AI infrastructure means new ways of looking at development. As training and inference are becoming more distinct, the mix of compute, power, connectivity and location will vary. Whilst training and inference may be using similar compute technology, the way in which they are run creates very different requirements. Training requires a large number of accelerators to be used to perform long running processes. These will require large clusters of accelerators to communicate frequently with consistent conditions. Inference will require applications to respond to users quickly, process data and return results in commercially acceptable timescales. This will mean considering where you can get land, how much power capacity you want and what infrastructure you can expect to recoup. This means that if you treat training and inference as interchangeable within your development strategy you may be making unnecessary capital investments.