Reaxis AI requalifies the stranded GPU fleets and re-contracts them as certified AI inference capacity, pointed at decode, the memory-bandwidth-bound half of inference that old silicon is still genuinely good at.