This is partially speculation, the currently shipping version has 40GB of HBM2 @ 1.6TB/s. Hypothetically future versions could have 96GB at a faster speed.
Time of GPU ownership of independent Deep Learning researchers is over. It didn't even last a decade!