Risk · R-138 · Operational Resilience
AI capacity and cost overrun
Usage, latency or spend exceeds what was planned, degrading the service or the budget, without any attacker being involved.
- Severity
- Medium
- Audience
- Heads-of · Practitioner
- Frameworks
- AML.T0034 · MAP 3
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-11
What it is
AI systems are charged by use and their use is hard to predict. A feature proves popular, prompts grow longer as context is added, an agent takes more steps than expected, a retry loop multiplies calls. Cost rises with each. Capacity is finite too: rate limits are reached, latency climbs, and other systems sharing the same quota are starved. This entry concerns ordinary, unplanned growth. Deliberate exhaustion by an attacker is a security risk covered elsewhere in the library.
How it happens
- The business case assumed a volume and a prompt size that both grew.
- An agentic workflow loops or retries without limit.
- Several systems share one quota with no allocation.
- Nobody is watching spend between monthly invoices.
Why it matters
Budgets are exceeded, sometimes many times over, before anyone notices. Latency degrades the customer experience. A cheaper model substituted in haste to control cost brings quality and validation problems of its own.
Warning signs
- Spend is reviewed only when the invoice arrives.
- No budget or usage alert is configured.
- Rate limit errors are increasing.
- Cost per transaction has drifted upward since launch.
Mitigating controls
Primary controls
- AI cost and capacity monitoring Detective
- Tracks usage, latency and spend against limits with alerts. Usage, latency and spend for each AI system are tracked against budgets and limits, with alerts and automatic caps before a limit is breached.
Supporting controls
- AI use-case intake and approval Preventive
- Tests the cost case before approval. Every proposed AI use is registered, screened, tiered and approved by a named authority before build or purchase begins, and again before it goes live.
- Inference usage quotas and rate limiting Preventive
- Caps consumption at source. Limits on request rate, input size, output size and spend are enforced for each user, tenant and system, with alerts as limits are approached.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| MITRE ATLAS | AML.T0034 | Cost Harvesting The adversarial form of the same exposure. |
| NIST AI Risk Management Framework (AI RMF 1.0) | MAP 3 | Capabilities, intended use, benefits and costs are understood Costs are understood before deployment. |
Sources
- NIST AI Risk Management Framework (AI RMF 1.0) NIST · framework
- MITRE ATLAS MITRE · framework
Related terms
How to cite
Axiom Verity (2026). AI capacity and cost overrun. Reference Library, R-138. https://axiom-verity.com/risks/ai-capacity-and-cost-overrun. Reviewed 2026-11-11.
Found an error or a better source? Report a correction.