As someone with more than a little bit of PLs: That's the wrong approach. Doing PLs is hard it's the ultimate naysayer's profession. Modern programming languages are built around a dozen or so major design decisions, or dimensions for the machine learning crowd, (that's why there are so many) - and many combinations, if not most of them, are simply incompatible. Like, you-end-up-with-undecidable-problems incompatible.

So doing a PL should focus solely on expressing computations. A different platform, even a very different platform like the cloud, does hardly justify the effort necessary to design and implement a language. In the end you won't do neither a good language nor good platform support.

Instead, develop a dialect using as much of the parent language as possible and only restrict the parts that are indeed incompatible with your goal.