At the 2026 OCP APAC Summit, debate over networking architectures in cloud AI data centers sharpened as industry players said the main bottleneck is no longer compute; beyond memory, networking has become a key constraint on how far AI infrastructure can scale.
The technical feasibility of co-packaged optics (CPO) has been demonstrated repeatedly at major industry conferences in recent years. At the 2026 OCP APAC Summit, however, Ayar Labs CEO Mark Wade shifted the focus from whether CPO works to whether it can scale commercially.
Qisda Chairman Peter Chen said infrastructure is the foundation that must first be built as AI compute demand surges. He said Qisda will not be absent from the fast-growing race, as the company steps up work on AI servers, data centers, 1.6T switches, and silicon photonics (SiPh).
ChipMOS said strong demand for automotive panels and memory products drove revenue and profit to record highs in the second quarter of 2026. Chairman Shih-Jye Cheng said the company will raise capital expenditure (capex) for two consecutive years, in 2026 and 2027, to support memory, advanced logic, and new product demand. The company plans to boost capex to more than 25% of revenue as it begins a new expansion cycle.
Global AI computing is rapidly moving beyond single-GPU systems toward AI Factory architectures, where agentic AI and inference workloads are pushing data centers to rely on networking, optics, packaging, cooling, power delivery and software co-design rather than raw accelerator counts. In a DIGITIMES interview, Nvidia Senior Vice President of Networking Gilad Shainer said the real unit of computing is now the entire data center, not an individual server.
