Hi Michael,
Scalability depends on how the framework is set up from the beginning and the quality of the code and the people working on the project. It might even be bad luck that the PHP ELO's I've seen were set up in such a way that when things got serious they weren't manageable any more.
The reason we chose for Java was the object orientation though, the principle of technical objects met the learning object philosophy wonderly well that I expect that Java, if used in the right way, will keep paying off.
The quote about the scalability comes from an article I wrote for the Polish magazine E-mentor. I've got extensive experience through my active involvement in implementing dozens of ELO's over the last 10 years, including PHP-based ELO's. Also, I'm board member of the Didactor Foundation, a Java based open source ELO and I do a lot of lectures about Open Source in education.
The remark about the scalability is not so much of a technical remark as well as a remark from an organisational point of view. PHP is easy to start of with, Java (and .NET) are typically not. There is a reason for this: the logicality of Java is harder to understand because it requires knowledge of objects and their relations. Once you've got a learning environment based on Java going, it's easier to make adjustments, add new components or, for instance, integrate IMS-ld functionality.
Why? because IMS-ld, as well as other new learning approaches, are based on the thought of objects (like knowledge, people etc), relations between objects, combined with a certain workflow. It's my experience that in this sense Java is more scalable because it's based on the relation metaphor.
This doesn't mean at all that PHP is not scalable, but looking to the future, I expect that it's not as flexible as it should be in order to adopt to the continuous changes based on didactic experiences.