Skip to content

Latest commit

 

History

History
67 lines (56 loc) · 4.56 KB

File metadata and controls

67 lines (56 loc) · 4.56 KB

AWS Run-Cost Model — EDU (K-12 & Higher Education)

The infrastructure "what does it cost to run" line items for the governed pattern, so a business case and a finance review use the same list across the portfolio. Illustrative framework, not a quote — prices change and vary by Region; populate each line with the AWS Pricing Calculator for your Region and volume. Pairs with the value-side model in offerings/ (ROI/TCO). Worked example below is for Agent 01 (Student & Family Concierge) at pilot scale.

Cost drivers (in rough order of impact)

  1. Amazon Bedrock inference (per token) — usually the dominant variable cost; scales with invocations × tokens/call × model price. The single biggest lever (model choice, prompt/context size).
  2. Human-in-the-loop wait time — Step Functions waitForTaskToken is cheap to hold, but the human is the real cost; see OPERATING-MODEL.md.

Line items — every service the pattern uses

# Service Billing basis Scales with Pilot magnitude*
1 Amazon Bedrock (inference) per input/output token invocations × tokens × model $$–$$$ (dominant)
2 Bedrock Guardrails per text unit evaluated every model call $
3 AgentCore Gateway or API Gateway per request (+ AgentCore units) tool calls / requests $
4 AWS Lambda requests + GB-seconds tool calls, workflow steps $
5 AWS Step Functions per state transition workflow steps × runs $
6 Amazon DynamoDB (on-demand) R/W request units + storage audit + approvals + jti writes $
7 Amazon S3 + Object Lock (WORM) storage + requests audit evidence retained $ (grows with retention)
8 Amazon CloudWatch logs GB + metrics + dashboards log volume, custom metrics $–$$ (watch log volume)
9 AWS WAF web ACL + rules + requests public edge traffic $
10 AWS PrivateLink (interface VPC endpoints) per-AZ-hour + data processed # endpoints × AZs (always-on) $$ (fixed, always-on)
11 Amazon Cognito monthly active users workforce users $ (free tier often covers pilots)
12 AWS KMS keys + requests CMK per data class + calls $
13 AWS Support % of monthly AWS spend (Business/Enterprise) total spend % add-on
14 Operations (customer staff) FTE time monitoring, approvals, IR, break/fix see OPERATING-MODEL

*Magnitudes are relative ($ ≈ tens, $$ ≈ hundreds, $$$ ≈ thousands / month at pilot scale) — replace with real figures from the Pricing Calculator. Data class for this pack: Student PII → PrivateLink for the record-touching variant.

Worked estimate framework (fill in)

Monthly Bedrock       = invocations/mo × avg_tokens/call × price_per_1k_tokens/1000
Guardrails            = invocations/mo × text_units/call × guardrail_price
Gateway/API GW        = tool_calls/mo × request_price
Lambda                = (tool_calls + steps)/mo × (request_price + GB_s_price × avg_GB_s)
Step Functions        = runs/mo × transitions/run × transition_price
DynamoDB              = (audit + approval + jti writes)/mo × WRU_price + storage_GB × storage_price
S3 Object Lock        = retained_GB × storage_price + requests
CloudWatch            = log_GB × ingest_price + custom_metrics × metric_price + dashboards
WAF                   = web_ACL + rules + (requests/mo × request_price)
PrivateLink           = endpoints × AZs × 730h × hourly_price + data_GB × processing_price   # ALWAYS-ON
Cognito / KMS         = MAU × price (often free tier) ; CMKs × key_price + requests
Support               = support_tier_% × (sum of the above)
------------------------------------------------------------------
Monthly AWS run-cost  = Σ above     ·  + human reviewer time (operations)

How to reduce it

  • Model & context size dominate — right-size the model, trim prompt/context, cache retrieval.
  • PrivateLink is a fixed always-on cost — share endpoints across agents in one VPC; don't provision per-agent endpoints if one set serves the account.
  • CloudWatch log volume is a silent driver — set retention, sample verbose logs.
  • Marginal cost per additional agent is low — the platform (gateway, audit, endpoints, WAF) is shared; agents 2..N mostly add Bedrock + Lambda/Step-Functions usage.

Illustrative only; not a quote or a guarantee. Combine with the value-side model in offerings/ for a full business case, and see NOT-CLAIMS.md for scope.