Solving the AI accelerator supply crunch with ABF substrates


Featured image Solving the AI accelerator supply crunch with ABF substrates

The Invisible Foundation: How Substrates Are Holding Up the AI Revolution

At the heart of every cutting-edge processor, from consumer PCs to colossal AI accelerators, lies a foundation often overlooked: the ABF substrate. These specialized insulating and wiring bases connect the tiny silicon chips above them to the massive printed circuit boards below, and their story is one of exponential demand colliding with complex physical and supply chain constraints.

When the artificial intelligence boom exploded, the demand for these substrates multiplied instantly. Training and inference for frontier AI models now require enormous computational power, driving data centers to scale into the gigawatt era. Companies like Nvidia ship millions of GPU packages, each relying on the ABF substrate, which has been critical to the semiconductor industry for decades. This dependence means that the seemingly abstract world of AI compute is resting entirely on this specialized material.

But this foundation is built on a remarkably narrow supply chain. The entire ecosystem—from the raw film material to the finished multilayer substrate—is highly concentrated. The linchpin is the Ajinomoto build-up film (ABF), a material supplied by a single company that accounts for a significant portion of the global market. This concentration introduces a significant vulnerability, demonstrating that the growth of AI compute is increasingly constrained by the limits of material supply.

The challenge isn’t just volume; it’s complexity. As AI accelerators evolve, designers are pushing for multi-die, multi-stack packages that pack compute and memory onto a single board. This architectural shift forces the substrate to expand dramatically in both the X-Y footprint and along the Z-axis. To accommodate these larger packages and the increased signal routing required, manufacturers must add more layers of ABF, further multiplying the demand across the entire ecosystem.

This physical expansion brings technical hurdles. Increasing the size of the substrate creates problems with warpage and yield, as materials expand differently when heated during assembly. Furthermore, adding more layers strains manufacturing capacity, demanding extremely tight tolerances and pushing the limits of what is physically manufacturable. The ecosystem faces a dual challenge: producing enough advanced substrates while simultaneously engineering them to scale without becoming impractical or unmanufacturable.

To push past these technical and supply chain bottlenecks, the industry is pursuing a multi-pronged roadmap. On the supply side, major players are investing heavily in capacity expansion. Companies like Ibiden and Unimicron are pouring billions into capital spending to expand their manufacturing capacity, aiming to double or triple current output to meet the demands of AI-server substrates. Meanwhile, Ajinomoto is also expanding upstream, building new facilities to increase film production capacity.

Simultaneously, innovation is addressing the physical limits of organic substrates. The focus is shifting toward new materials and architectures. Research is exploring alternatives like glass-core substrates, which offer superior dimensional stability and lower dielectric loss, potentially solving the warpage problem inherent in large organic substrates. This shift involves collaborations among substrate makers, glass specialists, and packaging houses.

The future of substrate technology involves a blend of capacity growth and material science breakthroughs. While the immediate focus remains on expanding manufacturing lines and securing the supply chain, the long-term vision points toward substrates that are inherently flatter, more robust, and capable of handling the extreme demands of future AI hardware. The evolution of the ABF substrate story is not just about materials; it’s about redefining the very limits of semiconductor packaging.

You may also like: