Amazon Says Switching AI Chips Is Easy. Its Checkbook Disagrees

Amazon Says Switching AI Chips Is Easy. Its Checkbook Disagrees

AWS sells the move off Nvidia as a one-line code change. A roughly $7 billion price tag on Decart says the real switching cost is far bigger, and that it never lived in the chip.

Amazon is in talks to buy Decart, a three-year-old AI startup founded by Israeli engineers, for around $7 billion, The Wall Street Journal reported on Sunday. Decart’s software makes AI models run faster across Nvidia GPUs, Google TPUs and Amazon’s own Trainium chips, which cuts the cost of moving a workload off Nvidia. In 2015 Amazon paid a reported $350 million for Annapurna Labs, the team that now designs Trainium. Same headcount, roughly 20 times the price: the software that helps customers use Amazon’s chip is valued far above the team that built it.

Amazon’s Trainium product page makes leaving Nvidia sound like fixing a typo. Change one line of PyTorch, point the model at .to('neuron'), and you are running on Amazon silicon. The page says it in eleven words: “Switch from GPU to Trainium by…

https://businessmodelanalyst.com/amazon-decart-acquisition-annapurna-switching-cost/