Portfolio 04 · whole-node capacity

Kubernetes GPU capacity and cost planner

Pack identical GPU pods against accelerator topology, CPU, and RAM; add HA and rollout reserve before checking autoscaler bounds and pricing whole nodes.

Decision this answers: How many whole GPU nodes must be rented, which resource binds, and can the configured autoscaler ceiling schedule the workload?

Scenario

Decision summary

Whole nodes rented

5exact-derived

Workload nodes plus HA and rollout reserve.

[1]

GPUs rented

40exact-derived

Node count multiplied by whole-node GPU capacity.

[1]

Monthly node cost

$113880.00exact-derived

No fractional pod allocation pricing.

[2]

Fleet plus control plane

$113953.00exact-derived

Whole-node cost plus separate cluster fee.

[2][3][4]

Direct-labeled result profile

Whole nodes rented GPUs rented Monthly node cost Fleet plus control plane

The tables below are the accessible source of truth; bar lengths never carry identity alone.

Managed Kubernetes worked examples

Managed Kubernetes worked examples
Candidate	Binding resource	Nodes	GPUs rented	Control plane	Availability claim	Evidence
GKE	GPU	5	40	Separate cluster-hour	Not checked	exact-derived[2][1]
AKS	GPU	5	40	Tier-specific	Not checked	exact-derived[3][1]
EKS	GPU	5	40	Separate cluster-hour	Not checked	exact-derived[4][1]

Official surface versus this site

Kubernetes GPU scheduling

Use the official surface to confirm native prices, counting rules, limits, and product eligibility.

This offline workbench

Kubernetes documents extended-resource scheduling. This planner makes homogeneous integer bin packing, HA reserve, rollout surge, control-plane fees, and whole-node cost explicit across managed services.

Sources and methodology

Every result-affecting reference is visible here without JavaScript and is retained in the JSON export.

[1] Kubernetes GPU scheduling

Authority: Kubernetes
Native identifier: extended resource nvidia.com/gpu
Unit: integer extended resource
Evidence: exact-derived
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: GPU requests are treated as indivisible scheduling constraints.

Open the authoritative source

[2] GKE pricing

Authority: Google Cloud
Native identifier: GKE cluster and node meters
Unit: cluster-hour plus node-hour
Evidence: negotiated
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Control-plane and editable whole-node rates remain separate.

Open the authoritative source

[3] AKS cost management

Authority: Microsoft
Native identifier: AKS tier and node meters
Unit: cluster and VM hour
Evidence: negotiated
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Tier-specific control-plane behavior is retained; node rate is editable.

Open the authoritative source

[4] Amazon EKS pricing

Authority: Amazon Web Services
Native identifier: EKS cluster
Unit: cluster-hour
Evidence: published list
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: The cluster fee is separate from EC2 node cost and extended-support pricing.

Open the authoritative source

Limits and non-claims

The opening 5-node plan is viable against the configured maximum: true.
This homogeneous planner does not solve mixed-node or arbitrary multi-dimensional bin packing.
Selected SKUs are not asserted available, quota-approved, or performance-equivalent.