By studying the intrinsic polyhedral structure of multistage stochastic linear problems (MSLP), we show that a MSLP with an arbitrary cost distribution is equivalent to a MSLP on a finite scenario tree. More precisely, we show that the expected cost-to-go function, at a given stage, is affine on each cell of a chamber complex i.e., on the common refinement of the complexes obtained by projecting the faces of a polyhedron. This chamber complex is independent of the cost distribution. Furthermore, we examine several important special cases of random cost distributions, exponential on a polyhedral cone, or uniform on a polytope, and obtain an explicit description of the supporting hyperplanes of the cost-to-go function, in terms of certain valuations attached to the cones of a normal fan. This leads to fixed-parameter tractability results, showing that MSLP can be solved in polynomial time when the number of stages together with certain characteristic dimensions are fixed.