Brain Machine Interfaces (BMIs) are emerging as an important research area in clinical therapy. A large range of potential BMI control signals can be found in the brain. In increasing order of volume of brain tissue being sampled, these signal includes recordings of electric discharges from multi unit activity (MUA), summed population activity of thousands of neurons via local field potentials (LFPs), and electrical activity recorded from either the surface of the brain via electrocorticograms (ECoGs) or the surface of the scalp via electroencephalograms (EEGs). While each of these signals have been studied separately, it has been difficult to compare the potential that each signal has for general prosthetic control across studies. Information theory has been proposed as an abstract measurement to bridge this gap, however the maximum information rates of any experiment is limited by the parameters defined by that experiment (e.g. inter-trial interval length, number of targets). Here we propose a different measure of information, which we call information capacity, which measures the maximum possible information rate that a signal can provide. An advantage of measuring information capacity is that it can readily be compared between different signals and different tasks. We show how to calculate information capacity making linear Gaussian assumptions, and we discuss more general possibilities. We present a case study involving a rat BMI task involving either MUA or LFP signals.