The CAPM is by no means a failure. While not a normative theory like MPT, but a positive one describing a capital market equilibrium under conditions of risk, the CAPM, if applied carefully, still provides a reasonable basis for methods to value uncertain cash flows, e.g. in capital budgeting. Further, it is at the core of models for derivatives pricing - in particular in cases where perfect continuous replication is not possible.
The failure consists rather in "testing" the CAPM with realized returns.
The CAPM is not meant to forecast that a high beta fund will perform better than a low-beta fund on average in the long term. Neither is it supposed to explain the past performance of these funds.
The CAPM describes relationships between expectations.
It is amazing that still so much effort has been spent on trying to test it with data about returns in the past.
Imagine you want to test what people are expecting today about the performance of GE vs. the Dow over the coming twelve months. So do you call PMs and ask? Do an online survey? Go through all pulished research you can find to try to see a consensus view? Do you analyze current mutual fund and portfolio holdings?
All these seem justifiable methods to assess expectations.
But instead your "test design" is to go on a holiday, come back in a year, and check what the Dow and GE have actually done...???
In principle, and simple terms, the famous Fama-French study tested if historical betas measured with realized returns in the past can explain differences between realized returns in the more recent past... While this was a thorough and valid test of historical relationships between realized returns, this test and all that followed, had and have nothing to do with expectations.
Again, the CAPM describes relationships between expectations. There is no link to realized returns. On the contrary. The CAPM describes a market in equilibrium, at a point in time. In equilibrium, there is no trading. At equilibrium prices, supply and demand are balanced. The CAPM is the Capital Asset PRICING Model. Not the Capital Asset Return Generation Model. There is no explanation at all in the CAPM for anything having any kind of realized return.
Sharpe seems to be a very diplomatic guy. He was relatively outspoken, though, about the distinction between the CAPM and models for the "return-generating process" in his speech when he got the Rijksbank prize for economics in memory of Alfred Nobel, which is downloadable at the Rijksbank web site (the speech, not the prize...).
A few years ago, Markowitz wrote a text called "The "Great Confusion" concerning MPT" - arguing against the common view that mean variance portfolio selection would require normally-distributed returns. One may or may not find arguments to disagree with Markowitz, but if that misunderstanding is as great as the title suggests - it still is dwarfed by the CAPM confusion.