"One VM to rule them all" reminds me of Parrot VM (for Perl 6). Looks like development slowed down a year or two ago. Back then, it got me excited as it looked like the "open source community" alternative to "Big Corporate Java".
(I'm having doubts if this comment belongs here. But it's my genuine reaction, for what it's worth.)
LLVM is providing most of the interop without strings attached.
I don't know a single person that's said a single good thing about Oracle, including people that work there. I know a ton of people that have been burned by Oracle and their lawyers.
Ref https://github.com/oracle/graal#license, appears that the parts you would want to distribute have the classpath exception which allows linking without messing with the license of your code (same as the JVM). The FAQ also covers this: http://www.graalvm.org/docs/faq/.
And that's just the top-level differences. It gets worse as you get into the details of API style and the costs of the indirection layers to convert into the local style, assuming conversion is even technically possible.
Even if you write with the lowest possible common denominator of C or Rust with no runtime, you'll still encounter a significant subset of those issues. And of course we still will use such things; all serious languages can communicate in some manner with a C implementation of a library. (Not for any theoretical reasons, really, C has more opinions than people often realize, but because it's still the baseline current systems are built on.) But it can't be as good as a native implementation.
Heck you will encounter those issues in a large project using only one language!
A lot of language specific libraries incorporate foreign language (often C) libraries and add additional functionality thst fits the ergonomics of the target language, rather than being a thin wrapper around a language neutral library. This is more than aesthetic when the target language has, for example, differences in type system or fundamental guarantees from the underlying library's language.
OTOH, sometimes there aren't C (etc.) libraries, and it's more straightforward to implement a library in (say) Python. But once you have a Python library, calling into it from Rust, C, Erlang, or even a broadly similar language like Ruby is, to put it mildly, non-trivial.
You mention Python that they say has an easy FFI with C. That's nice but it's usually used for the performance more than for reusing code.
The concept lives on in the form of two "new" Perl 6 projects, namely MoarVM (started in 2012; a multi language VM) and NQP (started around 2008; a virtual VM and multi language compiler toolkit).
These can appropriately be compared with Graal and Truffle. Imo their main weakness is they're not backed by Oracle but rather the Perl 6 project.
I think MoarVM/NQP are strikingly similar in their approach based on what I've learned so far about Graal/Truffle (which isn't much; perhaps a few hours of exploration; eg reading/watching the One VM to Rule Them All, One VM to Bind Them paper/video). 
Given the large number of Oracle employees working on Graal/Truffle (40 in 2016 according to the video) it doesn't surprise me that it's currently a lot faster.
Did you use Graal/Truffle friendly benchmarks (just as was done in the video)? Have you tried MoarVM friendly benchmarks to see where MoarVM shines? In particular, have you accounted for the new JIT that recently landed after 3 years of non-master branch development? 
Would you please elaborate a little on how MoarVM is insufficiently flexible?
TIA for any reply.
Does MoarVM use a meta-compilation technique? Pypy for example uses meta-tracing and Truffle uses Partial Evaluation to get from an interpreter specification to dynamically compiled code without duplicating the language logic. If yes, could you please point me to further information?
Meta-compilation is the primary innovation in Truffle. It allows to compose languages together and to compile them as one unit.
If you want to know more about the Truffle approach to meta-compilation see this PLDI paper from 2017.
Its NQP that does metacompilation, not MoarVM (afaik).
I linked to NQP's (spartan) readme above.
Some other points/links that might be helpful:
* nqp is written in nqp.
* nqp has a first class grammar construct. See wikipedia's Perl 6 Rules page. (nqp's grammar is defined using nqp's grammar construct. So is Perl 6's.)
* nqp and MoarVM implement a protocol for abstracting behavior (eg methods) and representation (memory layout and use) called 6model.
To try set your expectations appropriately, let me emphasize this note from the NQP readme: "NOTE: there's no end-user support for NQP and the behaviour can change without notice. It's a tool for writing Perl 6 compilers, not a low-level module for Perl 6 programmers."
NQP is a tool for writing compilers, not just Perl 6 compilers, and is in fact much more than that, but the Perl 6 project is currently very focused on the Perl 6 language and definitely doesn't want ordinary Perl 6 programmers thinking NQP is a tool for them.
Doc is sparse, again reflecting the focus on the Perl 6 language and the Rakudo compiler that builds atop NQP, analogous to your team focusing on just one of your guest languages.
The best doc for getting into NQP is not even in the NQP repo. It's the support materials for a 2-day workshop held in 2013. The abstract is "This intensive 2-day workshop takes a deep dive into many areas of the Rakudo Perl 6 and NQP internals, mostly focusing on the backend-agnostic parts but with some coverage of the JVM and future MoarVM backends also. During the course, participants will build their own small compiler, complete with a simple class-based object system, to help them understand how the toolchain works.".
Beyond that, there is a doc directory in the nqp repo.
Please write an update here on what you find out.
(nqp rules are Perl 6 Rules grammars but the general purpose DSL (!) for writing procedural code embedded in an nqp grammar is nqp's general purpose DSL rather than Perl 6's general purpose DSL.)
(MoarVM directly implements 6model -- MoarVM is named for "Metamodel On A Runtime". 6model is how one nqp based language is able to do magic like subclassing a class written in another language (whether nqp based or via an existing compiler) despite them having distinct OO semantics and implementation strategies.)
There is a community edition (CE) and an enterprise edition (EE) of GraalVM. The community edition is distributed under an open source license. It is free to use in production and comes with no strings attached, but also no guarantees or support. The enterprise edition is available from the Oracle Technology Network under an evaluation license. It provides improved performance and security for production deployments. If you are interested in using the enterprise edition in production, please contact firstname.lastname@example.org.
Except a VM with a more liberal license could easily beat it in popularity.
These licenses prevent Graal VM from being used inside e.g. OSes and browsers with a more liberal license.
The fact that some people will use the code without giving back is secondary to that.
That is not just my opinion, because if a more liberal library comes into existence (inevitable), it will prevail.
And on that day you will miss when GNU/Linux mattered.
But hey, we will always have freeware and public domain.