ONNX is a format that allows you to run AI models without Python in any language that implements ONNX, there's even an ONNX implementation in Go, meaning you can churn out even more performance out of AI models and waste drastically less resources (Go, Rust, C++, Zig, C, D etc could be used to squeeze performance). Think of it how Java produces a JAR file, well an ONNX file is a file that could be run by any runtime built for it. Another reasonable analogy would be WebAssembly, but to a degree.