Normal
Nvidia sell the dies. AIBs buy the RAM and other components, design and make the heatsinks (or make from design for Ref/FE), design and make the PCBs (or make from design for Ref/FE cards), and assembly (well, AIBs contract the OEMs who do the actual hand-on manufacture). As with every GPU, price-per-die drops throughout production as setup costs are amortised away and process yields improve through experience, with per-die costs at the end of life tending towards pure production cost.If AIBs estimated demand for 10k cards, order 10k dies form Nvidia, then decide "we don't want 10k dies now", then whether they are still stuck with 10k dies to sell depends on whether the purchase order they signed included a get-out clause. Likely not because Nvidia do not want to be stuck with extra dies because an AIB partner got their numbers wrong.
Nvidia sell the dies. AIBs buy the RAM and other components, design and make the heatsinks (or make from design for Ref/FE), design and make the PCBs (or make from design for Ref/FE cards), and assembly (well, AIBs contract the OEMs who do the actual hand-on manufacture). As with every GPU, price-per-die drops throughout production as setup costs are amortised away and process yields improve through experience, with per-die costs at the end of life tending towards pure production cost.
If AIBs estimated demand for 10k cards, order 10k dies form Nvidia, then decide "we don't want 10k dies now", then whether they are still stuck with 10k dies to sell depends on whether the purchase order they signed included a get-out clause. Likely not because Nvidia do not want to be stuck with extra dies because an AIB partner got their numbers wrong.
