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.
Speech-to-Text bills per minute of audio (rounded up), with higher rates for premium and specialized models and features like diarization. Long audio and premium models drive the bill. Here is the per-minute model and how to control it.
Redshift, BigQuery, and Synapse price data warehousing on different axes, provisioned nodes or serverless, per-TB-scanned or per-slot, per-DWU. The cheapest depends on query pattern and volume, not the platform. This compares the models.
DynamoDB, Firestore, and Cosmos DB price NoSQL differently, DynamoDB per capacity or request, Firestore per operation, Cosmos per request unit or serverless. The cheapest depends on read/write mix and traffic shape. This compares the models.
AWS RDS, GCP Cloud SQL, and Azure Database for PostgreSQL Flexible Server price managed Postgres similarly, compute per vCPU and memory, storage, and high-availability options. The cost drivers and levers are shared. This compares them.
t4g (Graviton, ARM) burstable instances cost about 20% less than t3 (Intel) for comparable size, often with better performance per dollar. The catch is ARM compatibility. Here is the comparison and when the switch is free money.
AWS instance families (general purpose, compute, memory, storage, and accelerated) target different workloads, and using the wrong family over-pays for resources you do not need. This is a practical guide to choosing the cost-effective family.
i4i storage-optimized instances offer faster NVMe local storage and better price-performance than the older i3 generation. For high-IOPS, low-latency local-disk workloads, the newer generation is usually the cost-effective choice. Here is the comparison.
AWS EBS, Azure Managed Disks, and GCP Persistent Disk price block storage similarly, by provisioned capacity and performance tier, with the fastest tiers charging for provisioned IOPS. This compares them and shows the shared optimization levers.
EKS, AKS, and GKE all charge for the worker nodes; the difference is the control-plane fee (EKS and GKE charge per cluster-hour, AKS's standard tier is free) and autopilot-style options. This compares them and where the real cost lives.
Polling burns compute and requests asking for changes that usually have not happened; event-driven architectures do work only when something occurs. For infrequent events, event-driven is dramatically cheaper. Here is the cost tradeoff.
SQS, Google Pub/Sub, and Azure Service Bus price messaging differently, SQS per request, Pub/Sub per TiB of throughput, Service Bus per operation or messaging unit. This compares the models and the surprises in each.
Compression shrinks the bytes you store, transfer, and scan, and cloud bills those bytes. Compressing data can cut storage, egress, and query cost at once, for a little CPU. Here is where compression pays and where it does not.
Most data cools over time: hot when new, cold within weeks, rarely touched within months. A tiered storage strategy moves data through cheaper tiers as it ages, cutting storage cost without losing access. Here is how to design one.
A cost anomaly, a runaway resource, a misconfigured loop, a leaked key, can add thousands before the monthly bill reveals it. Anomaly alerting detects unusual spend as it happens. Here is how it works and how it complements pre-deploy estimation.
Spot instances are cheap but reclaimable, and concentrating a fleet in one instance type risks a mass interruption. Diversifying across types and zones smooths that risk, letting you run more of a workload on Spot safely. Here is how.
MediaConvert bills per minute of output video, priced by resolution, frame rate, and features like HDR or advanced codecs. Encoding more than you need, in higher tiers than required, drives the bill. Here is how the per-minute model works.
IoT Core bills connectivity per million connection-minutes, messaging per million messages, and rules-engine actions, plus device shadow and registry operations. Chatty devices and heavy rule fan-out drive the bill. Here is the model.
Amazon Managed Service for Prometheus bills per metric sample ingested, per query processed, and for storage. High-cardinality metrics and short scrape intervals drive ingestion cost. Here is how the model works and how to control cardinality.
Amazon Managed Grafana bills per active user per month, by role (editor or viewer), plus optional add-ons. Cost scales with your user base, not dashboards or data. Here is how the per-user model works and how to keep it lean.
Keyspaces (managed Cassandra) bills read and write request units, storage per GB-month, and optional features, with on-demand and provisioned capacity modes. It scales to zero with no clusters to manage. Here is how it prices versus self-managed Cassandra.
Azure Container Instances bills per vCPU-second and memory-GB-second while a container runs, with no cluster to manage or idle. It suits bursty and short-lived containers; steady workloads are cheaper on AKS. Here is the model.
Data Lake Storage Gen2 prices like blob storage, hot, cool, cold, and archive tiers plus transaction and data-retrieval charges, with a hierarchical namespace. Matching tier to access frequency and minimizing transactions are the levers. Here is the model.
Apigee is a full API management platform priced by subscription tier (or pay-as-you-go per API call), with a significant baseline for the standard tiers. It is powerful but not cheap. Here is how it prices and when a lighter gateway suffices.
GCP Batch adds no charge of its own; you pay for the Compute Engine VMs it runs your jobs on, including Spot. Spot workers, right-sized machines, and jobs that scale to zero are the levers. Here is how to run batch cheaply.