DataCrunch is based in Finland and provides long time rental of bare metal servers, short time use of instances similar to EC2 and serverless container hosting. The latter is particularly interesting since they come with autoscaling and queue support out of the box. We have been using Fargate, but then you can’t go completely serverless with GPUs and the queue is a separate entity.
We deployed our first model using a vLLM docker image in days without having used the system before. We will probably moving existing model hosting from AWS to DataCrunch as well.
Agreed. At least we have ARM who might start to produce their own chips. We know how to design them, just not capable of actually producing them.