ChoiceSoftware deliveryCommunity

Cloud cost spike likely cause

Classify one sudden jump in a cloud bill or cost alert by its likely cause: expected growth, forgotten or idle resources, a misconfiguration or runaway job, data transfer, a pricing or billing change, or possible abuse, so small teams know who should look and how fast.

F@f

Limits to keep in mind

This only labels the likely cause. It does not stop, delete or change any cloud resource.

Billing data can lag by hours or a day; confirm with the provider's cost explorer before acting.

For any sign of compromise follow your security incident process and the provider's guidance, even if the label later turns out to be wrong.

Problem

Small engineering teams and founders get a cost alert or see the monthly cloud bill jump, but nobody owns cost full time. They lose hours guessing, panic over normal growth, miss a forgotten GPU instance or a log loop that keeps billing every hour, and notice leaked keys used for crypto mining only when the invoice arrives.

Goal

Label each cost spike with its most likely cause so expected growth is just noted, forgotten resources are cleaned up, runaway jobs and misconfigurations are fixed quickly, egress spikes are traced, pricing changes are checked against the provider notice, and possible abuse is treated as a security incident at once, with a human deciding anything unclear.

Input / state

One cost spike: what grew and by how much, related usage numbers, recent changes in the team's systems, and any billing notes from the provider.

cost_change
The service, account, region and cost before and after the spike, for example 'Compute eu-west-1: 410 USD/day to 1,950 USD/day since 2026-11-14'.
usage_metrics
Related usage numbers if available: instance hours, request counts, storage GB, data transfer GB, function invocations, log volume, active users or traffic.
recent_changes
Deploys, infrastructure or config changes, launches, experiments, migrations, new customers or marketing campaigns around the spike date.
billing_notes
Optional: provider emails, pricing change notices, expired credits or discounts, reserved capacity or savings plan changes, and security alerts.

Decision question

ChoiceQuestion key likely_cause

What most likely caused this cost spike? Use expected_growth when usage grew in line with real traffic or customers. Use forgotten_or_idle_resource when something keeps running with little or no use. Use misconfiguration_or_runaway when a change or bug makes a resource bill far more than intended. Use data_transfer_or_egress when traffic leaving a region or the cloud drives the jump. Use pricing_or_billing_change when usage is flat but the rate, credits or discounts changed. Use possible_abuse_or_compromise when the usage does not match anything the team runs.

Options

  • expected_growthUsage and cost rose roughly in proportion to real traffic, customers, data or a planned launch, and the unit cost per user or request stayed about the same.
  • forgotten_or_idle_resourceResources left running after a test, demo, migration or departed teammate: idle instances or GPUs, unattached volumes, old snapshots, unused load balancers or databases, or a dev environment that never scales down.
  • misconfiguration_or_runawayA deploy, config change or bug makes something bill far more than intended: retry or log loops, autoscaling with no maximum, a job stuck in a loop, debug logging left on, a cron running every minute, or a missing cache or lifecycle rule.
  • data_transfer_or_egressThe increase is mainly data transfer: traffic between regions or zones, internet egress, a missing CDN, large downloads or backups leaving the cloud, or a NAT gateway handling heavy traffic.
  • pricing_or_billing_changeUsage is roughly flat but the bill changed: a provider price change, free tier or startup credits ended, a reserved instance or savings plan expired, a new tax or a billing correction.
  • possible_abuse_or_compromiseUsage that nobody on the team recognizes: new instances in unused regions, sudden GPU or compute bursts, unknown users or API keys, crypto-mining patterns, or spikes right after a leaked key or security alert.

Allowed outcomes

Decided in your code from the answer and its confidence. Thresholds are yours to set and test.

  • Engineer reviews the billConfidence is below your threshold, the data is too thin to tell causes apart, several causes mix, or the input is not a cloud cost change.
  • Note as expected growthChoice is expected_growth and confidence meets your threshold.
  • Find the owner and clean upChoice is forgotten_or_idle_resource and confidence meets your threshold.
  • Fix the runaway job or config nowChoice is misconfiguration_or_runaway and confidence meets your threshold.
  • Trace where the data is goingChoice is data_transfer_or_egress and confidence meets your threshold.
  • Check the provider notice and contractChoice is pricing_or_billing_change and confidence meets your threshold.
  • Treat as a security incident nowChoice is possible_abuse_or_compromise and confidence meets your threshold.

Example scenarios

Predefined demonstrations. No model is called; the outcome shown is an illustrative judgment.

Example state

{
  "cost_change": "Compute and database: 320 USD/day to 540 USD/day since 2026-11-10",
  "billing_notes": "",
  "usage_metrics": "Requests up 70%, daily active users 4,100 to 7,000, cost per 1,000 requests flat",
  "recent_changes": "Public launch on 2026-11-10 with a Product Hunt post"
}

Outcome

Illustrative
Note as expected growth

Choice is expected_growth and confidence meets your threshold.

Why

Cost grew in line with real users after a planned launch and unit cost stayed flat.

Using it well

  • Compare cost growth with usage growth: cost rising much faster than traffic or customers points away from expected_growth.
  • Check the spike date against recent_changes first; a jump that starts within hours of a deploy or config change is usually misconfiguration_or_runaway.
  • If usage numbers are flat and only the rate or credits changed, choose pricing_or_billing_change, not a technical cause.
  • Resources in regions, accounts or services the team says it does not use, or compute bursts at odd hours with no matching deploy, are possible_abuse_or_compromise even if a benign explanation is possible.
  • A slow, steady rise from storage, snapshots or idle environments is forgotten_or_idle_resource; a sudden steep rise is more often runaway or abuse.
  • When data transfer GB grows much more than requests, prefer data_transfer_or_egress.
  • If the input mixes several services with different causes, or gives only a total with no breakdown, use needs_review and ask for a per-service breakdown.
  • Never suggest deleting resources or rotating keys as done; recommend what to check and who should act.

Request definition

POST https://api.typesafe.ai/v1/systemone
{
  "model": "jev-latest",
  "state": {
    "cost_change": "<cost_change>",
    "usage_metrics": "<usage_metrics>",
    "recent_changes": "<recent_changes>",
    "billing_notes": "<billing_notes>"
  },
  "questions": {
    "likely_cause": {
      "type": "choice",
      "instructions": "What most likely caused this cost spike? Use expected_growth when usage grew in line with real traffic or customers. Use forgotten_or_idle_resource when something keeps running with little or no use. Use misconfiguration_or_runaway when a change or bug makes a resource bill far more than intended. Use data_transfer_or_egress when traffic leaving a region or the cloud drives the jump. Use pricing_or_billing_change when usage is flat but the rate, credits or discounts changed. Use possible_abuse_or_compromise when the usage does not match anything the team runs.",
      "criteria": {
        "expected_growth": "Usage and cost rose roughly in proportion to real traffic, customers, data or a planned launch, and the unit cost per user or request stayed about the same.",
        "forgotten_or_idle_resource": "Resources left running after a test, demo, migration or departed teammate: idle instances or GPUs, unattached volumes, old snapshots, unused load balancers or databases, or a dev environment that never scales down.",
        "misconfiguration_or_runaway": "A deploy, config change or bug makes something bill far more than intended: retry or log loops, autoscaling with no maximum, a job stuck in a loop, debug logging left on, a cron running every minute, or a missing cache or lifecycle rule.",
        "data_transfer_or_egress": "The increase is mainly data transfer: traffic between regions or zones, internet egress, a missing CDN, large downloads or backups leaving the cloud, or a NAT gateway handling heavy traffic.",
        "pricing_or_billing_change": "Usage is roughly flat but the bill changed: a provider price change, free tier or startup credits ended, a reserved instance or savings plan expired, a new tax or a billing correction.",
        "possible_abuse_or_compromise": "Usage that nobody on the team recognizes: new instances in unused regions, sudden GPU or compute bursts, unknown users or API keys, crypto-mining patterns, or spikes right after a leaked key or security alert."
      }
    }
  }
}

Request shape from the System One API reference; a Kev server accepts the same body. Replace the placeholder state with your data. Requests need your own API key or server.

Version History

v1Current Versionabout 4 hours ago@f

Initial version

Comments (0)