of AI silicon, but it also introduces execution risk, cost overruns, and the possibility of overbuilding if demand normalizes. The Semiconductor Sub-Chain in Plain English The industry can appear opaque because several companies may contribute to one finished AI processor. The following sequence provides a simplified map. 1. Semiconductor equipment Equipment manufacturers build the specialized machines used inside semiconductor factories. These tools deposit, etch, measure, clean, and inspect microscopic layers on a wafer. Lithography equipment is used to print circuit patterns. Deposition and etching tools build and remove material. Metrology and inspection systems help manufacturers identify defects and improve production yields. Illustrative companies include ASML, Applied Materials, Lam Research, Tokyo Electron, and KLA. Advanced packaging and bonding equipment suppliers, such as BE Semiconductor Industries and Kulicke & Soffa, serve a different but increasingly important part of the chain. 2. Wafers and materials A wafer is a thin, highly purified slice of semiconductor material: usually silicon: on which circuits are built. Manufacturing also requires photoresists, specialty gases, chemicals, substrates, films, and other engineered materials. At advanced nodes, small variations in materials can affect yield, reliability, and performance. This gives specialized materials suppliers an important role even though they are less visible than chip designers. 3. Foundries A foundry manufactures chips designed by other companies. It operates expensive fabrication facilities, or fabs, and sells production capacity to customers. Taiwan Semiconductor Manufacturing Company (TSMC) is the most prominent pure-play foundry. Samsung operates both foundry and memory businesses, while Intel is developing its own foundry operations. GlobalFoundries and United Microelectronics serve important specialty and mature-node markets. A foundry’s competitive position depends on process technology, manufacturing yield, capacity, customer relationships, and capital discipline. 4. Advanced packaging After a wafer is fabricated, individual dies must be cut, connected, protected, and assembled into a finished package. Advanced packaging can combine several dies, chiplets, and HBM stacks in one system. This process is