While CUBLAS is available out there, it is not something that can be run on such a wide array of hardware. This is probably of great interest to startups wanting to do things on things on FPGA's and mobile devices. They now have a well optimized open source math library to use as potential building blocks, and it has an Apache License. CUBLAS is only good for NVidia cards. So I actually think this could be a big deal, and could potentially reduce the popularity of CUDA over time.
I sure hope it does. I have paid entirely too much for my predecessor's CUDA lessons through the premium NV charges for their cards.
Unfortunately, I'm pretty sure NV is entrenched at this point. My AMD card and my resolution to port everything I needed to use lasted about 3 months, and that was without any CUBLAS dependencies :(
