The $12M Cloud Arbitrage: Eliminating Enterprise Server Waste
Delivered live in London to enterprise CTOs, CFOs, and FinOps engineers, this keynote details the exact algorithmic discovery and elimination process that recovered twelve million dollars in recurring annual cloud spend at CAI. Arkajit presents the four foundational levers of cloud arbitrage: zombie asset identification, scheduled off-hours power management, compute instance rightsizing, and committed use financial engineering.
"Good morning. I want to begin with a simple financial reality that cloud providers spend millions of dollars in marketing trying to hide: the cloud is not cheaper than owning computers. The cloud is a convenience lease with variable pricing.
When software teams migrated from physical data centers to AWS and Azure over the past decade, finance departments rejoiced because capital expenditures (CapEx) became operating expenditures (OpEx). But what finance did not anticipate was that every software engineer with an AWS IAM key now had corporate purchasing authority. An engineer can spin up an r5.8xlarge instance on a Friday afternoon to run a quick benchmark, forget about it over the weekend, and run up thousands of dollars in server fees before anyone notices on next month's PDF invoice."
"When I took charge of the infrastructure governance review at CAI, our first task was not complex performance refactoring. It was simply finding zombie compute resources. What is a zombie resource? It is a virtual server, a database replica, or an attached storage volume that has had zero network input and zero CPU utilization above 1.5% for sixty consecutive days.
We discovered over 840 unattached Elastic Block Store (EBS) volumes. These were storage disks that had been unhooked when developers terminated temporary test servers three years earlier. But AWS does not automatically delete your storage disks when you shut down an instance unless you explicitly flag 'DeleteOnTermination'. Those detached volumes were sitting in us-east-1 accumulating charges twenty-four hours a day. We deleted them and saved 4.2 million dollars annually in our first sixty days."
"Why do staging and quality assurance environments run at 3:00 AM on a Sunday? There are 168 hours in a week. Your engineering team works approximately 45 hours in a standard work week. That means for 123 hours out of every 168 hours - almost 73% of total time - your non-production test servers are idling in empty office hours.
We wrote automated cron schedules that hooked into the AWS EventBridge and Lambda APIs. At 7:00 PM local time every evening, every EC2 instance tagged 'Environment=NonProduction' was automatically stopped. At 7:00 AM every Monday through Friday, the instances were powered back on. That single behavioral change eliminated 70% of our compute bill across all developer clusters. Engineers were delighted because their machines were fresh every morning, and the CFO saw instant bottom-line margin expansion."
"The next step was rightsizing. We monitored peak memory and CPU usage across our production database fleet. 80% of our database nodes had been provisioned for 128 GB of RAM because an architect three years prior was worried about a holiday traffic spike that never occurred. Their actual peak memory utilization never exceeded 22 GB.
We migrated those instances to appropriately sized generations and negotiated 3-year Savings Plans and Reserved Instances for our steady-state baselines. That accounted for the remaining 4.5 million dollars in recurring savings, bringing the verified total to 12.1 million dollars."
"To conclude: FinOps is not about being cheap. It is about capital discipline. Every dollar an engineering organization stops wasting on idle cloud servers is a dollar that can be reinvested into hiring top engineers, funding university scholarships, or lowering software subscription prices for customers. Stop letting convenience replace arithmetic. Measure your servers, turn off the lights when you leave, and build lean systems."