> The pre-trained models took an immense amounts of compute resources to build.

Oh definitely, but nobody is serving models from the same machine + process that they used to train them right? And solutions like ONNX exist (although TF and PyTorch’s support is inconsistent at best)

Additionally, those who built the models weren't necessarily developers and Python is an low-barrier to entry language.

It just feels like an engineering anti-pattern to build “down” to this level, instead of skilling people up, or standardising on some standard model serialisation and serving format, model serving tools exist, and they’re often written in faster/more optimised languages, so at that point, why bother with Python after actual model training at all.