Maybe so. Orly has a somewhat leveldb-like component called a Repo which might slot in well. A Repo is a log-structured merge storage system that also provides indexing, access to previous values, and, most importantly, consistent reads and writes, even across multiple Repos. It also makes good use of resources (RAM, SSD, HDD) to optimize performance and cost. (Also working on letting it run on GPUs, for insanely high performance for people with the power and air conditioning budgets.)

I've seen a couple of mentions of graph dbs on GPUs in this thread, and it does seem like a (somewhat) obvious fit. Anyone aware of any projects that make (good) use of that right now? Not necessarily a stand-alone service, but also things like an embedded graph db ("berkley db for graphs") or something like that?

BigData by systap is working on this combination see this blog post (http://blog.bigdata.com/?p=658)

We would clearly need to do a deep dive but since Orly was designed with modules in mind it should be feasible to use the storage engine underneath Cayley.