What It Costs to Run an AI SaaS: Real Numbers for 1,400 Users
Real monthly numbers from HomeoAI, the lean stack behind them, and how to estimate the running cost of your own AI app
Founders planning an AI product usually ask two cost questions: what will it cost to build, and what will it cost to run every month? The second one is often answered with guesswork, or with enterprise cloud bills that have nothing to do with an early-stage product.
HomeoAI, an AI assistant for homeopathic practitioners, gives a real answer. It serves more than 1,400 registered doctors across India, Pakistan, Bangladesh, the US, the UK, Canada, Italy and Australia. On the Codersarts Podcast, its builder Tejesh Vaish shared exactly what it costs to run: under ₹10,000 a month, roughly US$100–120, including LLM usage.
This post breaks down that number, the stack that makes it possible, what makes AI running costs grow, and a simple way to estimate your own. For the full architecture, read How HomeoAI Was Built.
The real monthly cost breakdown
Cost item | What it covers | Monthly cost |
AWS EC2 instance | One virtual machine running the whole app in Docker | About ₹5,000–6,000 |
LLM credits | GPT-4.1 API calls for doctors' queries | About ₹1,000, varying with query volume |
Total | Hosting plus AI usage | Under ₹10,000 (roughly US$100–120) |
Spread across 1,400+ registered users, that is about ₹7 (under US$0.10) per registered user per month. Not every registered user is active every day, so cost per active user is higher, but the order of magnitude is what matters: an AI SaaS at this stage can run for the price of a team lunch.
Tejesh initially considered charging doctors based on their LLM usage, then decided against it because usage costs stayed small. HomeoAI sells subscriptions instead, including yearly plans.
The stack behind the number
The low cost comes from a deliberately simple architecture:
Layer | Choice | Why it keeps costs low |
Interface | Fork of Open WebUI (Svelte, FastAPI, Python) | Chat, login, access control and voice come free with the open-source framework |
Hosting | One AWS EC2 instance | No managed cluster or per-service billing |
Deployment | Docker Compose: front end, back end, databases and Nginx | One command deploys everything to one machine |
App data | SQLite | No separate database server to pay for |
Practice data | Postgres | Runs in a container on the same machine |
Reference search | Chroma vector database | Open source, self-hosted, no per-query fees |
Retrieval service | Python Flask API | Lightweight, runs alongside the app |
Embeddings | Google Gemini embedding model | Chosen for quality at an acceptable cost |
LLM | GPT-4.1 via API | Pay only for queries actually made |
Everything runs on one machine. For a product with a couple of thousand users, Tejesh said a single VM is enough.
The four decisions that kept costs down
1. Reuse instead of rebuild
The biggest saving wasn't on the monthly bill; it was on the build. By forking Open WebUI, Tejesh got chat, authentication, role-based access, user groups, voice input and memory out of the box. That meant months of engineering he didn't have to do or pay for, and fewer moving parts to host.
2. Start on one server, not a cloud architecture
Many teams launch on managed Kubernetes, separate database services, a vector database subscription and a queue system before they have users. Each one adds a monthly bill and operational work. HomeoAI started on one EC2 instance with Docker and still runs there at 1,400+ users.
3. Self-host the open-source pieces
Chroma, SQLite and Postgres run in containers on the same machine. There are no per-seat, per-query or per-gigabyte fees from third-party services.
4. Choose models on quality and cost together
For embeddings, Tejesh tried three options. OpenAI's small embedding model was cheap but missed relevant results. The large model satisfied doctors but was expensive. Google's Gemini embedding model gave doctors the quality they wanted at an acceptable cost. He also added a fast endpoint for quick answers and a deeper one for detailed analysis, so users choose speed or depth per query.
What makes AI running costs grow
The numbers above are for one product at one stage. Your costs depend mainly on these factors:
Query volume: LLM cost rises roughly in line with the number of requests.
Tokens per query: retrieval-heavy apps send long context with every question. More retrieved passages mean more input tokens.
Model choice: larger, newer models usually cost more per token than smaller ones.
Agentic workflows: an agent that plans, calls tools and loops can make many LLM calls for one user request.
Data volume: a larger knowledge base needs more storage, more memory for search and a bigger one-time embedding job.
Concurrency: many simultaneous users may need a bigger machine or more than one.
Abuse: without rate limits and spend caps, a leaked key or fake accounts can multiply your LLM bill.
How to estimate your own AI app's monthly cost
Use this simple formula for the LLM part, which is usually the most variable:
Monthly LLM cost = queries per month × (input tokens × input price + output tokens × output price)
Worked example
Assume, for illustration only:
300 queries a day, so about 9,000 queries a month
4,000 input tokens per query (question, instructions and retrieved passages)
800 output tokens per answer
That gives:
Input: 9,000 × 4,000 = 36 million tokens a month
Output: 9,000 × 800 = 7.2 million tokens a month
Multiply each by your provider's current price per million tokens for the model you choose, and add them together. Prices change often and vary widely between models, so check your provider's pricing page before you plan.
Then add the fixed costs
Hosting: one virtual machine sized for your users
Domain, email and SSL
Backups and monitoring
Payment processing fees on each subscription
Occasional re-embedding when you add or change knowledge, or switch embedding models
And don't forget the biggest one
For most small AI products, the largest ongoing cost is not the cloud bill. It is engineering time: fixing bugs, updating dependencies, improving retrieval, handling security and supporting users. Plan for it from the start.
When to move beyond a single server
A single VM is the right starting point, not a permanent architecture. Signs it's time to scale:
Response times rise at peak hours because CPU or memory is maxed out.
You need high availability, because customers can't accept downtime during maintenance or a server failure.
Enterprise customers ask for data residency, dedicated environments or compliance controls.
The knowledge base outgrows memory and search slows down.
Many concurrent agent workflows make long-running requests that block others.
The usual next steps are a managed database, a separate vector database, a second app server behind a load balancer, and automated backups and monitoring. Add each one when a real need appears, not before.
AI SaaS running cost checklist
Area | Question to ask | Good answer |
Build | Are we building commodity features from scratch? | An open-source base for chat, auth and access control |
Hosting | Do we need a cluster at launch? | One VM with Docker until real load says otherwise |
Databases | Are we paying for managed services we don't need yet? | Self-hosted SQLite, Postgres and vector store at first |
LLM cost | Do we know cost per query? | Measured tokens per query and a monthly estimate |
Model choice | Was the model chosen on quality and cost? | Tested against expert-rated queries, cheapest acceptable option |
Speed vs depth | Does every query need the deepest model? | A fast path for simple queries, a deep path when needed |
Spend control | What caps our worst-case bill? | Hard LLM spending limit, rate limits and alerts |
Pricing | Does pricing cover usage at scale? | Subscription priced above cost per active user, reviewed monthly |
Scaling | What will trigger the next infrastructure step? | Named signals: latency, uptime needs, customer requirements |
Build and run your AI SaaS lean with Codersarts
HomeoAI shows that an AI product can launch and grow on a small, sensible budget when the architecture is chosen carefully. Codersarts helps founders do the same:
AI MVP development: we build your first version on a lean, production-ready stack, reusing proven open-source components where they fit. See https://build.codersarts.com.
Cost and architecture review: we estimate your real running costs, find where you're overspending and plan when to scale.
Managed hosting and operations: deployment, monitoring, backups, updates and incident response for your live app. See https://www.codersarts.com/solutions/operations.
LLM cost optimisation: model selection, prompt and retrieval tuning, caching and spend controls.
Talk to us: https://www.codersarts.com/contact
FAQ
How much does it cost to host an AI SaaS?
It depends on users, query volume and model choice. HomeoAI serves 1,400+ registered users on one AWS EC2 instance for about ₹5,000–6,000 a month in hosting, plus about ₹1,000 in LLM credits, under ₹10,000 (roughly US$100–120) in total.
How much do LLM API calls cost per month?
Multiply your monthly queries by the tokens each query uses and your provider's price per token, separately for input and output. Retrieval-heavy and agentic apps use more tokens per request. HomeoAI spends about ₹1,000 a month on LLM credits.
Can an AI SaaS run on a single server?
Yes, at the early stage. HomeoAI runs its app, databases, vector store and web server in Docker on one EC2 instance. Move to more servers when response times, uptime needs or customer requirements make it necessary.
What is the biggest hidden cost of running an AI app?
Engineering time. Maintenance, security updates, retrieval improvements and user support usually cost more than hosting and LLM credits for a small AI product.
How can I reduce LLM costs in my app?
Use the smallest model that meets your quality bar, send only the most relevant retrieved passages, offer a fast path for simple queries, cache repeated answers where it is safe, and set rate limits and a hard spending cap.
Can Codersarts build an AI MVP on a lean budget?
Yes. Codersarts builds AI MVPs on lean, production-ready stacks, reusing open-source components where they fit, and can manage hosting and operations after launch. Contact us at https://www.codersarts.com/contact.
Related reading:
Want to be a guest on the Codersarts Podcast? Email jitendra@codersarts.com with your name, role and the topic you'd like to talk about.
Comments