HPC Services Funding Models

As a followup to my post Researchers Need to Stop it with the Hardware, I note that I think the way HPC as a service is treated currently by granting agencies is bad.

Honestly the researchers and the granting agencies don't care. They just want to do the most they can.  So how would I construct the rules for an HPC service from funding agency on down?  Note that the following while not morally wrong does currently, (to the best of my understanding) violate IRS tax rules that are used to say what is capital and what pays indirects.

1. (Bad)Indirects pay for all HPC services. Who wouldn't love this. Admins are given a bucket of money build a service and researchers use it at no cost. Problems with this is, zero moderation between how much of an HPC service a given research group consumes. 
2. (Good)Grant money for HPC services are indirect free and the the fee for the service exposes the researcher to the entire price of the service.  That is admin, software, rack rent, hardware, etc.
3. (Best)Indirects pay things that are hard to measure, these would be public goods.  Admins and software are both hard to measure how much a user consumed. Power/cooling would also be in here if you can't measure it, or it isn't provided in a way that accurately reflects core usage. Items that are easy to measure are exposed to the researcher, this is a break down of a cost per core month of cpu time and similar items.

Note that in my above example I did not include disk. Disk is hard, I would love to have a guest post on how to organize services around storage for HPC data and funding that service. Leave a comment or tweet me @brockpalen if you have thoughts on this.