Deployments.ai Docs

Costs

The AWS resources a default installation runs, which of them cost money while idle, and the variables that shrink them.

A default installation runs a fixed set of always-on resources and bills everything else by use. This page lists the always-on resources the OpenTofu modules create with the sample var file's production posture, then the levers that shrink them. It names resource types, sizes, and counts from the modules; check current prices for your region on the AWS pricing pages. The sample uses us-east-2.

infrastructure/environments/customer also creates AWS Budgets and a Cost Anomaly Detection monitor. billing_daily_budget_usd defaults to 100, billing_monthly_budget_usd to 350, and billing_anomaly_threshold_usd to 100, and they alert billing_alert_email.

Always-on resources

ResourceFeatureSize and count with the sample posture
OpenSearch domainsMemory and BrainsTwo domains, each 2 × m7g.large.search with 50 GiB gp3 per node. shared_search_domain = true makes it one domain.
Agent workerAgent runtimeFargate ARM64, 1 vCPU and 4 GiB per task, at least 2 tasks, up to 20.
Model gatewayModel runtimeFargate ARM64, 1 vCPU and 2 GiB per task, bifrost_desired_count tasks (1). Never Spot.
Durable StreamsAgent runtimeFargate x86, 1 vCPU and 2 GiB, 1 task, plus an EFS file system with AWS Backup.
Stream gatewayAgent runtimeFargate ARM64, 0.5 vCPU and 1 GiB per task, 2 tasks with high_availability, behind a public Application Load Balancer.
FirecrawlWeb toolsFargate, 2 vCPU and 4 GiB, web_tools_desired_count tasks (1).
Chromium rendererWeb toolsFargate, 1 vCPU and 4 GiB, web_tools_chromium_desired_count tasks (1).
Internal Application Load BalancerModel runtimeOne, shared by the model gateway, Durable Streams, and web tools.
NATNetworkOne t4g.small NAT instance and an Elastic IP while web tools are on; otherwise one NAT gateway and Elastic IP per zone (2).
Observability fact streamObservability historyOne provisioned Kinesis shard, plus Firehose delivery to S3 Tables.
ClickHouse nodeObservability hot planeOne t4g.medium instance in unlimited mode, 20 GiB gp3 data and root volumes, and daily backups and snapshots kept 7 days.
Hot-plane serviceObservability hot planeFargate ARM64, 0.25 vCPU and 0.5 GiB, 1 task.
Console distributionConsoleOne CloudFront distribution.
KMS keysData, model runtime, identityCustomer managed keys for data, governance signing, model gateway tokens, and identity signing.
WAF web ACLIdentity serviceOne, with web_application_firewall.

Everything else bills by use: Lambda, DynamoDB on-demand tables, S3 and S3 Tables storage, SQS, SNS, EventBridge, API Gateway, SES, CloudWatch Logs, Athena queries, Bedrock tokens, and Lambda MicroVM sandbox time. Bedrock bills Cohere Embed v4 through AWS Marketplace.

The OpenSearch domains and the Fargate services are the largest fixed costs. Container Insights on every ECS cluster and detailed metrics follow enhanced_monitoring.

Observability hot plane

The hot plane's defaults are the smallest proven size. This table compares their monthly cost with the earlier defaults, at us-east-2 On-Demand prices: 730 hours a month, r7g.large at $0.1071 an hour, t4g.medium at $0.0336 an hour, gp3 at $0.08 per GB-month with its 3,000 IOPS and 125 MiB/s baseline included, and Fargate ARM64 at $0.03238 per vCPU-hour and $0.00356 per GB-hour. Prices change, so check the AWS pricing pages.

PartEarlier defaultDefault
ClickHouse noder7g.large: $0.1071 × 730 = $78.18t4g.medium: $0.0336 × 730 = $24.53
Data volume200 GiB × $0.08 = $16.0020 GiB × $0.08 = $1.60
Root volume50 GiB × $0.08 = $4.0020 GiB × $0.08 = $1.60
Hot-plane task0.5 vCPU and 1 GiB: (0.5 × $0.03238 + 1 × $0.00356) × 730 = $14.420.25 vCPU and 0.5 GiB: (0.25 × $0.03238 + 0.5 × $0.00356) × 730 = $7.21
Total$112.60$34.94, $77.66 or 69 percent less

The totals leave out what depends on use: backups, surplus CPU credits, the node's three custom CloudWatch metrics and its alarms, and data transfer.

Backups depend on how much data ClickHouse holds. Each daily native backup is a full copy in S3, so the bucket holds about one copy per day of retention; at S3 Standard's $0.023 per GB-month, that is retention × data × $0.023. EBS snapshots are incremental at $0.05 per GB-month: one full copy when nothing changes, up to one copy per day of retention when merges rewrite every part daily. With 16 GiB of data, the point where the 20 GiB volume's usage alarm fires:

Backup retentionNative backupsSnapshotsTotal
7 days, the default7 × 16 × $0.023 = $2.5816 × $0.05 = $0.80 to 7 × 16 × $0.05 = $5.60$3.38 to $8.18
90 days, the earlier default with the sample posture90 × 16 × $0.023 = $33.1216 × $0.05 = $0.80 to 90 × 16 × $0.05 = $72.00$33.92 to $105.12

A t4g.medium that runs above its 20 percent baseline for long enough pays for surplus CPU credits; Sizing shows what sustained load adds and how to scale up. Even both vCPUs at 100 percent all month, $24.53 + $46.72 = $71.25, cost less than the r7g.large.

Event analytics

Event analytics runs by default and costs about $0.60 a month while idle: six CloudWatch alarms at $0.10 each. Its projector Lambda, Firehose stream, S3 failure and result buckets, dead-letter queue, and Athena workgroup bill only by use. Each event costs one Lambda invocation and its bytes through Firehose into S3 Tables. Each operator query bills at least 10 MB scanned, and the workgroup stops any query past 10 GiB. The semantic event table exists in every installation whether or not the stack runs.

Price sandbox time

By default the agent worker meters Lambda MicroVM sandbox time without a cost, so sandboxes add nothing to an organization's spend. To price them, set the sandbox_usage_cost_policy input of module "platform". Each rate is a whole number of USD micros per hour, from 0 to 1,000,000,000. A USD micro is one millionth of a dollar.

RatePrices
base_hour_usd_microsEach hour a sandbox runs, the sandbox_ms meter.
vcpu_hour_usd_microsEach reserved vCPU-hour, the sandbox_vcpu_ms meter.
memory_gib_hour_usd_microsEach reserved GiB-hour of memory, the sandbox_memory_mib_ms meter.

This example passes on the Lambda MicroVM compute price in US East (N. Virginia) on Graviton, as AWS lists it in the Lambda pricing examples. AWS charges $0.0000276944 per vCPU-second and $0.0000036667 per GB-second. Per hour, that is 99,700 USD micros per vCPU and 13,200 USD micros per GB, rounded to the micro.

sandbox_usage_cost_policy = {
  vcpu_hour_usd_micros       = 99700
  memory_gib_hour_usd_micros = 13200
}

A default sandbox reserves 1 vCPU and 2 GiB, so an hour of it costs 99,700 + 2 × 13,200 = 126,100 USD micros, about $0.13. The policy prices the reservation only. AWS also bills compute above the baseline, snapshot reads, writes, and storage, and data transfer. Raise the rates if they should cover those too. Prices change, so check the AWS pricing page for your region.

OpenTofu rejects a rate outside that range at plan, and passes a set policy to the agent worker as DEPLOYMENTS_AI_SANDBOX_USAGE_COST_POLICY. Each priced meter event carries its costUsdMicros and funding: operator, and the usage materializer adds it to the organization's operator-funded month spend. Sandbox time then counts toward the same operator-funded spend ceiling as model usage on the installation's keys. The input is null by default, which leaves sandbox time unpriced and the worker environment unchanged.

Shrink a default installation

LeverEffect
operations.shared_search_domain = trueOne OpenSearch domain for memory and Brains instead of two.
operations.compact_sizing = trueSmaller Fargate tasks: worker 0.5 vCPU and 1 GiB, 1 to 4 tasks; model gateway 0.5 vCPU and 1 GiB; Durable Streams 0.25 vCPU and 0.5 GiB; Firecrawl 0.5 vCPU and 3 GiB.
operations.fargate_spot = trueFargate Spot for every ECS service except the model gateway.
operations.high_availability = falseOne stream gateway task instead of two.
operations.dedicated_public_load_balancers = falseCloudFront in front of the shared internal load balancer instead of a public load balancer for the stream gateway.
operations.enhanced_monitoring = falseNo Container Insights or detailed metrics.
web_tools_runtime_enabled = falseNo Firecrawl or Chromium tasks. Egress returns to NAT gateways unless you set network_nat_mode = "instance".
network_single_nat_gateway = trueIn gateway mode, one NAT gateway for every zone.
agent_observability_hot_enabled = falseNo ClickHouse node or hot-plane task. The observability API then answers 503 observability_unavailable.
agent_observability_history_enabled = falseNo Kinesis shard or Firehose streams. Needs agent_observability_hot_enabled = false too.
agent_email_enable_runtime = falseNo agent email workers. The SES domain identity stays.
event_analytics_lifecycle = "absent"No projector, Firehose stream, workgroup, or alarms. Only before the first deploy; later, see Lifecycle.

compact_sizing, fargate_spot, and turning high_availability off trade capacity and availability for cost. Keep the production posture for an installation that serves real traffic.

On this page