Amazon plans to deploy 1 million Nvidia GPUs by end of 2027, but is increasingly shifting to its own Trainium chips as custom AI accelerators gain traction

Amazon and Nvidia agreed during the previous quarter that Amazon will take delivery of 1 million Nvidia GPUs by the end of 2027, along with other Nvidia chips and networking equipment. This deal is projected to generate tens of billions of dollars in revenue over the next two years for Nvidia.

Amazon is standing up more Trainium servers (its own custom AI accelerator chip) than Nvidia servers and reports a $225 billion backlog for Trainium compute alone. CEO Andy Jassy stated that using its own AI accelerators instead of Nvidia GPUs will save the company tens of billions of dollars in capital expenditures.

Nvidia's market dominance in AI accelerators is eroding as cloud providers develop competing custom chips. Alphabet has seen strong demand for its TPUs and is selling them to select customers, while Microsoft intends to use its own Maia chips exclusively to expand capacity and control costs.