Case study · Moyai

easterisland,
anywhere.

€79a month, HA
3code edits
0AWS-only services
7/7phases live
Before

16 services.
2 you can't leave.

AWS eu-west-1 SaaS Locked in
computeECS Fargate
edgeLoad balancer
edgeWAF
edgeRoute 53
storageS3
AWS onlyS3 Tables
storageS3 Files
databaseDynamoDB
networkVPC
logsCloudWatch
identityIAM
databasePlanetScale
hard-wiredNebius
authWorkOS
analyticsPostHog
frontendVercel
From Moyai's Terraform02
The change

Zero rewrites.

3edits to the app
+86 −6
trace_lake/config.py - table_bucket_arn: str + rest_uri: str | None # any Iceberg REST duckdb_connection_manager.py + if rest_uri: attach_rest_catalog() pipeline_worker.py + LLM_BASE_URL # any OpenAI-compatible API
Unset the variables and it runs exactly as before03
Now on Ankra

Open source. Any cloud.

1
networkingTraefik
postgresCloudNativePG operator
2
monitoringPrometheus
3
dataSeaweedFS · Lakekeeper
llmllama.cpp
4
easterislandapp · Postgres × 3
harborregistry
5
observatorydemo UI
8 stacks · 5 waves · all in Git04
Architecture

How it fits together.

ANKRA CLOUD · eu-central-production · se-sto1 3 × 4 vCPU / 8 GB Usersbrowser Developergit push Ankra edgeIPv4 · WAF · DNS Ankra Pipelinesbuild · scan GitOps engineMoyai repo Traefikcert-manager · TLS observatory-demopublic UI easterislandAPI · no public URL Harborimage registry Postgres × 3CloudNativePG · HA SeaweedFSS3 API LakekeeperIceberg REST llama.cppOpenAI-compatible PrometheusServiceMonitors HTTPS tunnel GET /internal sync image swap via LLM_BASE_URL → Nebius · vLLM
Every box is a stack in Git05
Cost, production with HA

5× less.

AWS + vendors
$510
a month, before tokens and SaaS
Ankra Cloud, 3 nodes
€79
3-instance HA Postgres, S3, Iceberg and metrics included
Supabase Pro covers only Postgres, from $25 + compute06
Inference

Your own GPU is one URL away.

Demo

CPU

€0
Vendor

Nebius

per token
Agnostic

Rented GPU

~€214 / mo
Self-hosted

Own GPU

data stays
~1B tokens / monthis where your own GPU wins.
Same stacks, priced by Ankra

Any provider. Same YAML.

Hetzner€543 × cx33
Ankra Cloud€793 × cloud-native-4c-8g
Scaleway€1433 × DEV1-L
DigitalOcean€2033 × s-4vcpu-8gb
UpCloud€2113 × 4 vCPU / 8 GB
OVHcloud€3703 × c3-8
AWS€3923 × t3a.xlarge
GCP€4833 × e2-standard-4
3 × 4 vCPU / 8 GB nodes for HA Postgres + control plane · €/month · list prices08
On-prem

Your rack is config, not code.

01

All open source

02

No cloud APIs

03

Inference is one URL

04

Proxmox · Morpheus

Per customer: mirror images and models, internal CA, GPU choice. HA Postgres already runs.

Same stacks, same Git09
The path

Prove. Scale. Own.

Try the live demo
01

€79 a month

HA, on Ankra Cloud

02

Any provider

priced side by side before you move

03

Your hardware

and your GPUs

ankra.ai
Sources

How the numbers were made.

Before. Services from Moyai's Terraform (infra/). List prices, not Moyai's invoice.

AWS, $510. Fargate 2 × 4 vCPU / 16 GB in eu-west-1 at $0.04556 per vCPU-hour and $0.004997 per GB-hour (~$383, third-party listing; see AWS), PlanetScale HA ~$50–90 (pricing), load balancer, WAF, VPC and logs ~$50–65.

Ankra Cloud, €79. The live cluster after the HA upgrade on 17 Oct 2026: 3 × cloud-native-4c-8g workers (€47.97), one per Postgres instance, plus control plane (€23.97) and network (€6.98), at Ankra Cloud list prices.

Provider bars. Ankra's Compare Providers estimate for this cluster at 3 × 4 vCPU / 8 GB, like-for-like, 17 Oct 2026: nodes, control plane and network at each provider's list price. AWS in us-east-1, GCP in europe-west1; Scaleway excludes network.

Supabase. Pro $25 plus compute add-ons, per 2026 pricing guides.

Nebius. Per-token rates from AMCompute and Requesty; ~1B tokens/month assumes ~$0.20 per 1M blended.

GPU. Hetzner GEX45 at ~€214 a month (third-party report).

easterisland on Ankra11