“By the time AGI is achieved, inference GPU fleets will be large enough to run millions of AGI instances, equivalent to 100 million human researchers.”