Jibang Consulting: Cowos-L is expected to remain the mainstream packaging solution for AI chips by 2028

The Zhitong Finance App learned that according to TrendForce Jibang Consulting's latest semiconductor advanced packaging industry research, as GPU manufacturers and large-scale cloud service providers (CSP) continue to develop more powerful and multi-functional AI chips, one development focus of 2.5D packaging technology is to expand the package area. Among them, despite facing competition from Intel (Intel) EMIB-T, TSMC (TSMC) CowOS-L is expected to remain the mainstream packaging solution for AI chips until 2028 due to its technical maturity and yield advantages.

TrendForce Jibang Consulting said that with the support of strong AI demand, NVIDIA (Nvidia) is actively promoting integrated GB/VR cabinet solutions with high chip usage, and it is estimated that in 2026 its high-end GPU shipments will increase by about 30% per year. AMD (Chaowei) will push the MI400/MI450 in the second half of 2026 to help overall shipments.

For CSP's self-developed ASIC section, Google (Google) will have the strongest shipment volume and momentum in 2026, and it is expected that its TPU v7 will be the main product. AWS (Amazon Web Technology) is also actively developing. Starting in the second half of 2026, its chip and AI server systems have gradually advanced to Trainium v3. Currently, Microsoft (Microsoft) and Meta still mainly use GPU servers, and ASIC plans to have a smaller production capacity.

Demand for AI chips is strong, TSMC packaging production capacity is tight, and Intel EMIB-T is expected to receive spillover orders

TrendForce Jibang Consulting pointed out that in the past, AI server chips mainly used TSMC CowOS-S packaging technology to integrate HBM and Compute Die (Compute Die) with a silicon intermediate layer to achieve the computing power performance required for generative AI applications. However, as the chip area continues to expand and there is a need to integrate more HBM modules, silicon interlayer manufacturing faces physical limitations. In addition to changing the material of the intermediate layer, it is also necessary to switch to a CoWS-L solution with a larger exposure ratio for photomasks (also known as photomasks) to solve the packaging and transmission problems of Compute Die and HBM with greater flexibility.

Specifically, Cowos-S technology can integrate 2 SoC/Chiplets and 8 HBM modules, and will continue to be adopted by ASICs such as the Microsoft Maia 200. However, when you need to integrate more SoC/Chiplet and HBM modules, you need to use the Cowos-L solution. Currently, NVIDIA and AMD AI accelerators have adopted Cowos-L; in addition to the MTIA 400 mass-produced by Meta in 2026, the AI ASIC is estimated that AWS and Microsoft will also introduce CoWoS-L technology in 2027.

However, CoWoS-L increases the size of a single chip with a larger intermediate layer, and increases the cost of chip packaging. Coupled with the short supply of TSMC production capacity, Intel EMIB technology becomes an alternative. The EMIB solution reduces the expensive intermediary layer, and the overall cost is expected to decrease, but package module performance is limited by carrier board wiring density; in addition to strengthening Line/Space (L/S) and Microbump (microbump) specifications, Intel is also developing advanced solutions such as EMIB-T to shorten the signal path and enhance EMIB module performance by introducing TSV.

TrendForce Jibang Consulting observed that in addition to Google being expected to introduce the EMIB-T package in 2027, AWS is also testing EMIB technology. Although eMIB-T and CoWoS-L are competitive solutions, considering technical maturity and yield, the latter will still be the mainstream packaging technology used in AI chips until 2028.