Persistent demand for high-speed data transfer in AI data centers is escalating competition across the semiconductor supply chain to a new front.
According to a report, TSMC has recently indicated that its "COUPE on Substrate" solution, which integrates a compact universal photonic engine, is anticipated to commence mass production in the second half of 2026. The industry interprets this as a pivotal signal that AI chip manufacturing is transitioning from an "advanced node + advanced packaging" phase to a new stage of "co-packaged optics (CPO) + system integration," shifting the competitive focus further towards the integration of ABF substrates and CPO technologies.
As NVIDIA's next-generation Vera Rubin platform continues to enhance the integration of AI GPUs, HBM, and high-speed network interconnects, the subsequent potential bottleneck in NVIDIA's supply chain is likely to be ABF substrates. The company may employ measures such as framework agreements, advance payments, strategic partnerships, or even equity investments to prevent a recurrence of the tight supply situations previously seen with CoWoS and HBM. Market expectations suggest that, driven by concurrent demand from AI GPUs, ASICs, and CPO, high-end ABF substrates could face supply shortages again next year.
TSMC's extension of its CPO technology to the substrate level marks a significant evolution in AI chip packaging. The supply chain notes that placing optical components closer to the compute chip reduces data transmission distance, simultaneously lowering signal loss, latency, and power consumption. This is crucial as AI server competition moves from individual unit performance to efficiency at the rack and cluster level.
NVIDIA has also recently announced a deepened collaboration with Corning to expand its supply layout for optical components in AI data centers, securing capacity for future high-speed optical interconnects. Analysts believe these actions point in the same direction: as cloud service providers persistently pursue higher rack efficiency and lower power consumption, key segments like foundries, HBM, optical fiber, optical modules, and ABF substrates could all become strategic resources for which major AI players preemptively secure capacity.
Information from the supply chain indicates that compared to traditional CPU substrates, the substrate area and layer count required for AI GPUs and ASICs have increased substantially, raising ABF material consumption by 5 to 10 times. With the ongoing rise in demand for AI GPUs, ASICs, and high-end networking chips, the supply-demand balance for high-end ABF substrates is likely to remain tight structurally over the long term.
Analysis suggests that NVIDIA may preemptively secure capacity for high-end substrates through methods like long-term agreements, advance payments, or strategic partnerships to mitigate supply chain risks. This expectation aligns with NVIDIA's previous capacity-securing strategies in the CoWoS packaging and HBM memory sectors.
High-end ABF substrate manufacturers such as Kinsus, Unimicron, and Nan Ya Plastics are all positioned to benefit from this demand trend. Among them, Kinsus, which has entered the supply chain for NVIDIA's Vera Rubin platform, is expected to see relatively pronounced benefits. The industry anticipates that as CPO gradually becomes a mainstream architecture for future AI servers, the strategic importance of high-end ABF substrates will continue to rise, making the capacity and technological capabilities of related manufacturers key leverage in supply chain negotiations.
Comments