Always love Patrick's posts. Lambda has built a business on moving people from the cloud to on-prem and reducing the cost of cloud AI infrastructure. As some of the other comments have pointed out, for machine learning applications with GPUs, it almost always makes sense to run your workloads on prem, STH is showing that it can also true for traditional CPU workloads.
For example, we're able to provide GPU instances (https://lambdalabs.com/service/gpu-cloud) that are half the hourly cost of AWS hourly on-demand pricing. How? Because there are huge markups on clouds services.
We've done such extensive benchmarking and TCO analysis and the jury is out: it's simply less expensive to run on-prem. You're just paying for convenience when using a cloud. GPU or otherwise.
For example, we're able to provide GPU instances (https://lambdalabs.com/service/gpu-cloud) that are half the hourly cost of AWS hourly on-demand pricing. How? Because there are huge markups on clouds services.
We've done such extensive benchmarking and TCO analysis and the jury is out: it's simply less expensive to run on-prem. You're just paying for convenience when using a cloud. GPU or otherwise.
Sources:
- https://lambdalabs.com/gpu-benchmarks
- https://lambdalabs.com/blog/hyperplane-16-infiniband-cluster...