Anthropic says AI cut AMD chip migration to three days
Anthropic says it has reduced its dependence on Nvidia by training and running Claude across four chip families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Co-founder and Chief Compute Officer Tom Brown said the company treats capacity from these platforms as one interchangeable pool for research, model development, customer demand and internal use.
The company’s main test involved AMD GPUs. Anthropic says its frontier model Mythos ported or wrote nearly 100 percent of the required kernels, allowing Claude to run on AMD hardware in three days with almost no human intervention. The comparison is with the six to twelve months that conventional engineering approaches can require; the figures concern prompt processing on AMD MI355X GPUs.
The claims come from an Anthropic investor presentation and have not been independently verified. If accurate, the results could reduce the software advantage associated with Nvidia’s CUDA ecosystem and make it easier to select chips by price, availability or workload. Anthropic also describes research-based compute multipliers and its Anthropic Direct data-center program as ways to improve efficiency and lower ownership costs.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.