Back to home
Proprietary engine

K8s Shapley Attribution

Kubecost, Vantage, and Cloudability split shared cluster overhead by proportional resource share — oversimplified, gameable, and unfair to small workloads. We use game-theoretic Shapley value — provably fair.

For every pod: direct cost (CPU × memory × GPU × egress × PV) plus its Shapley share of cluster shared overhead (control plane + monitoring + load balancer + idle node capacity) — computed via Monte-Carlo sampled Shapley (Castro et al. 2009) over the coalition game.

Cluster monthly
$10.5K
Direct cost
$7.5K
CPU + mem + GPU + egress + PV
Shared overhead (Shapley)
$3.1K
400 samples
Pods analysed
30
Team rollup
Auto-derived from namespace → team mapping.
ml-platform$6.7K · 64%
platform$2.8K · 27%
growth$970.30 · 9%
Namespace rollup
Per-namespace monthly spend.
prod-ml64%$6.7K
prod-api27%$2.8K
prod-growth9%$970.30
Top 10 pods by attributed cost
Direct + Shapley share of shared overhead.
PodNamespaceCPU utilMem utilDirectShapleyTotalWaste
recommendation-engine-b59e12-0prod-ml92%55%$1.0K$250.16$1.3K26%
recommendation-engine-3004bf-1prod-ml69%83%$960.11$246.51$1.2K24%
embedding-worker-db2585-3prod-ml75%90%$895.48$165.23$1.1K17%
embedding-worker-097691-2prod-ml87%57%$727.81$128.73$856.5428%
embedding-worker-413d76-0prod-ml70%84%$698.30$121.43$819.7323%
embedding-worker-896222-1prod-ml85%46%$615.03$154.28$769.3135%
vector-db-e616e4-0prod-ml50%68%$275.72$146.98$422.7041%
api-gateway-ee46b0-2prod-api48%78%$292.99$102.69$395.6837%
api-gateway-46a991-1prod-api68%51%$307.54$88.09$395.6341%
api-gateway-52d729-0prod-api48%92%$293.28$80.79$374.0731%
Try with YOUR cluster

Upload `kubectl top pods` CSV

No agent install. No kubeconfig. Just paste your pod list and we'll attribute every byte of shared overhead via Shapley value in 2 seconds.

Required columns: pod_id, namespace, cpu_req_cores, mem_req_gb. Optional: gpu_count, egress_gb_per_day, pv_gb.

Cost dimensions

6 signals fed into the algorithm. K1–K4 + K6 are direct cost; K5 is the inventive step — shared overhead attributed via Shapley.

K1
CPU reservation × CPU $/hr
Direct CPU cost from pod's `requests.cpu` × node-pool CPU rate.
K2
Memory reservation × memory $/hr
Direct memory cost from pod's `requests.memory` × node-pool memory rate.
K3
GPU seconds × GPU $/hr
Direct GPU cost from observed GPU utilisation × GPU node rate.
K4
Network egress (sidecar-aware)
Per-pod outbound bytes × egress price; sidecar containers attributed to their parent pod.
K5
Shared overhead via Shapley
Cluster-wide costs (control plane, monitoring, load balancer, idle capacity) split via Shapley value over the coalition game.
K6
Persistent volume $/hr
Pod-mounted PV cost from EBS / GCE-PD rate × volume size.

We value your privacy. Cookies help us improve your experience. Learn more

Install CARTIE AI

Add to your home screen for quick access and offline support