Where to run each layer of the POC, why Global Accelerator waits, how cloud credits can zero out the cloud bill, and a running total you can track as the pilot moves along.
| Layer | POC home | Why |
|---|---|---|
| Dev portal (this site) | Cloudflare Pages + Access | Free static hosting, per-user password gate ≤ 50 seats |
| Web UI build (Phase 1) | Cloudflare Pages + Access | Testers click flows in a browser before any store/backend exists |
| Backend + admin + AI | AWS — 2 EC2 boxes | Code is already AWS-shaped; core box always-on, GPU box on-demand |
| Mobile app (device) | TestFlight / Play internal | Real push, camera, biometrics; reviewers validate here |
| Production (later) | AWS — 9-stack CDK | Aurora, ElastiCache, CloudFront, Lambda — when pilot proves it out |
Two instances in one VPC — no NAT, no Aurora, no CloudFront. The core box runs the existing Docker Compose stack; the GPU box runs the AI services and is started only when needed. Runbook: docs/CHEAPEST_DEMO_DEPLOY.md.
| Item | Spec | Cost | Notes |
|---|---|---|---|
| Core box (always-on) | t3.small 24/7 | ~$15/mo | ~$5 with a 1-yr savings plan |
| GPU box — around demos | g5.xlarge spot, ~2h/day | ~$20–30/mo | A10G 24GB; LLM on-GPU ≈ $0.30–0.40/hr |
| GPU box — review week | g5.xlarge spot, 1 wk then stop | ~$50–80 | Keep up during App Review |
| Cheaper GPU option | g4dn.xlarge, LLM via API | ~$0.16/hr | T4 16GB; API for LLM ~halves GPU cost |
| EBS storage | 2× ~40–80 GB gp3 | ~$8–12/mo | — |
| Route 53 / Elastic IP / Cognito / DynamoDB / S3 | demo volume | < $2/mo | Cognito free tier; EIP free while attached |
| POC backend — with GPU start/stop discipline | ~$30–50/mo | 24/7 GPU on spot would be ~$250–320/mo | |
New AWS accounts frequently have 0 vCPU quota for G/VT instances (spot and on-demand). Request an increase to ≥8 vCPUs on day one — it can take 24–48h and is the most common cause of a stalled GPU deploy.
You linked Global Accelerator's pricing. It's a production-scale networking product (static anycast IPs + edge routing for lower latency / multi-region failover). The POC does not need it — a single region with a Route 53 record and Let's Encrypt is enough. Noted here so the number is on record for later:
| Charge | Rate | Notes |
|---|---|---|
| Fixed accelerator fee | $0.025 / hr | ≈ $18/mo per accelerator, always-on |
| Data-transfer premium (DT-Premium) | $0.007–$0.105 / GB | By route; e.g. US↔EU $0.015, US→Australia $0.105; charged on the dominant direction only |
| Public IPv4 | standard IPv4 fee | Per the VPC IPv4 pricing |
Verdict: add at production scale for multi-region / low-latency, not for the POC.
The POC can run at near-zero cash cost by landing startup credits. Two live paths (verified July 2026):
Self-serve Founders Hub — up to $150K Azure credits, no funding requirement, startup must be < 7 years old. Tiers: Ideate $1K → Build $5–25K → Grow $25–50K → Scale $120–150K.
Azure Accelerate adds activity-based funding of $5K–$175K per engagement for assessments / migration / modernization.
Apply: startups.microsoft.com
Credits typically $1K–$100K depending on tier / accelerator affiliation. Since the code is already AWS-shaped, Activate credits map directly onto the demo box + eventual production stacks with zero re-platforming.
Apply: aws.amazon.com/activate
Ship the POC on AWS now — the app is built for it and the cheap-demo path is documented. In parallel, apply for both AWS Activate and Microsoft for Startups credits. Take whichever lands; Activate keeps you on AWS with no rework, and Microsoft credits are worth keeping as a strategic option (and a fallback if you ever want a cost-neutral second region). Don't re-platform to Azure just to chase credits during the POC.
The one-time / annual cost of getting on the stores, and the brand-lock rules that decide what you must freeze first.
| Tier | Cost | What it gets you |
|---|---|---|
| Apple Account (free) | $0 | On-device dev testing only — 10 App IDs / 3 devices, all 7-day expiry. Good for the earliest sanity check. |
| Apple Developer Program | $99 / yr | TestFlight (up to 10,000 external testers), App Store Connect, App Store distribution. Individual or org (org needs D-U-N-S). |
| Apple Developer Enterprise | $299 / yr | Internal-only distribution to your own employees. Not for public/pilot testers — not what we want. |
| Google Play Developer | $25 one-time | Play Console, internal/closed/open testing tracks, Play Store distribution + identity verification. |
| POC store entry | $124 | Apple $99/yr + Google $25 one-time |
google-services.json — push, bound to packageName or logo change → a normal ~1–3 day App Review / Play graphic-review pass on a new build. Bundle-ID / package change → a brand-new app on both stores: new listing, new review, and no way to migrate users, ratings, or history. Lock display name + bundle ID (com.mifamilias.app) + logo before generating any signing or push keys.
What it costs to get through the pilot, split one-time vs. monthly. Credits (above) can offset most of the AWS line.
| Phase | One-time | Monthly (cash) | Monthly (with credits) |
|---|---|---|---|
| Dev portal + Web UI test (Cloudflare) | — | $0 | $0 |
| Apple + Google accounts | $124 | — | $8 |
| AWS demo backend (GPU start/stop) | — | $30–50 | ~$0 (credits) |
| Domain (if not owned) | ~$12/yr | ~$1 | ~$1 |
| POC + pilot | ~$136 | ~$40–60 / mo | ~$10 / mo |
The full 9-stack CDK runs ~$90–100/mo idle (Aurora ~$43, NAT ~$32, ElastiCache ~$12, CloudWatch ~$3, rest pay-per-use). Dev-mode levers cut this materially: NAT instances (~$5 vs $32), Aurora Serverless scaling to 0 idle, CloudFront PRICE_CLASS_100. Don't stand this up until the pilot justifies it — the two-box demo is enough to launch and validate.
Pricing captured July 2026 from public AWS / Apple / Microsoft pages — reconfirm before committing spend.