[londons_explore]

With nearly everyone using inference accelerators, the pool of hardware is no longer shared between training and use.