Information transfer through electromagnetic waves is vital to computing, imaging, telecommunications, and power management. A central figure of merit is the Shannon capacity, the maximum error-free bit-rate over a noisy channel achievable via encoding. Prior works on optical information transfer have treated waves propagating through fixed photonic structures such as vacuum, where the communication channels are predetermined and the Shannon capacity can be evaluated via a convex water-filling procedure, or relied on simplifications of relevant physics via circuit theory and far-field assumptions. In this article, we present limits on the significantly more difficult problem of maximizing Shannon capacity given the freedom to structure the photonic environment, allowing for the optimization of optical communication channels subject to wave propagation constraints. By combining information theory, wave scattering, and convex relaxations, we formalize the relationship between electromagnetic and information-theoretic quantities, showing how device considerations such as material choice, system size, and geometry impact optimal channel power allocations and maximum bit-rates. Our results generalize prior attempts to interpret optical communication using the spectral decomposition of the electromagnetic Green s function, and provide a foundation for furthering understanding of how to optimally engineer information-processing photonic devices such as antennas, MIMO space-division multiplexers, and metasurface imaging kernels.