With the advance of Artificial Neural Networks, GEMM has become a dominant arithmetic operation in compute-intensive applications. Multipliers contribute to a major portion of area and power consumed by conventional systems. This work proposes T-MAC to improve the power and area efficiency of machine learning accelerators by exploiting a new dimension of computation reuse empowered by Hybrid Unary-Binary computing paradigm. Leveraging the seamless binary interface and internal unary computing kernel, T-MAC achieves higher hardware efficiency and better scalability compared against conventional binary hardware.