This is nice gesture, but in my experience, especially for medium/large experiments, it almost definitely won't be applicable to analysis software, private datasets/raw data/metadata, and more specifically middleware. It doesn't really help with reproducible results in the cases which might be most interesting. For medium to big data experiments, most source code is often useless without the (often custom or proprietary) middleware used for processing the data. For very large experiments, there is likely no canonical dataset and all "datasets" are actually generated on-demand (at least those might be covered under this).

For software, and especially middleware, some universities will be very reluctant to give that up, especially because many times the majority of the software/middleware isn't actually funded through the NSF, but maybe through a computing division, foreign institution, or international collaboration, for example. At the very least, they'll likely want to assert a restrictive copyright on the license and likely even attempt to patent some of the methods (a la Google). Researchers themselves may be reluctant to give up code: Some (maybe most?) universities have profit sharing too.

At the very least, however, it is a policy researchers can point to.