How I Run Production AI for Less Than a Netflix Subscription

There’s a persistent myth in the AI space that I am personally committed to dismantling: that production AI is expensive.

It’s not. The bill for my entire AI infrastructure stack — the thing that runs real client workloads, real embedding pipelines, real LLM orchestration — is $12 a month. That’s less than Netflix Standard. Less than a Spotify family plan. Roughly the cost of two cups of coffee at an airport.

The expensive part isn’t the infrastructure. The expensive part is bad architecture that scales the wrong things — usually because someone started with a managed SaaS platform, got comfortable, and never asked what they were actually paying for.

Let me show you what the cheap version looks like, because cheap and production-grade are not mutually exclusive.

The Myth: AI Needs Expensive Infrastructure

This myth persists partly because enterprise AI vendors benefit from it. If you believe that AI requires specialized compute, proprietary databases, and managed platforms with five-figure contracts, you’ll sign up for their products. And some use cases genuinely do require that — serious model training on massive datasets, low-latency inference at millions of requests per day, the stuff that runs at hyperscaler scale.

But most SMB AI use cases? They’re not that. They’re document processing pipelines. RAG-based chatbots over internal knowledge bases. Workflow automation that calls an LLM to classify, summarize, or generate. Structured data extraction from unstructured inputs. None of that requires a datacenter or a cloud AI platform.

What it requires is a reasonably sized server, a well-chosen database, and code you actually control.

The Stack

Here’s what I run, in full.

Compute: Hetzner Cloud VPS

Hetzner is a German cloud provider that is absurdly affordable by American standards. A CX31 instance — 2 vCPU, 8GB RAM, 80GB SSD — runs €10.90/month, roughly $12 at current exchange rates. That’s the entire foundation. It’s more than enough for most AI workloads that don’t involve running models locally. (For LLM inference, I’m calling the OpenAI or Anthropic API — more on that math below.)

If you want a US-based provider with similar pricing, Hetzner has US data centers now, or you can look at Vultr and Linode. The point is: commodity VPS hosting is dirt cheap and completely adequate for most AI applications.

Container Runtime: Docker + Docker Compose

Every service runs in a Docker container. This is non-negotiable for me — it means the entire stack is reproducible, portable, and version-controlled. If I need to move to a different server, I copy a docker-compose.yml and an Ansible playbook and rebuild in 20 minutes. If something breaks, I can roll back by checking out the previous version of the config.

Docker Compose handles local development. When it goes to production, the same config runs on the VPS. No surprises between environments.

Database: PostgreSQL + pgvector

Postgres is the most battle-tested open-source database on the planet, and with the pgvector extension it becomes a vector database capable of storing embeddings and doing similarity search. For 95% of RAG applications, this is all you need. You don’t need Pinecone. You don’t need Weaviate. You don’t need a specialized vector database with its own pricing model and its own lock-in story.

pgvector handles hundreds of thousands of vectors without breaking a sweat on this VPS. It’s fast enough for real-time RAG retrieval at reasonable dataset sizes. And when you outgrow it — which is a good problem to have — migrating to a dedicated vector store is straightforward because your data is in a standard format in a database you control.

Orchestration: LangGraph

LangGraph is the orchestration layer I use for agentic workflows — building pipelines with multiple steps, conditional branching, human-in-the-loop checkpoints, and retry logic. It’s built on top of LangChain but gives you actual graph-based workflow control rather than the chain abstraction that starts to fight you once things get complicated.

For simpler workflows, n8n handles the job — but LangGraph is what I reach for when the workflow needs to branch, loop, handle state, or orchestrate multiple agents with shared memory.

Interface: Discord

Discord has become my default control surface for AI-powered internal tools. The API is excellent, webhook support is built-in, bots are easy to deploy, and almost everyone already has it installed. For approval loops, status notifications, draft reviews, and ad-hoc commands — Discord is fast to build on and surprisingly robust.

This is not glamorous architecture. It’s not a polished SaaS dashboard with a beautiful UI. It’s a private Discord server with purpose-built bots for specific workflows. And it works better than the alternatives I’ve tried, at a fraction of the cost.

The LLM API Math

Since I’m calling OpenAI or Anthropic rather than running models locally, the compute costs are usage-based. Let me give you real numbers.

GPT-4o runs at $2.50 per million input tokens and $10 per million output tokens. For a typical RAG query — say 2,000 token context window plus a 500-token response — that’s roughly $0.005 per query. At 1,000 queries a day, that’s $5/day or $150/month. At 10,000 queries a day, it’s $1,500/month, which is where you start thinking harder about model selection and caching.

For most SMB applications, 1,000 queries per day is on the high end. Most internal AI tools are running 100-300 queries per day. At that volume, your LLM API bill is $15-45/month. Add $12 for the VPS. You’re at $60/month for a production AI system.

Claude Sonnet is comparable in price and often better at reasoning tasks. For embedding models, OpenAI’s text-embedding-3-small runs at $0.02 per million tokens — cheap enough to essentially ignore in your cost model until you’re processing millions of documents.

The total infrastructure cost to get a real RAG-based AI application running in production: somewhere between $25 and $150/month depending on usage. Not $2,000/month. Not a five-figure enterprise contract.

What You’re Actually Paying for with AI SaaS

When you pay $500/month for a managed AI platform, here’s what you’re paying for that isn’t on the Infrastructure bill:

• Their margin on the same Hetzner (or AWS) compute you could rent directly
• Their engineers’ salaries to build the dashboard you’re using
• Their sales and marketing costs, amortized across your account
• A premium on the same LLM API calls that flow through their platform
• And — the critical one — the switching cost moat they’re quietly building

I’m not saying managed platforms have zero value. They abstract away infrastructure complexity, and that’s genuinely useful for teams without engineering resources. But you should know what you’re buying with that premium — and make an intentional choice about whether it’s worth it.

For clients who want AI capabilities without building an engineering capability: I build and hand over the stack. You get a system on infrastructure you own, documented so you could hand it to any engineer, with no ongoing platform dependency. The setup cost is a one-time project. The monthly bill is $12.

That’s the model. And that’s why the expensive thing isn’t the infrastructure — it’s the architecture decisions you make before you touch a server.

If you want to understand the business case for owning your infrastructure rather than renting it, that’s the core argument in Why I Don’t Sell AI SaaS. And if you build on this stack and want to understand how to make the workflows themselves reliable — the guardrails, retry logic, and approval gates — that’s covered in How to Build AI Systems That Don’t Go Rogue.

If you’re currently paying for AI SaaS and you’ve started to wonder what you’re actually getting — or if you want to understand what a self-hosted AI stack would look like for your specific use case — let’s talk.