We propose a formal definition for (valid) speculative computations, which is independent of any implementation technique. By speculative computations we mean optimization mechanisms that rely on relaxing the flow of execution in a given program, and on guessing the values read from pointers in the memory. Our framework for formalizing these computations is the standard operational one that is used to describe the semantics of programming languages. In particular, we introduce speculation contexts, that generalize classical evaluation contexts, and allow us to deal with out of order computations. Regarding concurrent programs, we show that the standard DRF guarantee, asserting that data race free programs are correctly implemented in a relaxed semantics, fails with speculative computations, but that a similar guarantee holds for programs that are free of data races in the speculative semantics.