One of the cool things about AWS Lambda is that you can use it as a short-term supercomputer with thousands of threads. We (Stanford/UCSD ExCamera) are using AWS Lambda to do massively parallel low-latency video encoding with 4,000+ concurrent threads. Our colleagues at U.C. Berkeley who built PyWren are doing machine learning on Lambda with similar numbers of threads. AWS has been encouraging to us for these kind of workloads, even if they aren't what Amazon initially envisioned.
So a hard limit of 400 (that can't be increased) would be a real bummer. I wonder if we can get around this by defining 20 identical functions and then invoking each one 400 times in parallel...