Rising layer counts multiply ABF film demand faster than accelerator shipments

The reason the film is unavoidable is geometric. Connection pads on a die sit micrometers apart, while the traces on a motherboard sit hundreds of micrometers to millimeters apart ^[2]. Something has to fan one pitch out into the other while distributing power and ground and holding the package rigid, and the default part that does that job is a glass-reinforced resin core with successive build-up layers of copper wiring and insulating film on top ^[3]. Practically every advanced logic and AI chip from Intel, AMD and Nvidia is built on one ^[9].

The demand figure that matters is units multiplied by build-up layers multiplied by substrate area, and all three terms are rising ^[18]. Tom's Hardware puts Nvidia's Blackwell shipments at an estimated 3.2 million packages through the end of 2025, every one of them on an ABF substrate ^[5]. Gigawatt-era sites are being specified at millions of accelerators each ^[6], which makes a single such site the same order of magnitude as the entire cumulative Blackwell package run to date ^[19]. On top of that, packages now carry compute, memory and support components together, so the X-Y footprint is growing ^[11], and layer counts are growing to route the extra signals and power, with each added layer requiring another ABF layer and more manufacturing time ^[12].

So nameplate capacity understates the real draw. Making the substrate bigger brings warpage, yield loss and electrical losses inside the part itself ^[13], so the ecosystem is solving two problems at once: build enough advanced substrates, and re-engineer them so they keep scaling without becoming unmanufacturable ^[14]. Ajinomoto and Ibiden have outlined plans to expand material and manufacturing capacity respectively ^[15], and Intel, Samsung and SK's Absolics are exploring glass-core substrates to get past the limits of organic ABF ^[16]. Exploring is the operative word.

The concentration claim is the strong part of this story. Each substrate maker, whether Unimicron, Ibiden, Kinsus, Shinko, Samsung Electromechanics or Nan Ya PCB, laminates its build-up layers using dielectric film from Japan's Ajinomoto, reported to hold 95% or more of the global market ^[7][8]. A company better known for food seasoning sits at the base of the stack.

The scheduling claim is the weak part. The available account says demand is now growing beyond what the supply chain can comfortably supply ^[10], and that is a qualitative statement. There are no substrate lead times, no allocation splits and no prices in it ^[20]. To conclude that substrates rather than wafer capacity set buildout dates, you would need substrate lead time to exceed accelerator lead time at a named customer, or substrate cost per package to move enough to show up in a quote, and neither figure is on the table ^[20].

What is on the table is a better question to ask a vendor. Layers per substrate, panel size and which substrate house is building it convert an accelerator order into film draw ^[18]; a unit count does not ^[12].