Thermostatically controlled loads (TCLs) have promising regulation potential and are regarded as valuable flexible resources. Due to the relatively lower operating power, TCLs are always aggregated to participate in various regulation programs. However, the intrinsic response delays of aggregation always result in relatively slower speed and limit further implementation in the fast frequency regulation (FFR) service, which is urgently needed by renewable power systems without sufficient inertias. Therefore, a non-aggregate method based on edge intelligent terminal (EIT) is proposed to utilize TCLs to provide FFR without unacceptable rebound effects. Firstly, the FFR infrastructure, consisting of the distributed frequency response and centralized recovery optimization, is developed based on functions of the EIT, such as local measurement, edge computation, and flexible control. Then, the equivalent mirror recovery strategy is proposed to mitigate power fluctuations and reduce occupations of other regulation resources. On this basis, the optimal recovery problem of TCLs is formulated to achieve flexible power control without violations of FFR requirements. To overcome the curse of dimensionality, the dynamic clustering method is proposed to divide non-aggregate TCLs into tens of groups, thereby the optimization problem is solved by EITs separately and the solution speed is independent of system scales. Finally, the proposed FFR is validated based on several simulation cases, and results show that TCLs significantly contribute to enhancing frequency resilience for low-inertia power systems.