Amazon intends to make its Trainium AI chips available outside Amazon Web Services (AWS) in late 2027, CEO Andy Jassy said on October 29, 2026. The plan follows customer requests, according to Amazon’s statement.
What Amazon says about demand
Amazon reported that revenue commitments for Trainium and its Graviton processor family had each grown more than 20 times year over year. It also said short-term capacity for prior-generation Trainium instances sold at a premium compared with the first half of 2026. An AWS instance is computing capacity delivered through the cloud; the reported premium concerned that service, not a physical chip price.
Separately, Amazon said customers using its custom silicon saved more than $2 billion in Q3 2026. That figure covers Amazon’s custom silicon overall, not Trainium alone.
Why Amazon may keep chips in AWS
Jassy described a trade-off between selling chips and using them to power AWS services. An external sale can bring revenue sooner, when a chip leaves inventory, he said. Deploying chips in AWS, including for EC2 instances or Bedrock, can generate revenue and free cash flow over five to six years of use. Jassy said Amazon would probably reconsider the balance each year.
The AWS route also depends on getting infrastructure into service. NeoTeo previously examined the timing risks around AWS server-rack deployment.
What TorchNeuron does for PyTorch developers
Amazon says its open-source software TorchNeuron lets developers run existing PyTorch code on Trainium without rewriting it. That gives developers a way to use the framework with Amazon’s chip family, alongside the company’s plan to make Trainium available beyond AWS.
