An A100 is still a place to serve a model.

A100 is still what a lot of clusters actually have. Lambda lists A100 40GB and A100 80GB. Amazon EC2 lists A100 40GB. Nebius’s row in the catalog does not. Your own server can join with whatever A100 is in it.

denv does not rent A100s and has not published a tokens-per-second result for them. Use the H100 write-ups as method, not as a forecast.

A 70B model at 16-bit does not fit in 40GB or 80GB of weights alone. At 8-bit the weights are on the order of 70GB before KV cache, which is the difference between a 40GB card and an 80GB card. That arithmetic is the rule of thumb in the inference guide. The recipe is what decides precision and GPU count. MIG on A100 is real hardware, and DRA is how Kubernetes asks for a slice. A recipe that was only tested on H100 is not an A100 recipe.

Repair and failover do not care that the card is a generation older. They care that the node reports faults and that another replica exists.