“By the end of the decade, GPU fleets numbering in the hundreds of millions will be able to run billions of superintelligent agents that can think orders of magnitude faster than humans.”