C++ through CUDA is by far the most popular option. There is some support in other languages but the support and ecosystem is far from what exists for CUDA and c++.

Python via RAPIDS.ai . There first bc most data science community for prod + scale is in it. It feels like the early days Hadoop and Spark.

IMO golang and JS are both better technical fits (go for parallel concurrency and js for concurrency/V8/typed arrays/wasm), and we got close via Apache arrow libs, but will be a year or two more for them as a core supporter is needed and we had to stop the JS side after we wrote arrow. Python side is exploding so now just a matter of time.