Two time-invariant rateless code constructions are developed for efficient communication over multi-input multi- output (MIMO) Gaussian channels. Both architectures employ layering, dithering, and repetition as key ingredients, and convert the MIMO channel into a scalar channel to which classical Gaussian base codes can be applied. Both constructions are convolutionally structured - one is based on faster-than-Nyquist (ftN) signaling, while the other on a diagonal layering (DL) structure. Moreover, both employ successive cancellation decoding. We show that ftN rateless codes are asymptotically capacity achieving at any signal-to-noise ratio (SNR) and induce a time-invariant scalar channel. We also show that DL codes are capacity achieving at any SNR, and induce a particular time-varying scalar channel to which standard LDPC base codes can be applied without significantly sacrificing performance.