AWS Cost & EC2 Right-Sizing¶
What the cloud lab costs to run — per range and at con scale — and why each node is the size it is. All prices are us-east-1 on-demand, Linux (list, mid-2026); your region/discounts will differ.
Right-sizing rationale¶
The lab looks heavy (31 services across the estate) but almost all of it is lightweight Python mocks. The only real compute is Ollama local inference on dev + ds (~5 tok/s at 2 vCPU); ml runs 13 mock services + a lightweight Ray scheduler; app/attack are proxies/tooling; the k8s node is a single-node k3s. So the sizing is driven by two facts:
- Ollama is CPU-bound → dev + ds use
c6i.xlarge(non-burstable, guaranteed 100% CPU) so sustained multi-attendee inference never hits the t3 CPU-credit cliff. - Everything else is idle-light → ml drops from the original 4×-oversized
t3.xlargetot3.large; app tot3.small; k8s tot3.medium. The attack box ist3.large(8 GB), bumped fromt3.smallto host ~25 concurrent seat shells plus the tool. - ARM/Graviton is blocked — Weaviate and Qdrant ship x86-only binaries and ds needs both — so all nodes are x86. (Graviton would otherwise save ~20%.)
| node | instance | vCPU / RAM | disk | $/hr | why |
|---|---|---|---|---|---|
| ailab-dev | c6i.xlarge |
4 / 8 | 20 GB | 0.170 | Ollama — non-burstable |
| ailab-ml | t3.large |
2 / 8 | 40 GB | 0.083 | 13 mocks + Ray (down from t3.xlarge) |
| ailab-ds | c6i.xlarge |
4 / 8 | 20 GB | 0.170 | Ollama + Weaviate/Qdrant (x86) |
| ailab-k8s | t3.medium |
2 / 4 | 30 GB | 0.042 | single-node k3s "ACME cluster" |
| ailab-app | t3.small |
2 / 2 | 20 GB | 0.021 | LangServe/Streamlit/A2A proxies |
| ailab-attack | t3.large |
2 / 8 | 20 GB | 0.083 | operator tooling + lab-listener + ~25 seat shells |
| per range | 6 nodes | 16 / 38 | 150 GB | ≈ 0.569 compute + 0.015 EBS = 0.58 |
Cost at scale¶
| scenario | cost | note |
|---|---|---|
| 1 range, running | ~$0.58/hr | ~$4.7 for an 8-hr workshop day |
| 5 ranges, running | ~$2.90/hr | ~$23 / 8-hr day; ~$70 / 3-day con (needs quota 80) |
| 5 ranges, left on 24/7 (month) | ~$1,290/mo | the real cost trap — never leave them running |
| 5 ranges, STOPPED between events | $0 compute | pay only EBS ≈ $58/mo (or terminate + keep AMIs ≈ $6–12/mo) |
| testing / non-demo, Spot | ~$0.12/hr/range | stateless nodes tolerate Spot (~65% off) |
The takeaways¶
- At con scale the dollar cost is trivial — ~$16/day for all 5 ranges. The lab is cheap to run.
- The only real cost risk is leaving instances running idle. Compute for a stopped EC2 instance
is $0 — you pay only ~$0.80/mo of EBS per node. So between events, stop the ranges
(
aws ec2 stop-instances); after the con,terraform destroy. - Right-sizing is roughly cost-neutral, not a saving — dropping ml
t3.xlarge→t3.largefunds thec6iupgrade on the two Ollama hosts. You buy predictable performance, not a smaller bill. - The one big lever is the GPU tier — real inference (
g5.xlarge≈ $1.0/hr each) would multiply the bill, which is exactly why it stays an optional tier, off for the talk. - Reset between waves is free —
aws-reset-wave.shrestores from the bakedlab-readyAMIs + re-seeds; the instances keep running, so there's no extra spend to reset the estate for a new wave.