A provider-by-provider cost model for the circular-materials Digital Product Passport platform — AWS vs Google Cloud vs Cloudflare across Postgres, storage, egress, compute, backup/DR, and the AI-token menu. Companion to the v1.4 Technical SoW (§5, §9). Pick a provider per line and every tier re-totals.
D1 is SQLite, not Postgres-compatible. Hyperdrive is a free pooler that fronts a Postgres you host elsewhere (Neon, RDS, Cloud SQL). So the database always carries a partner bill — Cloudflare can't own that line, and it isn't offered as a pick.
Scans scale into the millions; at Consumer that's ~$1,200/mo of egress on AWS/GCP versus $0 on Cloudflare — and the same free egress makes backup restore / cross-region replication free. Its weakness: Containers (its only always-on compute) costs ~3× a VM.
Round SoW-keyed estimates drive every number below. The interactive controls sit where they act — the provider picker and AI model/switch on the Stack picker rows, the egress payload in Network egress, the DR standby in Backup & DR.
Provider per line — the picker on each row of Stack picker. ★ = the recommended default: a two-provider stack (AWS + Cloudflare), not three.
AI model & on/off — on the AI assistant row of the builder; switching the model shows how sharply cost scales — Flash → Opus is ~12×.
Egress payload — the slider in the Network egress section.
DR standby — the toggle in the Backup & DR section.
Round order-of-magnitude estimates (~1k batches/org·mo, ~10 scans/batch, ~5 MB/batch, ~5M AI-tokens/org·mo), scaled 10× per tier. Pilot volume is an open SoW item; storage is cumulative (climbs over a product's life); AI-tokens/org is the softest input. Treat as a frame to challenge, not a forecast.
Cheapest provider per tier is highlighted. Use these to see the spread; the Stack picker section lets you pick one per line and see the total move.
| Provider | Pilot | Early | Growth | Consumer |
|---|
Cloudflare is not a row — it has no managed Postgres (Hyperdrive is a free pooler fronting an external PG, billed ≈ the AWS column). Between the two shown: GCP runs ~40% higher at Consumer — Cloud SQL SSD storage is $0.22–0.34/GB vs AWS gp3 $0.115.
| Provider | Pilot | Early | Growth | Consumer |
|---|
R2 is cheapest at every tier (~35% under S3) — and it's the same bucket that serves egress for free.
| Provider | Pilot | Early | Growth | Consumer |
|---|
The decisive line at scale. CloudFront's 1 TB/mo free tier holds AWS at $0 through Growth; GCP's CDN bills from 1 GB. Cloudflare is $0 everywhere — the advantage bites once reads pass ~1 TB/mo, i.e. at Consumer.
| Provider | Pilot | Early | Growth | Consumer |
|---|
Cloudflare Containers is ~3× a VM — it bills per-10ms of active runtime, built
for bursty jobs, not a 24/7 process. GCP e2 is ~10–15% cheaper than AWS Graviton. The table above
compares the three clouds; the Stack picker adds a fourth compute option — AWS Fargate
(the chosen target). Fargate carries a ~1.3–1.5× premium over on-demand EC2 (Growth: ~$340 vs
~$224; and ~2–3× vs reserved EC2) — but reserved-EC2 economics need steady committed
capacity, the opposite of one-command ephemeral {env}.rezethub.dev environments.
Fargate's scale-to-zero, zero-patching, per-task model is that sweet spot; the premium (~$120/mo at
Growth) is repaid in ops. Move prod's always-on node to reserved-EC2 later if Consumer-scale
compute cost bites — a config change, non-prod stays on Fargate.
| Provider | Pilot | Early | Growth | Consumer |
|---|
Managed Postgres already does the daily backup. RDS / Cloud SQL include automated daily snapshots + point-in-time recovery, free up to 1× the DB size — you don't need R2 for daily backup. This line is the DR delta: surviving a whole region or provider. AWS/GCP = native cross-region snapshot copy (inter-region transfer ~$0.02/GB). Cloudflare = a cross-provider dump to R2 — $0.01/GB to store and $0 to restore, but it pays the DB cloud's egress (~$0.09/GB) to export the data out. So R2 is the right target when the DB is self-managed (today's VPS — no managed backup, S3-compatible, free restore) or when you specifically want off-AWS resilience; on RDS it's an add-on for provider-independence, not a replacement for RDS's own daily backup. Backups vs failover: object-storage backups give a low RPO but restore in hours (RTO); the hot standby toggled above is the fast-failover layer — a second running replica, promotable in minutes, added as its own DR standby line (~0.8× the DB) in the stack below.
AI is a per-token choice, not an AWS/GCP/CF line — the model dropdown sets the provider (Gemini→Vertex, GPT→OpenAI, Claude→Anthropic). The code default is Gemini 2.5 Flash. Rates per 1M tokens; blended assumes an input-heavy 4:1 assistant. Prompt-caching cuts cached input to ~10%; batch paths are 50% off.
| Model | Input /1M | Output /1M | Blended /1M | Pilot | Early | Growth | Consumer |
|---|
Why this matters commercially: expose Flash as the margin-safe default and offer premium models as a priced upsell — never as the baseline. Flash → Opus is $3.7k vs $45k/mo at Consumer. GPT-5 is the model named in scope; OpenAI's live flagship has since moved to the 5.4/5.5/5.6 line.
Pick a provider for each line; ★ marks the cheapest option. The default stack is RezetHub's actual choice: Fargate compute + AWS database, Cloudflare for storage, egress and backup, and Datadog observability. Toggle the assistant and switch its model right on the AI assistant row. The sticky total and contrast cards follow your picks; picking AWS/GCP for egress or backup also makes the payload lever move the total.
Observability = Datadog (OTLP): APM traces, infra metrics, logs and RUM in one portal — "what is our infrastructure doing." Estimated at ~$50/host (APM + infra) plus log ingest and RUM; the swing factors are log retention and RUM session volume, so the public passport is sampled, not full-fidelity. Set the line to None to see the total without a portal. (Instrumentation is vendor-neutral OTLP, so the portal is a config swap; the previous error-only tracker is retired in favour of this.)
| Line · provider | Pilot | Early | Growth | Consumer |
|---|
The recommended picks save ~30–40% vs committing to one hyperscaler at Consumer scale — almost entirely from R2's $0 egress and cheaper storage/DR. Sanity check: Pilot production ≈ $65/mo (~A$100); across prod + reduced staging + small test that lands inside the SoW's A$300–900/mo three-environment envelope.