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.
You can call a hosted model API and pay per token, or run your own model on GPU instances and pay for uptime. Each wins at a different scale. Here is how the two cost models compare and where they cross over.
NAT gateways charge both an hourly fee and a per-GB data processing fee, so a chatty private subnet can rack up a large bill nobody expected. Here is why it happens and how to cut the cost.
Observability data (metrics, logs, and traces) is billed by volume, and unmanaged it can rival your compute bill. A cost-aware approach controls what you collect, how long you keep it, and how you sample. Here is the playbook.
Migrating from a data center to AWS has a one-time project cost and a new ongoing run cost that replaces your capital spend. Understanding both, and how they compare to on-prem, is essential before you commit. Here is the breakdown.
Managed OpenSearch bills for data nodes, storage, and often dedicated master and warm nodes, so a large or over-provisioned cluster gets expensive fast. Here is how to right-size and tier it without losing search performance.
AWS PrivateLink gives private, secure access to services without traversing the public internet, but interface endpoints bill an hourly fee per endpoint per AZ plus a per-GB data processing charge. Here is how the cost works and when it pays off.
Self-hosted Prometheus trades a per-metric SaaS bill for the cost of running and operating your own monitoring stack. Which is cheaper depends on scale, cardinality, and how much operational time you can spend. Here is the comparison.
A retrieval-augmented generation pipeline spends on embeddings, a vector database, retrieval compute, and the LLM tokens for augmented prompts. Each stage has its own cost driver, and context size ties them together. Here is the full breakdown.
X-Ray bills about $5 per million traces recorded and about $0.50 per million traces retrieved or scanned, with a free tier. Sampling is the main lever, since you rarely need every request traced. Here is how the tracing model prices.
Traffic Manager bills per million DNS queries plus a monthly fee per monitored endpoint, with external endpoints costing more to check. It is a cheap DNS-based global router. Here is how it prices and where cost hides with many endpoints.
Secret Manager bills about $0.06 per active secret version per location per month and about $0.03 per 10,000 access operations. It is cheap unless you store many versions or access secrets on every request. Here is the model and the caching lever.
DataSync bills about $0.0125 per GB of data copied, plus any underlying storage and transfer at the endpoints. Cost tracks how much you move, so incremental transfers and filters keep it down. Here is the simple model.
AWS Batch adds no charge of its own; you pay for the EC2, Fargate, or Spot compute your jobs run on. Compute environment choice and Spot are the big levers. Here is how to run batch workloads cheaply.
AppSync bills GraphQL query and data-modification operations per million, real-time subscription messages per million, and connection-minutes. Chatty clients and broad subscriptions drive the bill. Here is how the managed-GraphQL pricing works.
Azure Standard Load Balancer bills per load-balancing rule and per GB of data processed, plus outbound. The Basic tier is free but limited and retiring. Here is how the Standard tier prices and how to keep it lean.
Cloud Composer bills environment compute (per-hour, always on) plus the underlying GKE, database, and storage. The always-on environment fee is significant, so consolidating DAGs and right-sizing the environment matter. Here is the breakdown.
The right index turns a full scan into a fast lookup, cutting both query latency and the compute or I/O you pay for. But every index adds write overhead and storage. Here is how indexing trades cost against performance and how to get it right.
Databases are often the biggest single line on the bill and the most over-provisioned, sized for a peak that rarely comes or copied from another environment. Right-sizing to real CPU, memory, and IO need cuts cost without hurting performance. Here is how.
Running in multiple regions cuts latency for distributed users and adds resilience, but it multiplies compute, replicates data, and adds cross-region transfer. Knowing when the latency and availability are worth the multiplied cost is the decision. Here is the framework.
Redshift Serverless bills per Redshift Processing Unit (RPU) hour while queries run, plus managed storage, with no cluster to pause. It suits variable, intermittent analytics; steady heavy load is cheaper on a provisioned cluster. Here is how to size it.
Amazon MQ bills broker instance hours by size and deployment (single or active-standby), plus storage. Active-standby doubles the instance cost for high availability. Here is how the managed-broker pricing works and when to use it.
WorkSpaces bills per user by bundle, in either an always-on monthly mode or an hourly mode with a small monthly base plus per-hour usage. Choosing the right mode for each user's usage pattern is the main saving. Here is how to decide.
Transfer Family bills about $0.30 per hour for each enabled protocol endpoint (SFTP, FTPS, FTP) plus about $0.04 per GB uploaded and downloaded. The always-on endpoint fee is the surprise for low-volume use. Here is the breakdown.
Azure VPN Gateway bills a fixed hourly fee by SKU (from about $0.19/hour for VpnGw1 to several dollars for higher SKUs) plus outbound data transfer. The SKU, chosen for throughput and tunnels, dominates the bill. Here is how to size it.