Portfolio optimalisation depends heavily on the estimation of the moments (and therefore has HUGE estimation uncertainty).
Even though it's useful for comparing and analysing different existing strategies, I think practitioners are moving more towards the usage of factor portfolios for the strategies themselves (e.g. Fama-French). Also because the exploitation of such anomalies have been proven to be quite persistent and relatively profitable.
To give you an example of the estimation uncertainty that goes together with portfolio optimalisation, a simple plug-in Markowitz regression on the S&P 500 (1997-2006) using the delta method yields a weight of w=0.5 and a standard error of 0.4! So the estimation says that we should invest half our portfolio in the risky asset (S&P 500), with a standard deviation of 0.4. You can imagine that a 95% confidence interval would range all the way from investing nothing in the market to investing everything. So what is really optimal? The estimation uncertainty can be improved using shrinkage methods but you get the picture.
Another disadvantage of such portfolio optimalisation is that it doesn't really capture anomalies like factor models can. It's of course possible to combine the optimalisation of weights with the results from factor models to sort of 'get the best of both worlds'.
Besides, most portfolio optimalisation models can't beat the 1/N portfolio out of sample. Even if they do beat the benchmark, they often still have transaction costs which need to be corrected for. After correction, you will find again that the models are quite useless and you're better off investing in the 1/N portfolio instead. Like I said before, this is not true for cross-sectional anomalies. It has been shown numerous times that various factor models and predictive variables perform quite well. See for example Fama and French (2008), Campbell and Thompson (2008) and Goyal and Welch (2008).