One of the most significant developments that has recently caught my attention regarding AI infrastructure is that the bottleneck is no longer limited to just GPUs and HBM.

TrendForce’s AI server supply chain monitoring indicates that supply constraints persist for ABF substrates, DRAM, and HDDs. Specifically, lead times for ABF substrates are currently in the 48–56 week range, whereas a balanced market typically sees lead times of around 12 weeks.

So, what is an ABF substrate, and why is it so important? ABF stands for "Ajinomoto Build-up Film." The crucial point here is that ABF is not the substrate itself; rather, it is an ultra-thin insulating film used in high-performance semiconductor package substrates.

To put it simply:

GPU / CPU / ASIC → ABF-based package substrate → PCB → Server

A substrate sits between the chip and the motherboard, carrying dense electrical connections. ABF provides electrical insulation between the layers of this substrate and enables the creation of high-density copper interconnects within very small areas. Ajinomoto notes that ABF is a key material for forming the micro-circuits of build-up substrates used in high-performance CPUs.

The issue on the AI ​​front is this:

AI accelerators are getting larger.

The denser integration of components—such as GPUs, ASICs, and HBM—within a single package causes the package substrate to grow physically and necessitates a greater number of layers.

Larger package → more wiring
More wiring → more substrate layers
More layers → more ABF
Larger substrate → longer production time and risk of lower yield

Consequently, even if the number of AI chips remains constant, the amount of ABF and substrate capacity required per chip can increase. Therefore, viewing the ABF bottleneck simply as a matter of "more GPUs being produced" is an incomplete perspective.

As the package size and layer count of modern AI accelerators increase, substrate manufacturing becomes increasingly difficult. Production issues such as warpage, yield, and electrical losses become more critical with larger substrates.

And the concentration of the supply chain here is noteworthy.

Ajinomoto is the primary supplier of ABF material. Further down the line, companies such as Ibiden, Unimicron, Kinsus, Shinko, Samsung Electro-Mechanics, and Nan Ya PCB utilize ABF and other materials to manufacture high-performance package substrates.

In other words, there is another critical layer behind the GPU in the AI ​​supply chain:

GPU → HBM → Advanced Packaging → ABF/Package Substrate → PCB → Server

For this reason, I consider TrendForce’s report of a 48–56 week lead time for ABF to be a highly significant data point.

Moreover, the issue is not limited to ABF.

On that same supply chain radar:

ABF: 48–56 weeks
HDD: ~50 weeks
DRAM: ~20 weeks

...lead times are observed at these levels. GPUs, with a lead time of approximately 30–40 weeks, do not appear to be as constrained as these other items.

The HDD sector is also particularly noteworthy. Due to the storage demand generated by AI data centers, it is estimated that the supply crunch for HDDs could persist until 2027.

In short, the AI ​​story is not just about a single chip.

GPU → HBM/DRAM → ABF/Packaging → Networking → Storage

The entire chain must invest in capacity simultaneously.

From my perspective, the most important point is this:

When it comes to AI infrastructure, simply asking "how many GPUs will be produced?" is no longer sufficient. The real question is whether the entire system surrounding that GPU can be produced at the same pace.

The GPU might be ready, but without ABF substrates, the package cannot be completed.

The package might be ready, but without DRAM or HBM, the server cannot be completed.

The server might be ready, but if HDD or networking capacity is insufficient, the entire data center might not be able to come online at the planned speed.

That is why I believe we need to keep a closer eye on the companies and components that act as "bottlenecks" in AI investment in the coming period.

The ABF sector is particularly interesting to me because the rise in demand there isn't driven solely by the need for more AI chips; the substrate area and layer count per chip are also increasing.

In short:

As AI chips grow larger, the infrastructure supporting them must expand as well.

ABF is precisely that kind of sector right now.