Couldn't you get at least similar storage performance from EBS Provisioned IOPS SSDs in AWS? And maybe "scale out," instead of "scaling up," to maintain performance within AWS constraints? The advantage obviously is that you avoid the capex and if you don't achieve the sales that you foretasted then you can just rapidly scale down your AWS footprint, thus avoiding financial catastrophe. You could spring for the fancy custom bare metal hardware when you reach a "hyperscale" amount of workload where it makes sense to switch to your own hardware, like Google or Facebook. And if your customers really want the fancy Infiniband clusters right now, why not license your software and make them put up the capital to buy the hardware themselves? Just my two cents, maybe I'm missing something and your approach makes sense.