We ran the compute numbers today, and our internal “large VM” came out to about $36 a month. The comparable Amazon price in our model was $260.
That was a striking gap, but it was an early compute comparison, not the final answer to private cloud versus AWS.
The number reflected how much more efficient our infrastructure had become over the previous year. We had more than halved our compute cost and, by this estimate, roughly doubled efficiency. That was meaningful progress toward slowing the capital growth we had planned for at the start of the year.
The next email in the thread immediately named the missing pieces: power, cooling, top-of-rack switching, and more complete personnel costs. A server does not become a virtual machine by itself. Somebody buys it, installs it, networks it, operates the cluster, replaces failed parts, and maintains the software that lets the rest of engineering request capacity.
I still wanted the comparison.
Cloud economics can become a religious argument when people compare a visible hourly price on one side with an incomplete internal budget on the other. The better approach is to make the model explicit, improve it, and keep using it to challenge decisions.
AWS carried costs that did not appear in our direct compute number, but it also provided flexibility and services we would otherwise own. Our internal platform had a lower modeled unit cost because Shutterstock ran enough steady infrastructure to use hardware heavily and had already invested in the team and automation around it.
The correct answer did not have to be all private or all public. We were already connecting physical environments with AWS when extra capacity made sense. The model helped us decide which workloads belonged where.
I called the result “very good progress,” which is still how I read it. Thirty-six dollars was not a universal truth or a promise that every hidden cost had been captured. It was evidence that the work on utilization, automation, and platform design was bending the curve in the direction we intended.
More comprehensive data still needed to come. The first number was good enough to make that next pass worth doing.