Now, the current leaders in compute are AWS, GCP, Azure and so forth.
AWS makes a yearly revenue of roughly 40 Billion based on projected 2020 revenue while GCP makes a revenue of 10 Billion yearly. Now a lost of costs are associated with hosting servers for other companies. However, some of these computers may be equipped with GPU’s to train models for customers.
Some, such as Google’s TPU offer up to 100+ petaflops of performance. Now Tesla comes in with over an exaflop, equivalent to 1000+ petaflops. As an example, training a model can range in price and in the case of XLNet model, it cost $245,000 to train on Google’s TPU V3. A model that could potentially not work, but costs so much compute power. Now if Tesla could leverage this insane supercomputer to lease to other companies for training purposes, this could not only speed up training time, but also leverage existing resources for revenue streams. This seemingly harmless tweet of Musk could potentially turn into another profit puppy for Tesla.
