You can run multiple inferences in parallel on the same set of weights, that's what batching is. Given enough parallelization it can be almost entirely compute-limited, at least for small context (max ~10GB per request apparently, but that's for 1M tokens!)