We introduce a two-step blocking and regularization approach for the estimation of high-dimensional covariances using high-frequency data. In a first step, we group assets according to their observation frequencies and then estimate the covariance matrix in a block-wise manner using realised kenels with group-specific time scales. In a second step, the covariance matrix is regularized using random matrix theory. The resulting “RnB” estimator is positive-definite, well-conditioned and makes more efficient use of the data than the standard multi-variate realized kernel estimator of Barndorff-Nielsen, Hansen, Lunde, and Shephard (2008a). The performance of the new estimator is analyzed in an extensive simulation study mimicking the liquidity and market microstructure features of the S&P 1500 universe. The regularization and blocking procedure yields significant efficiency gains for varying liquidity settings, noise-to-signal ratios, and dimensions. An empirical application to the forecasting of daily covariances of the S&P 500 index confirms the simulation results.