Hmmm. I'd pay very careful attention to the fine print on that. The way their scheme works is that each node chooses a random time to start its transmission, which means that all their chipping sequences are hopefully desynchronised. Then the AP tries decoding from all possible starting points at once. If I'm understanding spread-spectrum correctly, each parallel receiver pulls out only the signal from the node (or nodes) whose chipping sequences are synchronized with it, while all the other node's signals appear as noise to it.

Which is the catch. Run too many nodes at once, and the signal-to-noise ratio will rise and your signals will start to fail to decode. What this generally means is that nodes at the edge of the network - the ones with the lowest SNR budget - start to drop out first. I'd be interested to see what their link budget and range is with 1600 devices on the same node, or conversely how many devices they can fit on a node at the advertised link budget and range. (Also, just how well they cope with a mixture of near and far nodes.)