A Chinese startup developed an AI accelerator on a 14nm process, asserting it achieves performance comparable to 4nm-class chips for specific inference workloads, though it lags in training capabilities. The claim remains unproven, with the underlying economic rationale being the primary focus of the analysis.
A 14nm chip achieving 4nm-class inference performance, if validated, would significantly alter compute cost curves and reduce reliance on leading-edge fabs.