Faithful simulation of the Madden-Julian Oscillation (MJO) can significantly improve extended-range prediction, but the numerical models have difficulty in simulating the MJO. One of the main challenges is how to simulate the MJO propagation speed properly. In this study, we examine the propagation speed of MJO during boreal winter in CMIP6 models and reveal the key factors controlling the MJO propagation speed in models. The result shows that most of CMIP6 models can well simulate the MJO’s eastward propagation over Indo-Pacific warm pool, but the simulated MJOs in different models have diverse propagation speeds. It is shown that the Kelvin-wave response and Rossby-wave response to the MJO heating are the key circulation factors affecting the simulated MJO propagation speed, with stronger Kelvin-wave response and weaker Rossby-wave response corresponding to faster propagation. The variation of the MJO circulation structure among models is attributed to the diverse background sea surface temperature (SST). The models simulating faster MJO tend to have higher SST over Central Pacific (CP) and Western North Pacific (WNP). The warming over CP affects the MJO speed in two ways. First, this zonal expansion of the warm pool can increase the MJO horizontal scale, leading to stronger Kelvin-wave response that favors faster MJO propagation. Second, it increases the moisture content over CP, which weakens zonal moisture gradient over West Pacific, resulting in acceleration of the MJO through enhancing the zonal moisture advection. Different from previous studies, we find that the SST warming over WNP also affects the MJO speed. The SST warming over WNP leads to a more symmetrical SST distribution about the equator, which enhances the moisture content over Maritime Continent, leading to more equatorial symmetric distribution of the background specific humidity during boreal winter. This symmetric distribution of moisture is conducive to the development of Kelvin wave response. The results here not only have implications in understanding the MJO dynamics, but also indicate that realistic simulation of MJO requires fidelity in simulation of background states.