At https://github.com/mapbox/rasterio/blob/master/rasterio/__init__.py#L191, a big GeoTIFF is created in RAM. Then at https://github.com/mapbox/rasterio/blob/master/rasterio/__init__.py#L199 that GeoTIFF is read into memory *again* so that it can be written to the Python file object. There will be two copies in memory. It's terribly inefficient, but I don't want to spend the time to optimize this case when I should be documenting the limitations instead.