At ITA Software, we use SBCL for one large project, and CCL for one even larger project. One constraint at ITA is that we are committed to using open source implementations. We switched from CMUCL to SBCL somewhat before I joined the company (three years ago). These two implementations are closely related, but SBCL has several improvements. We use Linux exclusively.
The reasons for using SBCL on one product and CCL on the other are rather complex and involve things like differences of opinion within ITA, which I'd rather not talk about in public. Both implementations are excellent. One important advantage of CCL is that the compiler runs very quickly, which is very helpful for compile-edit-debug cycles that cannot be done incrementally, including running complete system tests. Also, we currently get very good information from the debugger for CCL, whereas when I was last using SBCL, it was not as good (I don't know how much it might have been improved since then).
We also feel good about the fact that CCL is maintained by an actual company with full-time employees who are dedicated to improving the product. SBCL is maintained by an informal group of maintainers, who work on SBCL on their own time, or when they are engaged as consultants. Let me emphasize that the SBCL maintainers include some truly excellent developers, and I don't inten to be disrespectful to them.
The product that uses SBCL (QPX) is very compute-intensive, and we believe that SBCL's compiler generates faster code for this application. Actually, the standard Gabriel micro-benchmarks don't say that either of these implementations is across-the-board faster than the other. Micro-benchmarks have many limitations, and can't be relied upon to predict the relative performance of real applications.