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.
AWS Lambda, Azure Functions, and Google Cloud Functions all bill invocations plus compute (memory times duration), with a free tier. The models are close; the real cost differences come from configuration and the surrounding services. Here is the comparison.
S3 Express One Zone charges far higher storage (around $0.16/GB-month) but much lower per-request costs than S3 Standard, in a single AZ for low latency. For request-heavy, latency-sensitive workloads it can be cheaper overall. Here is the math.
Global tables replicate a DynamoDB table across regions, so you pay replicated write capacity, storage in every region, and cross-region replication transfer. The multiplier is the number of regions. Here is when multi-region is worth it.
EFS charges storage plus, depending on throughput mode, either nothing extra (Bursting), per throughput used (Elastic), or per provisioned throughput. Picking the wrong mode over-pays or throttles. Here is how the modes price and when to use each.
CloudFront Functions cost about $0.10 per million invocations for lightweight viewer logic; Lambda@Edge bills per request plus duration for heavier compute across more event types. Choosing the lighter tool where it fits cuts edge cost. Here is the comparison.
Front Door Standard has a modest base fee plus usage; Premium costs several times more but bundles managed WAF rules, bot protection, and Private Link. Paying for Premium features you do not use is the overspend. Here is how to choose.
BigQuery storage bills active data (recently modified) at about $0.02/GB-month and long-term data (untouched 90 days) at about half that, with an option to bill physical (compressed) bytes instead of logical. Here is how to cut storage cost.
Cloud SQL now has two editions: Enterprise and Enterprise Plus. Enterprise Plus costs more per vCPU and memory but adds a data cache, better performance, and near-zero-downtime maintenance. Here is how the editions price and when the upgrade pays off.
A Kubernetes cluster bill is one number; the hard part is attributing it to namespaces, teams, and workloads. This is how cost monitoring works, what it needs (requests, labels, allocation), and how it complements pre-deploy estimation.
Rightsizing done once drifts back to waste as workloads change. Automating it, through recommendations, policies, and pre-deploy checks, keeps resources matched to real usage over time. Here is how automated rightsizing works and where it fits.
Serverless bills only for what you use, which can be cheap or expensive depending on memory, duration, invocation volume, and idle provisioned capacity. This is a practical guide to keeping Lambda, Fargate, and serverless databases economical.
Direct Connect bills a port-hour fee by connection speed plus data transfer out at rates well below internet egress. For high, steady traffic to AWS it pays back the port fee; for low volume it does not. Here is the model.
Site-to-Site VPN bills about $0.05 per connection-hour (roughly $36/month) plus standard data transfer out. It is a cheap, quick way to connect on-premises to AWS over the internet. Here is the model and when to prefer Direct Connect.
Storage Gateway charges for the AWS storage it writes to (S3, EBS snapshots, or virtual tapes), requests, and data transfer, plus a per-gateway fee on some types. The bulk of the bill is the backing storage. Here is the model.
QuickSight bills per author (a monthly subscription) and per reader (session-based with a monthly cap), plus SPICE in-memory capacity beyond the included allowance. Matching author and reader licensing to real usage is the main lever. Here is the model.
Bedrock bills per 1,000 input and output tokens on-demand, with rates varying by model, plus a provisioned-throughput option (per model-unit-hour) for steady high volume. Prompt and output length drive on-demand cost. Here is how to control it.
Azure OpenAI bills per 1,000 tokens by model on the standard pay-as-you-go option, or per Provisioned Throughput Unit (PTU) for reserved, predictable capacity. Model choice and token volume drive standard cost. Here is how to control it.
Azure Cache for Redis bills per hour by tier (Basic, Standard, Premium, Enterprise) and cache size, provisioned regardless of use. Basic has no replica; Standard and above add availability. Here is how the tiers price and how to choose.
AlloyDB bills compute per vCPU-hour and memory, plus regional storage per GB-month with no separate I/O charge, plus backups. It is a PostgreSQL-compatible database aimed at demanding workloads. Here is how it prices versus Cloud SQL.
Cloud TPU bills per chip-hour by TPU version and configuration, with on-demand, Spot (preemptible), and committed-use pricing. Idle TPUs and full-price on-demand for interruptible training are the expensive mistakes. Here is how to cut cost.
Inf2 (Inferentia) and Trn1 (Trainium) instances are AWS custom ML accelerators priced per instance-hour, often far cheaper per inference or training step than GPU instances. The catch is model compatibility. Here is when they cut cost.
Timestream bills writes per million, storage in a fast memory tier (per GB-hour) and a cheap magnetic tier (per GB-month), and queries per GB scanned. Tuning retention between tiers and scanning less are the levers. Here is the model.
Cosmos DB serverless bills per request unit (RU) consumed plus storage, with no provisioned throughput to pay for when idle. It suits spiky or low-traffic workloads; steady high traffic is cheaper on provisioned throughput. Here is how to choose.
Synapse bills dedicated SQL pools per DWU-hour, serverless SQL per TB scanned, and Spark pools per vCore-hour, plus data integration. Pausing dedicated pools and using serverless for ad-hoc queries are the big levers. Here is the breakdown.