Disclosure: I work on Google Cloud. Not necessarily. The DGX-1, for example, has pretty poor perf/$$ but reduces the time a data scientist spends waiting. For

Disclosure: I work on Google Cloud. Not necessarily. The DGX-1, for example, has pretty poor perf/$$ but reduces the time a data scientist spends waiting. For some organizations, their people time is so valuable that what matters is “what gets me my answers back faster”, because that employee is easily $100/hr+. That’s actually why the 8xV100 with NVLINK is so attractive (and why the TPUs also have board to board networking, not just chip to chip).