Skip to content

Global Capacity Orchestrator

Guidance for EKS AutoMode Clusters with Global Capacity Orchestrator on AWS · Official AWS Solutions Guidance

One API. Every Accelerator. Any Region.

GCO turns a fleet of EKS Auto Mode clusters into one accelerated-compute platform. You hand it a Kubernetes manifest; it validates the manifest, finds a Region with capacity, provisions matching NVIDIA GPU, AWS Trainium, AWS Inferentia, or CPU (amd64 and arm64/Graviton) nodes, runs the workload, and keeps the outputs after the pods are gone. Inference endpoints deploy to every configured Region with one command, and an AI agent can drive all of it for you through the project's own MCP server.

Try it

Five commands take you from a clean machine to a job running on GCO. AWS charges start at the third one, and the last one removes everything again.

git clone https://github.com/aws-solutions-library-samples/global-capacity-orchestrator-on-aws.git && cd global-capacity-orchestrator-on-aws
./scripts/setup-dev-alias.sh && source ~/.zshrc                   # dev container + `gco` shell function (or ~/.bashrc — the script says which)
gco stacks deploy-all -y                                          # billable from here: global control plane + every Region in cdk.json
gco jobs submit-sqs examples/simple-job.yaml --region us-east-1   # your first job
gco stacks destroy-all -y                                         # tear it all down

The Get started page walks through each step, shows what success looks like, and points at the fixes when it does not.

Checking GPU capacity for g5.xlarge in us-east-1 through the GCO MCP server

Or just ask: the GCO MCP server answering a GPU-capacity question inside an AI-powered IDE. gco autopilot sets this up for Claude Code, OpenAI Codex, or OpenCode in one command.

Is GCO for you?

GCO fits teams running accelerated workloads — LLM training and inference, batch ML, HPC, and everyday CPU jobs — who want:

  • Capacity-aware placement instead of checking each Region by hand: spot placement scores, spot price history, capacity reservations, and Capacity Blocks behind gco capacity, with auto-Region submission built on them.
  • One IAM-authenticated API for every Region, with health-based failover through Global Accelerator in the commercial aws partition and IAM-authenticated regional APIs in aws-cn and aws-us-gov. No kubeconfig distribution.
  • Inference endpoints in every Region from one command — vLLM, SGLang, Triton — with rolling updates, scaling, canaries, and model weights synced from a central S3 bucket.
  • Outputs that outlive the pod: shared EFS by default, FSx for Lustre and per-Region S3 buckets when you need them.
  • An agent-first front door: gco autopilot launches Claude Code on Amazon Bedrock (gco autopilot --engine codex launches OpenAI Codex and gco autopilot --engine opencode launches OpenCode), grounded by the GCO MCP server and recommended companion MCPs.

The README's Why GCO? has the side-by-side comparison with running clusters yourself, and Core Concepts explains the ideas behind it.

What it costs

A single-Region deployment with default settings carries roughly $210/month of fixed platform cost (EKS control plane, NAT gateways, the internal ALB, Global Accelerator, CloudWatch); GPU instances dominate real spend and scale with usage, and multi-Region deployments scale linearly. Optional add-ons cost nothing until you enable them. The README's sample cost table carries the itemized numbers and their pricing date.

Where to go next

You are… Start with
Ready to run it Get started — the fast path, with the point where billing begins marked
Evaluating it, or planning a real deployment Evaluating & deploying — requirements, costs, customization, upgrades
Wondering what workloads it supports What you can run — schedulers, training, inference, observability
Trying to understand the architecture How it works — the control plane, a Region, a job's path, the security posture
A developer exploring the codebase Repo tour and How we build & test
Ready to contribute or fork Contributing
Looking for the AWS Solutions Library listing Official AWS Solutions Guidance — the overview, benefits, and reference architecture AWS publishes for GCO

Every page here is a short orientation over the authoritative documentation on GitHub — the README and the documentation index — so the deep material is always one click away and never duplicated here. New to Kubernetes itself? The Learning Path adds a primer and role-based tracks. Every trade-off in the project is resolved against ten prioritized tenets, beginning with Protect Workloads, Data, and Accounts.