An AI MVP in 2026 costs roughly $450 to $6,400 per month to run at MVP-1 scale, across five infrastructure lines — model API spend, vector DB, hosting, observability, and on-call. The range is wide because two lines scale with usage and three do not. Founders who get surprised by the bill are the ones who assumed everything scaled the same way. This guide decomposes each line at MVP-1 scale, gives a defensible 2026 monthly range, flags three hidden costs that show up in months 3 to 6, and traces the year-one curve so a non-technical founder can budget the running cost the way a CFO budgets cloud spend.
This is the founder’s explainer that accompanies the AI MVP economics playbook, which decomposes the build line. This piece decomposes the running cost line, and sits within the broader idea-to-product manifesto.
What MVP-1 scale means
MVP-1 scale in this guide means: one workflow live in production, roughly 1,000 user queries per day (about 30,000 per month), a single tenant or small pilot list, no enterprise SSO or compliance overlay, and a small team (one or two engineers plus the founder). This is the post-launch state of a $150K AI MVP — the typical bracket described in how much does an AI MVP cost in 2026.
The bill at MVP-1 is not shaped like the bill at scale. At MVP-1, fixed overhead (hosting baseline, observability baseline, on-call retainer) usually outweighs variable cost (model API tokens, vector DB queries). That inverts between 10,000 and 50,000 daily queries — the point at which model API spend starts to dominate. Most founders budget as if they were already at scale and underestimate fixed overhead. The decomposition below avoids both mistakes by tagging each line flat or variable.
The 5 infrastructure lines, decomposed
Line 1. Model API spend
What it is: the per-request cost paid to OpenAI, Anthropic, Google, or whoever serves the language model. Token-priced — input tokens are cheaper than output tokens, and frontier-tier models (Claude Opus 4.8, GPT-5, Gemini 2.5 Pro) cost roughly 5x to 10x more per token than workhorse-tier models (Claude Sonnet 4.6, GPT-5 Mini, Gemini 2.5 Flash).
Why it varies: spend is a function of which tier you route to, prompt length, response length, and how many model calls happen per user query. A RAG workflow that calls the model once per query at workhorse tier costs a fraction of an agentic workflow that calls the model 5 to 10 times per query at frontier tier.
Defensible 2026 monthly range at MVP-1: $200 to $2,000. Low end: ~30K queries per month at ~1K total tokens per query on workhorse tier. High end: ~5K tokens per query and a mix of frontier and workhorse routing. Public 2026 pricing puts workhorse tier near $0.50 per million input tokens and $2 per million output; frontier tier near $5 input and $20 output.
Flat or variable: variable. Scales linearly with usage.
Line 2. Vector DB
What it is: a database optimized for semantic search. If the product retrieves information from a knowledge base (documents, transcripts, support tickets), it needs one. If it is a pure conversational interface with no retrieval, it does not. Vector databases explained is the founder-readable primer.
Why it varies: vendors charge for two things — storage (size of the index) and queries (how often the index is searched). Pinecone, Weaviate Cloud, Qdrant Cloud, and Turbopuffer all price along that curve. Postgres with pgvector is the cheapest option at MVP-1 and is the default I recommend until there is signal the workload needs a managed vector DB.
Defensible 2026 monthly range at MVP-1: $0 to $300. $0 if pgvector runs on the application Postgres. $50 to $150 for a serverless managed vector DB at MVP-1 volume. $150 to $300 for a pod-based managed offering with higher throughput guarantees. See the AI project storage tax for how storage accumulates.
Flat or variable: hybrid. Storage is flat; queries are variable. At MVP-1, storage dominates.
Line 3. Hosting (application and database)
What it is: the compute and database the product runs on — web tier, application server, operational Postgres, and any background workers. The conservative default is a serverless or platform-as-a-service host (Vercel, Railway, Fly.io, Render) plus a managed Postgres (Supabase, Neon, RDS).
Why it varies: hosting at MVP-1 is dominated by the baseline tier, not the usage. A small Vercel team plus a small Supabase instance is roughly $100 to $250 per month regardless of whether you have 100 or 1,000 users. Once you outgrow the baseline — usually at 5K to 10K daily queries — you pay for compute hours, database storage tiers, and bandwidth above the included quota.
Defensible 2026 monthly range at MVP-1: $150 to $600. Low end: a single small instance and a small managed database. High end: a redundant setup with a separate background-worker tier and a higher-tier database.
Flat or variable: mostly flat. Becomes variable above the MVP-1 envelope.
Line 4. Observability
What it is: the tooling that records what the model did on every request — prompt, response, latency, token count, and any human or eval feedback. Helicone, Langfuse, Langsmith, Arize Phoenix, and Datadog LLM Observability play in this space. Observability is the difference between “the product is broken and we have no idea why” and “the product regressed on the legal-citation eval after we routed traffic to Sonnet 4.6 instead of Opus 4.8; here are the 12 failing traces.” See the AI agency observability stack we install on day one.
Why it varies: vendors price along three axes — number of traces, retention period, and trace depth. At MVP-1, the cost driver is usually retention: 30-day is the default, 90-day is roughly 2x to 3x.
Defensible 2026 monthly range at MVP-1: $100 to $500. Low: free-tier or self-hosted Langfuse plus a small managed Postgres for trace storage. Mid: a Helicone or Langfuse Cloud entry tier. High: Datadog or a similar enterprise-grade vendor with a longer retention window.
Flat or variable: hybrid. A flat subscription with a trace quota plus per-trace overage. At MVP-1, the flat tier is almost always sufficient.
Line 5. On-call
What it is: the cost of someone being available to fix the product when it breaks. For an AI MVP shipped by an agency, this is typically a 2 to 4 week paid retainer ($2K to $4K per week) that ends when the founder takes over or signs a longer support contract. Solo founders absorb it themselves or pay a fractional engineer ($1K to $3K per month).
Defensible 2026 monthly range at MVP-1: $0 to $3,000. $0 if the founder absorbs it; $1K to $3K if a fractional engineer or the build agency provides paid coverage.
Flat or variable: flat. A retainer, not a usage fee.
What scales with usage vs. what is flat
| Line | Type | Why it matters |
|---|---|---|
| Model API spend | Variable | Doubles when usage doubles. The line that overruns at product-market fit. |
| Vector DB | Hybrid (storage flat, queries variable) | At MVP-1, storage dominates. Goes variable past 50K queries per day. |
| Hosting | Mostly flat | The baseline tier holds until 5K to 10K daily queries. Then tiers up. |
| Observability | Hybrid (flat tier + trace overage) | At MVP-1, the flat tier is enough. Overage hits at scale. |
| On-call | Flat | A retainer or zero. Does not scale with usage. |
The founder takeaway: at MVP-1 scale, the bill is mostly flat. The variable line (model API spend) is roughly $200 to $2,000 of a $450 to $6,400 total — typically 20 to 50 percent of the bill at MVP-1, climbing toward 70 to 80 percent at scale. Founders who project “the bill scales linearly with users” overestimate the early months and underestimate the later ones.
The 3 hidden cost lines most founders miss
These three lines do not show up in vendor proposals and surprise founders in months 3 to 6.
1. Re-eval after model updates. Every 8 to 12 weeks a new frontier model ships. The founder either stays on the current model and watches the cost-per-quality curve worsen, or migrates and re-runs the eval suite. A 200-sample re-run with hand-graded sub-criteria takes 6 to 12 engineer-hours plus $100 to $400 in inference. Budget $400 to $1,200 per quarter. Vendor proposals quote the build cost of the eval suite, not the re-run cost.
2. Observability storage growth. Traces accumulate faster than founders expect. At 30K queries per month, a trace is 5 to 20 KB — 150 MB to 600 MB per month. A 90-day retention window means a steady-state of 450 MB to 1.8 GB. Vendors price flat tiers around this envelope; growth above it triggers overage typically between months 4 and 9, and the line jumps 50 to 200 percent.
3. Dev/staging fragmentation. The ranges above assume production infra only. Real teams run three environments — dev, staging, and production. Each has its own version of each line. The realistic baseline is 1.5x to 2x the production line for hosting, vector DB, and observability. Founders forget this because vendor proposals quote one environment.
Add these together and a $150K MVP at MVP-1 should budget roughly $400 to $1,200 per quarter for re-eval cycles, $0 to $1,000 per quarter in observability overages, and a 50 to 100 percent overlay on hosting, vector DB, and observability for dev/staging. That bumps the monthly envelope from $450–$6,400 to roughly $700–$9,000 as the realistic total infra bill.
The year-one cost curve in three phases
The monthly bill is not flat across the year.
Months 1 to 3 — debug-heavy, low usage. Observability spend is at the high end of the range (the team is shipping fixes; traces are read often, retention is generous). Model API spend is at the low end. Hosting baseline is locked in. Envelope: roughly $500 to $4,000 per month.
Months 4 to 6 — first re-eval and model migration. A new SOTA model ships. The team re-runs evals, decides whether to migrate, and typically migrates one workflow. Add $400 to $1,200 of one-time re-eval cost. Envelope: $700 to $5,500 per month, plus the one-time re-eval. This is also when the capex-to-opex shift starts to matter — the team is paying recurring OpEx instead of one-off CapEx.
Months 7 to 12 — operating at growing scale. Model API spend begins to dominate. Hosting tiers up at least once. Observability overage starts to appear. Envelope: $1,500 to $9,000 per month, with model API spend 50 to 70 percent of the bill.
Aggregate year-one infra cost at MVP-1-to-just-past-MVP-1: roughly $15K to $80K on top of the build cost. A $150K MVP with a typical infra envelope lands at $165K to $220K of total year-one spend.
How to sanity-check a vendor proposal
Three checks a non-technical founder can run in under 10 minutes:
- Are all 5 lines named? A proposal that lists “hosting” and “model spend” and stops is hiding the other three. Ask for vector DB, observability, and on-call explicitly.
- Is each line tagged flat or variable? If a vendor quotes one monthly number with no decomposition, the founder cannot model 3x usage. Ask for the variable-line unit rate (per-query model API cost, per-query vector DB cost).
- Are dev/staging environments included? The proposal should state which environments are funded. If only production is funded, the founder will hit the dev/staging gap in month 2 and ask for a change order.
A vendor who can answer all three in writing is pricing on a 2026 cost frame. A vendor who cannot is pricing on a 2018 SaaS frame, and the first cost overrun will be a budget dispute the founder cannot win.
Frequently asked questions
Q: What is the cheapest defensible monthly infra bill for an AI MVP?
A: Roughly $450 per month at MVP-1 — model API at $200, pgvector inside the application Postgres ($0 marginal vector DB cost), a small Vercel-plus-Supabase stack at $150, free-tier or self-hosted Langfuse at $100, and founder-absorbed on-call. This is the floor. Below it, one of the five lines is being skipped, not optimized.
Q: Why does observability cost as much as hosting? A: At MVP-1, observability is the bottleneck on iteration speed, not on serving. Every weekly fix is gated by a replay of failing traces; that replay needs a trace store with enough retention to find the failure. Cutting it below $100 per month means losing the eval-replay capability — see the hidden cost of AI evals.
Q: Can I run an AI MVP without a vector DB?
A: Yes, if the product is purely conversational or the knowledge base fits in the model’s context window on every call. Most workflows that search documents, transcripts, or a support corpus need one — but pgvector inside the application Postgres is a defensible default at MVP-1. Reach for a managed vector DB only when query latency or recall on the Postgres path becomes a bottleneck.
Q: How do I budget for model API spend if I do not know my user volume? A: Use a per-query envelope. Budget $0.02 to $0.10 per query on workhorse tier and $0.10 to $0.50 per query on frontier tier. Multiply by expected queries per month. The single biggest lever is the frontier-vs-workhorse split — evals on a frontier model, production traffic on a workhorse model. That is the most common 3x to 5x cost saving available in 2026.
Q: Is self-hosted infra cheaper at MVP-1? A: Almost never. Self-hosting trades dollars for engineer-hours. The typical $200 to $500 saved per month costs 10 to 20 engineer-hours per month in maintenance. A bad trade unless the founder has a senior engineer with capacity to spare. Self-hosting becomes defensible past $2K per month of managed-vendor spend.
Q: What happens to the bill at 10x MVP-1 scale (10K queries per day)? A: Model API spend goes up roughly 10x (from $200–$2K to $2K–$20K). Vector DB query cost goes up 10x but storage stays flat — $200 to $2K. Hosting tiers up once or twice, $400 to $1,500. Observability tiers up, $300 to $1,200. On-call graduates to a fractional or full-time engineer, $3K to $10K. Total: roughly $5,900 to $34,700 per month. Model API spend has crossed the 50 percent threshold.
Q: When do I need a dedicated platform or DevOps engineer? A: When the monthly infra bill crosses roughly $4K to $5K — around 5K to 10K queries per day. Below that, the founder and the build engineer can manage infra part-time. Above it, configuration drift, observability tier-ups, model migrations, and dev/staging fragmentation consume 10 to 20 hours per week of engineer time — the threshold at which a dedicated platform role pays for itself.
Q: Should I pick the cheapest vendor on every line?
A: No. Pick the cheapest defensible vendor on hosting and observability (commodity lines, moderate lock-in). Pick the most flexible vendor on model API spend (multi-provider via OpenRouter or a thin internal abstraction; this is where lock-in is most expensive). Pick pgvector at MVP-1; reach for managed only when the workload outgrows Postgres.
Where to go next
To sanity-check the build line that sits above this running cost, read the AI MVP economics playbook, which decomposes the $50K-to-$250K build envelope line by line.
For pure price-bracket framing, what does $50K, $100K, $250K buy you in an AI MVP maps the four common brackets to scope. The idea-to-product manifesto is the master read.
For a free weekly note covering this kind of decomposition — vendor pricing changes, year-one cost curves, eval-replay practice — subscribe to the SFAI Labs newsletter. One email per week, written for non-engineer founders shipping AI products.
Dirk Jan van Veen, PhD
