Self-Hosted LLMs: Full AI Capabilities Without Cloud API Costs

The Problem With Cloud AI APIs

Cloud AI APIs work well for prototyping. For production use at any real volume, the economics stop making sense quickly. Every query costs money. Every document sent to an external model leaves your infrastructure. Every API key is a dependency on a vendor's pricing, availability, and terms.

• Per-token costs scale directly with usage, so the more value you get, the more you pay
• Sensitive business data, customer records, and internal documents pass through third-party infrastructure
• Rate limits and outages in a vendor's API become your production incidents
• Model versions change or deprecate on the vendor's schedule, not yours
• No ability to fine-tune or customize behavior without expensive enterprise contracts

Self-hosted models solve all of these. The tradeoff used to be quality. That gap has closed significantly. Modern open-weight models run well on modest hardware and handle the majority of business use cases without touching a cloud API.

chat.steelcitysolutions.io

We run our own AI chat platform at chat.steelcitysolutions.io, built on Open WebUI backed by locally hosted models. It serves as both our internal tooling and a live reference deployment we use when scoping similar work for clients.

Real Example: Receipt Scanning in the Fairway H&C App

The Fairway H&C dispatch platform includes receipt tracking tied directly to jobs. Technicians upload a photo of a receipt from the field. A vision-language model reads the image and extracts the relevant data automatically.

1. 1
   Technician snaps a photo of the receipt

   No special hardware, no scanner. A phone camera is all that's needed. The image is uploaded directly through the job detail page in the web app.

2. 2
   Vision-language model reads the image

   The uploaded image is passed to a locally hosted vision-language model. The model reads the receipt and extracts vendor name, date, line items, and total. No cloud API call, no external service, no per-image fee.

3. 3
   Structured data is written to the job record

   The extracted fields are formatted and saved directly to the job. The receipt is stored as an attached image alongside the parsed data. The job record now has a complete, searchable expense entry with zero manual data entry.

4. 4
   Available for review, audit, and reporting

   The receipt image and its parsed data stay attached to the job permanently. Admins can review, export, or audit any receipt at any time. The original image is always available to verify the extracted data against.

The result: a technician spends three seconds taking a photo. The system handles everything else. No manual entry, no lost receipts, no end-of-month scramble to reconcile expenses.

What This Can Be Applied To

Receipt scanning is one example of a broader pattern: take something a human is currently doing by reading a document and typing into a system, and replace the reading and typing with a vision or language model. The human takes the photo or uploads the file. The model does the rest.

• Invoices parsed and entered into accounting systems automatically
• Inspection reports photographed in the field and filed with extracted findings
• Handwritten forms digitized without manual transcription
• Email threads summarized and routed to the right team automatically
• Internal knowledge bases made searchable through document-grounded chat (RAG)
• Customer support drafted automatically from ticket history and product docs
• Meeting transcripts summarized into action items without a human reviewing the recording

How a Deployment Is Built

A self-hosted AI deployment for a business typically involves three components: the model server, the interface or integration layer, and the application logic that connects it to the workflow.

Cost Comparison

The economics shift significantly at moderate usage. A self-hosted deployment has a fixed infrastructure cost. A cloud API has a variable cost that grows with every query.

• A small VPS with 32GB RAM runs around $40 to $80 per month and handles hundreds of queries per day without issue
• Processing the same volume through a cloud API typically costs several hundred dollars per month at modest scale
• At high volume, the difference becomes an order of magnitude
• For vision tasks like receipt scanning, cloud vision APIs charge per image. A self-hosted VLM has no per-image cost
• The fixed cost stays the same whether you run 100 queries a day or 10,000

Cloud APIs still make sense for low-volume or irregular workloads where standing up dedicated infrastructure isn't justified. The right answer depends on your actual usage pattern. We scope both options honestly before recommending one.