Page 117 - 電路板季刊第111期
P. 117
電路板季刊 2026.4 產業脈動 115
The supplier' monopoly: is Nittobo's tible to warpage causing the assembly
position similar to Ajinomoto's ? to bow or twist during reflow, reducing
In the semiconductor industry's yield. T-Glass, with its CTE of approx-
collective memory, the 2021 ABF mate- imately 2.8 ppm/°C, and a next-gener-
rial shortage marked a lesson of how a ation variant targeting 2.0 ppm/°C by
single material can paralyze an entire 2028, provides the dimensional stabil-
supply chain. ity that no standard E-glass (CTE of
5.5 ppm/°C) can match. Without it, the
Proposal to make it more dynamic: largest and complex AI packages sim-
Ajinomoto Build-up Film (ABF), a dielec- ply cannot be manufactured at accept-
tric layer most designers or VPs had able yield.
never considered a strategic asset,
The parallel to Ajinomoto's monop-
suddenly became a symbol of systemic oly on ABF film is instructive but imper-
fragility when substrate manufacturers fect. Therefore, the differences make
ran out of qualified capacity. But simply the T-Glass situation more precarious.
adding more ABF film couldn't solve the
bottleneck, the real constraint was fac- ABF was a proprietary dielectric
tory throughput, yield compounding, and film from a single supplier, substrate
long qualification cycles. makers could, and eventually did, qual-
ify alternative build-up materials from
Four years later, the industry finds Sekisui Chemical, Sumitomo or start-
itself staring at a remarkably analo- ups like Thintronics and others, even if
gous weakpoint. it took almost a year.
Only this time, the material in ques- T-Glass, by contrast, is not a propri-
tion is even less known and the sup- etary product. It is the output of an entire
plier concentration even more extreme: specialty glass-melting and fiber-drawing
T-Glass, a low coefficient of thermal infrastructure that Nittobo has refined
expansion fiberglass cloth produced over decades.
almost exclusively by Japan's Nitto Building a new glass furnace requires
Boseki, better known as Nittobo. approximately two years of construction
T-Glass matters because of a sim- and commissioning. Producing yarn with
ple physics problem. As AI chip pack- the required uniformity, dielectric prop-
ages and substrates are getting larger erties, and coefficient of thermal expan-
for higher compute power and include sion (CTE) also demands deep process
more HBM memory and chiplets' expertise.
approach, the IC substrates that sup- Weaving that yarn into cloth with
port them must also expand. Larger IC the precise thread counts and surface
substrates are more and more suscep- finishes needed for high-end cop-
