Movies aren't made by software, they're made by massive teams of people who share overlapping expertise.

Do you want to write Lua for Kerbal Space Program, or Perl for NASA?

At studios I've been at, we did tests with various renderers. Sure, we could have rendered a film in Cycles, but it would have taken much longer.

Let's say you want to render realistic skin using Cycles. You can't, the only BSSRDF profiles that Cycles offer are simple bicubic and gaussian profiles, these wouldn't work for rendering realistic skin. Cycles doesn't even offer a simple dipole. Renderman on the other hand offers a dipole, it offers the state of the Photon Beam Diffusion, it offers Pixar's recent approximate BSSRDF, which is almost as accurate as PBD and is as easy to compute as a gaussian or a bicubic profile. So you'd have to implement your own BSSRDF shader, which is only doable for large studios such as Weta which have their own R&D department.

So let's say you want to render hair. You can do this with Cycles, but their importance sampling code is not state of the art (last time I checked), which means that you can probably expect double the amount of noise in Cycles. You could still use this for rendering, but computing time is quite limited (especially if artists want to compute previews). You'd either have to implement your own hair shader, or you'd have to purchase a ton of extra hardware.

Let's say you want to render a scene with tons of triangles and textures, you probably won't be able to do that in Cycles. They don't use special tricks for quantizing those triangles in the memory. Their code for caching textures (sometimes multiple terabytes for a single scene), is also not as good as Renderman, especially when using the GPU (a GPU isn't very good at constantly streaming terabytes of data into it), which means that your texture date either needs to be limited, or you'd need to throw a lot of rendering machines at it.

Cycles can do most stuff that Renderman (or Arnold) can, but that's not the point. If you'd pay me to work on a renderer for you fulltime for a year, I could produce a feature complete renderer, but it won't be optimized. The code won't be optimized, but probably more importantly the ray intersection code and the importance sampling code won't be optimized, which means that renders will be slow and noisy. Pixar has a whole team working on optimizing their code and whole team of researchers working on improving the importance sampling. Cycles is made by hobbyists who are doing an excellent job (Cycles is an amazing renderer for amateur users), but it's just not in the same ballpark as Renderman or Arnold.

Cycles is an absolutely amazing renderer, but it just lacks a lot of stuff that you'd during production of feature movies. If you want to use Cycles for your own use I can absolutely recommend it, it will have everything you need and it's open source!

From your username, would you happen to be Marcos or someone else from Solid Angle? If so, big fan of Arnold. :)

This is %100 not true. Forward path tracing is not adequate for anything except for direct illumination. Any enclosed space that has most of it's illumination from some sort of bounce is not feasible without some sort of illumination caching. Bidirectional path tracing and VCM do not carry any addition complexity from an artist's point of view. Also avoiding SDS paths is not always something that artists can simply avoid, and any illumination that is missing is room for improvement.

Do you have links to Solid Angle's patents? Renderman actually samples emissive geometry exceptionally well.