Quick answer
Cross-referencing published 2026 breakdowns with what we see in founder builds, the all-in monthly infrastructure bands for a typical AI SaaS look like this:
- ~1,000 users: roughly $30–$150/month. Most services sit in free tiers; the LLM API is your only real variable cost.
- ~10,000 users: roughly $450–$1,300/month lean, up to ~$3,000 if you picked premium everything. Free tiers end here and architecture sins start billing you.
- ~100,000 users: roughly $4,000–$14,500/month. The spread between disciplined and undisciplined stacks is now a salary.
The single biggest lever at every scale is the same: LLM API spend is 40–80% of the bill, and it is set by two decisions — which model handles which task, and whether you cache.
Last updated: August 2026. Every figure below is directional, compiled from public pricing pages and published teardowns — re-verify against current rate cards before budgeting.
The line items
| Layer | Lean default | Typical cost (10K users) | Premium alternative |
|---|---|---|---|
| LLM API | Small/nano-class model + routing | $70–$150 | Frontier model for everything: $900–$1,800 |
| Vector DB | pgvector on existing Postgres | $0–$25 | Managed serverless (Pinecone-class): $15–$210 |
| Hosting / compute | Cloudflare / small VPS | $10–$30 | Hyperscaler default config: $30–$120 |
| Auth | Supabase Auth or self-managed | $0–$25 | Per-MAU providers: $275+ at 10K users |
| Observability | Langfuse/Helicone free tier → paid | $0–$50 | Full paid stack: $100–$200 |
| Payments | Stripe (% of revenue, not fixed) | ~2.9% + fees | Merchant-of-record: ~5% |
Payments deserve one honest note: on percentage fees, that line scales with revenue rather than users, which is why it dwarfs infrastructure in some published totals. It is a cost of selling, not a cost of running AI — keep it out of your infra budget mental model.
The three traps that blow up the bill
1. One model for everything. Nano-class models in 2026 cost cents per million tokens; frontier models cost dollars. The published spread is roughly 12x for the same query volume. Classification, extraction, and simple Q&A do not need the frontier model — route them down and save the expensive calls for reasoning chains that earn it.
2. The auth curve. Per-MAU auth providers are free at 1K users and hundreds per month at 10K. It is the steepest pricing curve in the stack, and it arrives exactly when you can least afford surprises. Supabase Auth (generous free tier) or boring self-managed sessions dodge it entirely.
3. No cache. Published teardowns and our own builds agree: a semantic or Redis cache in front of LLM calls cuts token spend 30–50% for products with repetitive query patterns. It is an afternoon of work that pays for itself in the first month at any real volume.
Vector storage: pgvector until proven otherwise
Under ~500K vectors and modest query rates, pgvector on the Postgres you already run costs nothing extra and performs fine with HNSW indexing — published B2B teardowns call it the right answer 90% of the time at that scale. Above that, self-hosted Qdrant on a $10–$20 VPS carries tens of millions of vectors with no per-query fees, while managed serverless options trade money for zero ops. Full comparison: Pinecone vs Weaviate vs Qdrant.
Observability is not optional
Every published production teardown lands on the same lesson: without request logging, latency by query type, and error rates by prompt version, you cannot diagnose cost spikes or catch quality regressions before users do. Free tiers (Langfuse, Helicone) cover early volume; budget $40–$120/month once you have real traffic. Our tooling comparison: LangSmith vs Phoenix vs Braintrust.
The four defaults Cipher applies in founder builds
First-hand, from client work rather than theory — these are the defaults we wire before launch, because retrofitting each of them costs more than building it in:
- Model routing from day one. A cheap classifier decides which model each request deserves. Even a two-tier split (nano for structure, frontier for reasoning) holds the LLM line to the bottom of its band.
- Cache before scale. Semantic caching on the hot paths ships with the MVP, not after the first scary invoice.
- pgvector first. No dedicated vector infrastructure until the eval harness — not vibes — says retrieval quality or latency demands it.
- Founder-owned accounts. Every subscription, API key, and cloud account belongs to the founder from day one. Cost control starts with being able to see the bills.
Stack selection details live in the AI stack for a one-person SaaS, and the wider budgeting picture in tech stacks for scaling AI businesses.
FAQ
What should a solo founder budget for an AI MVP's first six months? Infrastructure alone: $30–$150/month pre-traction is realistic on a lean stack. Your real costs are build time and distribution, not servers.
When do AI products become unprofitable? When per-user token cost exceeds per-user revenue — usually caused by heavy users on flat-rate plans with a frontier model behind every request. Hybrid pricing (base subscription plus usage overage) plus model routing is the standard fix.
Is self-hosting models cheaper than APIs? Almost never below very large volume. A single always-on GPU instance costs more per month than most startups' entire API bill. Revisit only when your monthly token spend rivals a GPU cluster invoice.
Do these bands apply in Australia? Yes with a currency caveat — most AI infrastructure bills in USD, so AUD budgets carry FX exposure. Sydney-region hosting adds little premium; the LLM line is identical.
Who helps set this up properly? Cipher Projects builds launch stacks for solo founders with cost controls wired in from the start — flat rate, on accounts you own.
Conclusion
Running an AI product in 2026 is cheap if you make three decisions early: route models by task, cache aggressively, and keep vectors in Postgres until measurement says otherwise. The bands are knowable, the traps are documented, and the difference between a $450 month and a $3,000 month at the same scale is discipline, not luck.