Service Quotas: The Limit You Will Hit First

Quotas are more specific than people expect

A quota is almost never “instances”. It is typically scoped by all of:

• Resource type — vCPU, instances, IP addresses, buckets, load balancers.
• Family or class — standard instances, memory-optimised, GPU.
• Region — a quota in Frankfurt says nothing about Singapore.
• Sometimes availability zone.

The classic surprise: an AWS account with a 32 vCPU standard-family quota in eu-west-1 cannot launch eight c6i.2xlarge instances, because that needs 64 vCPU of the compute-optimised family in that region. Two separate quotas, neither of which was the one you checked.

Where to look

How to get an increase approved

Quota requests are read by people or by systems trained on people’s answers. A clear, specific, unexciting request is approved faster than a vague one.

1. Ask for a specific number, not “more”. State the resource, the family, the region and the target value.
2. Explain the workload in one sentence, plainly. “Running a 12-node Kubernetes cluster for a web application” is ideal.
3. Say when you need it. A date makes it a scheduling question rather than an open-ended one.
4. Ask for a realistic headroom above what you need, not ten times it. Very large jumps trigger review.
5. Have some usage history first if you can. An account that has run modest workloads for a few weeks is a much easier approval.

If a request is declined, run something real for a while and ask again. Demonstrated normal usage moves quotas more reliably than argument.

Design around quotas rather than fighting them

• Spread across regions if a single-region quota is the constraint and latency permits.
• Use fewer, larger instances where the quota counts instances, or more, smaller ones where it counts vCPU.
• Check quota before an autoscaling group is allowed to grow; hitting a cap mid-scale is worse than not scaling.
• Use separate accounts or projects per environment where quotas are per-account — it also gives cleaner blast radius.
• Watch for quotas on things that are not compute: static IPs, NAT gateways, VPCs, security groups per network interface.

The quotas people forget

Compute quotas get the attention. These are the ones that bite unexpectedly:

• Elastic or static IP addresses per region — small by default, and easily exhausted.
• VPCs or virtual networks per region.
• Rules per security group, and security groups per interface.
• Concurrent function executions on serverless platforms.
• API request rates, which throttle rather than fail outright and are harder to diagnose.
• Number of load balancers, which is lower than most people assume.

Common questions

Where to go next

For what determines an account’s starting quotas, see what verified actually means. Configurations are described by their vCPU and instance capacity in the catalogue.