Beyond AI Chips: How AI Demand Is Reshaping Semiconductor Supply

Read Below

• AI Demand Extends Beyond Accelerators: Growth in AI infrastructure is reshaping demand for HBM, server DRAM, enterprise storage, advanced packaging and power-conversion technologies.

• Capacity Growth Does Not Guarantee Availability: Manufacturing complexity, packaging constraints and shifting supplier priorities mean higher investment may not translate directly into greater component availability.

• Engineering-Led Supply Planning Reduces Exposure: McKinsey Electronics supports customers with authorized sourcing, component selection and technically evaluated alternatives to address AI-driven capacity pressures before they affect the BOM.

AI-driven semiconductor demand is often measured through GPUs, accelerators and leading-edge process nodes. But the supply impact extends much further.

Deploying AI at scale requires high-bandwidth memory, server DRAM, enterprise storage, advanced packaging, networking and increasingly complex power-conversion systems. As manufacturers direct investment toward these technologies, AI is beginning to influence capacity and development priorities across a much wider part of the semiconductor industry.

For design engineers, this creates a less visible form of exposure: a product does not need to contain an AI accelerator to be affected by the AI supply chain.

AI Is Changing the Economics of Memory Capacity

Memory is one of the clearest examples of this shift.

SEMI expects worldwide investment in 300 mm memory-fab equipment to reach $52 billion in 2026, an increase of 29% year over year. DRAM equipment spending is projected to account for $37 billion, supported by demand for high-bandwidth memory (HBM) and DDR5. Investment in 3D NAND is expected to reach $14 billion as AI infrastructure increases storage requirements.

Higher investment, however, does not translate directly into an equivalent increase in available memory. SEMI notes that effective capacity growth is moderated by technology migration and manufacturing complexity, including the transition to advanced-node DRAM, HBM and higher-layer NAND.

This distinction is important. Wafer starts remain a useful capacity indicator, but they do not show how much usable output a manufacturing line can produce, which devices are being prioritized or where additional complexity may reduce throughput.

As a result, headline capacity figures may not fully reflect the availability of specific memory products.

HBM Changes the Value of Advanced DRAM Capacity

HBM makes this change particularly visible.

Unlike conventional DRAM, HBM vertically integrates multiple memory dies to deliver very high bandwidth close to the processor. Its production combines advanced DRAM processes with multi-die stacking, high-density vertical interconnects such as through-silicon vias (TSVs), and advanced packaging. These requirements increase both manufacturing complexity and the value of capacity capable of supporting HBM.

The effects are already visible across the memory market. In July 2026, SK hynix reported that AI demand was supporting sales of HBM, AI-server DRAM and enterprise SSDs. The company also stated that it had entered long-term agreements with around ten customers to improve medium- and long-term supply stability.

Micron's 2026 AI portfolio reflects the same expansion across the memory hierarchy. It includes HBM4, DDR5 server memory, LPDDR5X-based SOCAMM2 low-power data-center memory and PCIe Gen6 data-center SSDs. This is significant because AI workloads do not depend on HBM alone. Training and inference systems also require larger system-memory pools and high-performance storage.

AI demand therefore propagates beyond the memory connected directly to the accelerator. It can influence investment and capacity priorities across several DRAM and NAND categories.

Advanced Packaging Is Now Part of the Capacity Equation

AI is also narrowing the traditional distinction between semiconductor fabrication and packaging.

In leading AI processors, the package is part of the system architecture. TSMC's Chip-on-Wafer-on-Substrate (CoWoS) technology integrates compute dies and HBM through advanced interposer structures. The company reports that demand for CoWoS increased significantly with the emergence of generative AI and continues to expand the platform to accommodate larger interposers, more silicon and additional HBM. TSMC is now producing 5.5-reticle-size CoWoS and is developing even larger versions for future AI systems.

This changes how semiconductor capacity should be assessed. A finished AI accelerator depends on several tightly connected stages: leading-edge logic fabrication, advanced memory production, interposer manufacturing, assembly and high-density packaging.

Additional wafer capacity at one stage will not necessarily increase finished-device output if another stage remains constrained. For engineers and sourcing teams, front-end wafer capacity can no longer be viewed in isolation. Advanced packaging capacity is increasingly part of the supply picture.

Where AI Demand Extends Beyond the Accelerator

The practical question is not only whether a component is consumed directly by AI systems. Engineers should also consider whether their bill of materials shares manufacturing processes, packaging resources or supplier investment priorities with technologies being expanded for AI.

AI Is Creating a Semiconductor Market Behind the GPU

Power electronics provide another example of the wider effect.

As compute density rises, data centers must convert and deliver significantly more power with greater efficiency. Infineon's 2026 AI data-center platforms illustrate this transition. The company has introduced an 18 kW three-phase power-supply reference design for 50 V rack architectures and a 30 kW three-phase interleaved T-Type power-factor-correction (PFC) evaluation board designed for 800 VDC or ±400 VDC rack architectures. Infineon's broader AI data-center power portfolio spans silicon, silicon carbide (SiC) and gallium nitride (GaN) power devices, together with gate drivers, controllers, sensors and other power-management ICs.

Infineon expects future AI processors to require approximately 2–4 kW per GPU. Delivering this power efficiently—from the grid to processor rails below 1 V—requires more sophisticated conversion architectures and a broader range of semiconductor content.

This does not establish that AI will cause a general shortage of SiC, GaN or power-management components. Current evidence does not support such a broad conclusion. It does show that AI infrastructure investment extends into power conversion, sensing, protection and control—not only compute silicon.

The New Supply Risk Is Second-Order Exposure

Traditional supply analysis focuses primarily on direct exposure: whether a component serves a high-demand application, whether its manufacturer is constrained and whether its process node is heavily utilized.

AI adds another question: What else is competing for the manufacturing ecosystem behind this component?

A product may contain no AI processor yet still rely on memory from a supplier prioritizing AI servers, packaging resources receiving heavy high-performance-computing investment, or power technologies whose roadmaps are increasingly shaped by high-density data centers.

SEMI's July 2026 silicon-wafer shipment report reinforces the breadth of this effect. The organization reported that AI-related demand was contributing to shipment growth not only in advanced logic and memory, but also in power devices and photonics.

This does not mean every component category will become constrained. It means that effective supply analysis increasingly depends on understanding relationships between capacity, technology and supplier priorities—not simply the end application listed for a component.

What This Means for Design and Supply Planning

AI is changing where memory manufacturers invest, increasing the strategic importance of advanced packaging and reshaping data-center power architectures. These shifts can affect technologies used well beyond AI systems.

For engineers developing industrial, automotive, communications or embedded products, the objective is not to assume that AI will create shortages across the market. It is to identify where a design shares manufacturing resources with the areas absorbing exceptional AI-driven investment.

This requires earlier visibility into memory technologies, packaging dependencies, manufacturing processes and alternative components. It also makes close coordination between engineering, procurement and authorized supply partners more important when reviewing availability and lifecycle risk.

McKinsey Electronics supports customers across the Middle East, Africa and Türkiye with authorized sourcing, component selection and technically evaluated alternatives. In a market increasingly shaped by second-order capacity effects, this engineering-led approach helps customers assess supply conditions without treating every market signal as a shortage, or waiting until disruption reaches the BOM.