Notes from the team building C3X. Cost estimation for Terraform, the economics of cloud infrastructure, and how to ship FinOps tooling without a SaaS gate.
Every cluster carries a fixed overhead before a single pod runs: control plane, system nodes, load balancers, and monitoring agents. Running eight clusters at 15 percent utilisation is expensive in a specific, measurable way.
A load balancer with no targets and a public IP with no association each cost a few dollars to tens of dollars a month. Individually trivial, collectively thousands. Here is how to find them, verify they are dead, and remove them safely.
Autoscaling is sold as a cost control and often operates as a cost floor. Minimum counts set for comfort, thresholds that never trigger, and cooldowns that hold capacity long after demand fell all quietly set a bill you never chose.
Third party software bought through a cloud marketplace lands on the same invoice as compute but obeys different rules. It is rarely audited, often over-licensed, and frequently the largest line nobody in engineering owns.
Due diligence looks at gross margin, unit economics, and whether infrastructure spend scales with revenue. A cost cleanup before the process starts changes the numbers investors see. Here is what to fix and in what order.
When spend triples overnight, the priority is stopping the burn, not understanding it. Here is an hour by hour runbook: contain first, diagnose second, then decide what to reverse and what to keep.
Graviton4 instances (R8g, M8g, C8g) deliver meaningfully more performance per core than Graviton3, at a similar or slightly higher hourly price. Whether the newer generation is cheaper for your workload depends on how well it converts extra performance into fewer or smaller instances. Here is the comparison.
The C-family (compute-optimized) gives more CPU per dollar with less memory, while the M-family (general-purpose) balances CPU and memory at a higher price. Picking the wrong one wastes money on RAM you do not use or forces oversizing to get enough CPU. Here is how to choose on cost.
The R-family (memory-optimized) packs 8 GB of memory per vCPU, costing more per hour than general-purpose but far less than oversizing an M-instance to reach the memory. For in-memory databases, caches, and analytics, it is the cheapest way to buy RAM. Here is when it pays off.
An Auto Scaling group can save money by matching capacity to demand, or waste it by scaling on the wrong metric, holding too much warm capacity, or never scaling down. Here is how to tune an ASG so you pay for the load you have, not the peak you fear.
Lambda is cheap for spiky, low-utilization workloads and expensive for steady, high-utilization ones, where an EC2 instance running full-time wins. The breakeven depends on request volume, duration, and how busy the compute would be. Here is how to find the crossover.
Fargate Spot runs your ECS tasks on spare capacity for up to about 70 percent less than regular Fargate, at the cost of possible interruption. For fault-tolerant, stateless, and batch workloads it is one of the biggest container savings available. Here is how it works and when to use it.
Spot and Savings Plans are not either-or: the cheapest fleets layer them, committing to a steady baseline with Savings Plans and filling the elastic and interruptible portion with Spot. Here is how to combine them without wasting a commitment on capacity Spot should cover.
Idle EC2 instances, running at single-digit CPU with no real traffic, are one of the most common and avoidable cloud costs. Finding and stopping them, or scheduling them off, often cuts a compute bill by 20 to 40 percent. Here is how to hunt idle compute down.
Stopping an EC2 instance halts compute charges but keeps EBS billing; hibernating preserves memory to disk for fast resume but stores that memory on EBS, adding cost. For part-time workloads, knowing which to use and what still bills is worth real money. Here is the comparison.
Development, test, and staging environments rarely need to run outside business hours, yet many run 24/7. Scheduling them off nights and weekends cuts their cost by roughly 65 to 75 percent with no loss of function. Here is how to do it and what it saves.
Lambda cost per million invocations depends on two levers, the flat per-request fee and the per-GB-second duration charge, and the duration part usually dominates. Working through real numbers shows exactly where the cost goes and which lever to pull. Here is the breakdown.
Two numbers decide whether Savings Plans save money or waste it: coverage (how much of your usage the commitment covers) and utilization (how much of the commitment you actually use). Over-commit and utilization drops; under-commit and coverage is low. Here is how to balance them.
EC2 offers four ways to pay, on-demand, Spot, Savings Plans, and Reserved Instances, each with a different discount, commitment, and risk profile. Using the right one for each part of a workload is worth 50 to 80 percent off the naive on-demand bill. Here is the full comparison.
App Runner and Fargate both run containers without managing servers, but they price and scale differently. App Runner bundles more of the operational layer at a per-resource rate, while Fargate gives finer control and can scale to zero more cheaply. Here is the cost comparison.
Burstable T-family instances earn CPU credits when idle and spend them when busy, but under sustained load they exhaust credits and either throttle or, in unlimited mode, charge extra. Understanding when unlimited mode saves money and when it silently inflates the bill is worth knowing. Here is how it works.
ECS capacity providers decide where your tasks run and in what mix, and that decision drives the cost of an ECS cluster. Splitting between on-demand and Spot, tuning target capacity, and choosing Fargate versus EC2 providers can cut container spend substantially. Here is how they work on cost.
Fargate bills for exactly the vCPU and memory you allocate to a task, so over-allocating is pure waste that scales with task count. Right-sizing tasks to real usage often cuts Fargate spend by 30 to 50 percent. Here is how to size tasks to what they actually use.
AWS Batch schedules containerized batch jobs across compute you configure, and because most batch work is interruption-tolerant, running it on Spot can cut compute cost by up to 90 percent. Here is how to configure Batch for Spot savings without losing jobs to interruptions.