I suspect part of the reason is the long time frames for dev of this tech. I suspect it is at least 2 years for this to see the light of day. That is forever in this space.
Intel failed with Larrabee and itanium. Maybe this will go better?
How many machine learning strategies is Intel going to try? Does it even know what it's doing? Spending billions of dollars left and right on totally different machine learning technologies kind of looks like it doesn't, and it's just hoping it will get lucky with one of them.
And even if you think that's not a terrible strategy to "see what works", there's still the issue that they need to have great software support for all of these platforms if they want developer adoption. The more different machine learning strategies it adopts, the harder that's going to be for Intel to achieve.
But I bet that branching instructions (various variants of search) still play a big role when you go beyond classifiers to reinforcement learning etc. so there is need for other architectures beyond GPUs.
I do have high hopes for their memristor initiative, but that's got to be years out.
The GPU move is smart for Intel.
The later is further proof that if you sell to dinosaurs, you won't survive the Big One (in this case, x86 growing up courtesy of AMD and Arm spending 15 years washing the footing underneath both by moving personal computing to mobile devices). This should be a big warning sign to the OpenPower guys. You need to start small and scale/price up, not the other way around.
Maybe they are building upon AMD's core tech based on that other licensing deal? If so I would bet on them succeeding.