We analyze the performance of massive MIMO systems with $N$-antenna users. The benefit is that $N$ streams can be multiplexed per user, at the price of increasing the channel estimation overhead linearly with $N$. Uplink and downlink spectral efficiency (SE) expressions are derived for any $N$, and these are achievable using estimated channels and per-user-basis MMSE-SIC detectors. Large-system approximations of the SEs are obtained. This analysis shows that MMSE-SIC has similar asymptotic SE as linear MMSE detectors, indicating that the SE increase from having multi-antenna users can be harvested using linear detectors. We generalize the power scaling laws for massive MIMO to handle arbitrary $N$, and show that one can reduce the multiplication of the pilot power and payload power as $\frac{1}{M}$ where $M$ is the number of BS antennas, and still notably increase the SE with $M$ before reaching a non-zero asymptotic limit. Simulations testify our analysis and show that the SE increases with $N$. We also note that the same improvement can be achieved by serving $N$ times more single-antenna users instead, thus the additional user antennas are particular beneficial for SE enhancement when there are few active users in the system.