ASIC Shipments to Surpass GPUs for the First Time by 2027; ABF Substrates Emerge as New Bottleneck
nashnova research
DIGITIMES projects ASIC shipments will surpass GPUs for the first time in 2027, driven by Google, Amazon, and Huawei; but as packaging capacity eases, the bottleneck has shifted to ABF substrates and high-layer-count PCBs, with tight supply expected through 2028.
ASICs overtaking GPUs — what does this mean for the market?
DIGITIMES forecasts that ASIC shipments will exceed GPU shipments in cloud AI accelerators for the first time in 2027. This means → the AI chip market is no longer Nvidia's alone; custom silicon is taking real share.
The push comes from Google, Amazon, and Huawei scaling up in-house ASICs — chips designed for specific tasks, unlike general-purpose GPUs.
The competitive landscape shifts to "Nvidia vs. Google at the top, everyone else chasing." In plain terms = one dominant player becomes a two-leader race with a crowded pack behind.
Packaging is easing — so where is the bottleneck now?
TSMC's 3 nm process plus CoWoS-L advanced packaging — a method that tiles multiple chips onto a single interposer — is expected to dominate high-end cloud AI chip production through 2027.
TSMC, ASE, and Amkor are aggressively expanding CoWoS capacity; the packaging supply gap is narrowing.
This means → the bottleneck has migrated upstream. The new constraints are T-glass fabric, ABF substrates, high-layer-count PCBs, and HBM memory (HBM3E, HBM4, HBM4E). In plain terms = the queue used to form at the packaging door; now it forms at the materials warehouse.
ABF substrates and high-layer PCBs — what exactly is tight?
ABF substrates — the key base material connecting a chip to its circuit board inside a package — will remain tight through 2027–2028. Ibiden, Unimicron, Kinsus, Nan Ya PCB, Samsung Electro-Mechanics (Semco), and Shinko have all restarted major capacity expansions.
High-layer-count PCBs face similar strain. Layer counts have jumped from 12–20 to 40–50, with some designs using HDI-HLC hybrid lamination. This means → manufacturing difficulty has surged, and fewer fabs can do the work — advanced capacity is consumed fast.
PCB suppliers span Taiwan (Unimicron, ZDT, CGE and others), mainland China (Shenghong, Shennan Circuits and others), South Korea (ISU Petasys), and the U.S. (TTM Technologies).
Why is expansion so slow?
Supply-chain sources say a PCB capacity build takes at least two to three years — from land acquisition and construction to equipment installation. Capital outlays run into the billions. In plain terms = it is not a lack of willingness; building a single fab is slow and expensive.
This reflects a structural mismatch: demand doubles on a yearly cadence, but capacity takes two to three years to come online. Supply elasticity is severely limited.
What test do new ASIC players face?
For fast-rising ASIC makers, the core challenge is that Nvidia, AMD, Google, and Amazon have already locked up most available capacity.
The industry is watching whether these newcomers can qualify new suppliers ahead of schedule or co-invest to secure capacity. This means → chip design is only step one; the ability to mass-produce depends on who secures the substrate and packaging "ticket" first.
市场有风险，内容仅供研究参考，不构成投资建议。