That's...probably not true at this moment. They're using AMD GPUs, which generally support the sorts of workloads folks need, and AFAIK they haven't opened up support for AMD GPU spot instances to users.
The OpenCL ML ecosystem is really lacking at the moment.