Right now we keep our paid customer models' in memory. Models in free accounts are just evaluated every time (these models are usually much smaller). For now we

Right now we keep our paid customer models' in memory. Models in free accounts are just evaluated every time (these models are usually much smaller). For now we want to avoid the complexity of distributed systems and just run this on very powerful machines with lots of RAM. For reference, one of our enterprise competitors has a calculation engine that can do 40B cells on a single host (using Java!).