The hidden bottleneck beneath the AI boom

Beneath the shiny surface of the artificial intelligence boom lies a quiet crisis that could throttle the industry's next phase of growth. Most high-end processors, including the powerful accelerators driving modern data centers, rely on a specialized insulating base known as an ABF substrate. These components act as the essential bridge between tiny silicon chips and the larger circuit boards they sit on. As demand for AI compute explodes, the supply chain for these substrates is buckling under the weight of exponential growth, creating a bottleneck that hardware manufacturers are struggling to bypass.

The stakes are high because the technology is not easily replaced. Almost every advanced logic chip made by major players like Intel, AMD, and Nvidia depends on this specific packaging method. With millions of new accelerators expected to enter the market, the industry faces a dual challenge: producing enough of these components to meet demand, and engineering them to handle the increasing physical size and complexity of modern AI packages without failing during manufacture.

A food giant controls the chip supply

The most surprising aspect of this supply chain is its extreme concentration. According to reporting from Tom's Hardware, a single company, Ajinomoto, controls roughly 95 percent of the global market for the specific dielectric film used in these substrates. Ajinomoto is better known for producing food seasoning, yet it sits at the base of one of the most critical supply chains in computing. This dependency means that hundreds of millions of semiconductor devices rely on material from a single source, creating a vulnerability that few other parts of the industry share.

This monopoly is now becoming a strain point. As data centers expand to house millions of accelerators, the demand for Ajinomoto’s film is outpacing the capacity of the substrate makers who laminate it. The narrowness of this supply chain means that any disruption or delay in material production can ripple through the entire industry, slowing down the deployment of new AI infrastructure and delaying the hardware that powers large language models.

Larger packages create new physical limits

The problem is not just about quantity; it is also about physics. Modern AI accelerators are becoming larger and more complex, packing multiple compute and memory units onto a single board. To accommodate this, substrates must grow in size and add more layers to route signals and power. Each additional layer requires more of the specialized film, multiplying the material demand. However, making these substrates larger and thicker introduces technical difficulties, including warpage and electrical losses, which can reduce the number of functional chips produced from each wafer.

Manufacturers are finding that the traditional materials are hitting a ceiling. The organic film that has served the industry well for decades is struggling to maintain structural integrity as packages expand. This creates a trade-off where increasing the size of the chip package can lead to lower yields and higher costs, forcing engineers to find new ways to manage heat and signal integrity without breaking the material itself.

Industry looks to glass for alternatives

To break this deadlock, the industry is looking beyond the current standard. Suppliers like Ibiden and Ajinomoto are planning to expand their manufacturing capacity to keep up with immediate demand. Simultaneously, major players such as Intel, Samsung, and SK are exploring glass-core substrates as a potential alternative. Glass offers greater rigidity and can handle larger packages better than the current organic films, potentially solving the warpage and yield issues that are currently plaguing the supply chain.

However, transitioning to new materials is a slow process that requires significant re-engineering of production lines. Until these alternatives are fully viable and scaled, the industry remains dependent on the existing ABF substrate ecosystem. This means that for the near future, the growth of AI infrastructure may be limited not by the speed of the chips themselves, but by the ability to produce the physical bases that hold them.