Three fixes for IP pressure, solving different constraints.

Per-node Pod limits and subnet-wide address exhaustion are different problems with different remedies.

Prefix delegation

Hand out /28 blocks

The CNI assigns whole 16-address prefixes to an ENI instead of single IPs, raising Pods per node a lot (up to the recommended cap). Fixes the per-node limit.

Custom networking

A separate CIDR for Pods

Pods use ENIs in subnets from a secondary VPC CIDR (for example 100.64.0.0/16) with an ENIConfig. Fixes VPC-wide exhaustion.

IPv6 cluster

Effectively unlimited Pod IPs

An IPv6 cluster gives Pods IPv6 addresses. Needs IPv6 support in your VPC and dependencies. A bigger design change.

Pods per m5.large (illustrative)

default
29

prefix delegation
up to about 110

# enable on the VPC CNI add-on
ENABLE_PREFIX_DELEGATION="true"
WARM_PREFIX_TARGET="1"

Plan subnets first

Use generous subnets (/19 or larger for nodes and Pods in busy clusters) and reserve ranges. Prefix delegation needs contiguous /28 blocks, so fragmented subnets can still fail.

Running out of IPs: prefix delegation, custom networking and IPv6

Two different shortages are easy to confuse. The first is per-node: an instance type simply cannot hold more than a limited number of Pod IPs. The second is subnet-wide: your VPC has run out of free addresses. Prefix delegation addresses the first. Instead of attaching individual secondary IPs to an ENI, the CNI attaches /28 prefixes (16 addresses each), so one node can run far more Pods, up to the recommended maximum (commonly 110 for mid-sized nodes). It also increases node density, which can lower node count and cost.

Custom networking addresses the second. It separates the node's primary ENI from Pod ENIs and places Pods in subnets taken from a secondary CIDR block associated with the VPC (commonly from the 100.64.0.0/10 range), configured with ENIConfig objects. It is the right tool when the main VPC CIDR is nearly exhausted across the account. The two techniques are often combined.

A third path is an IPv6 cluster, where Pods get IPv6 addresses and IPv4 exhaustion disappears, at the cost of ensuring your VPC, load balancers and every dependency work with IPv6. Treat it as a deliberate design choice rather than a quick fix.

The bar chart is illustrative: exact numbers vary by instance type and cluster settings, so check the EKS maximum-Pods guidance for your instance types. Gotcha: enabling prefix delegation on a cluster whose subnets are already fragmented can fail to allocate prefixes. Plan address space before you grow.