Roger Guerrero Plana

Home / Services / Self-hosted AI

Self-hosted AI and private APIs

Open language models running on your own server or a dedicated GPU, with a private API for your applications. Your data never leaves your infrastructure, and the cost stops growing with every call.

Let's talk about your case See AI audit
When it makes sense

Three reasons not to call someone else's API

The data. This is the main one. If you work with clinical records, a law firm's internal documentation, HR files or anything you cannot justify sending to a third party, self-hosting removes the problem at the root instead of managing it with a contract.

The cost. With an API you pay per use, so your product's success inflates your bill. With your own infrastructure the cost is fixed: past a certain volume, each new user is practically free.

The control. Nobody deprecates your model, changes its behaviour overnight, raises the price or rate-limits you at the worst possible moment. The version you validated is still there tomorrow.

And when it doesn't: if your volume is low and the data isn't sensitive, the API is cheaper and simpler. I say this before you buy anything. Self-hosting because it sounds good is money burned plus one more thing to maintain.
What's included

From the decision to the running service

  • Honest sizing. Which model, which GPU and what it really costs, with the break-even point against API pricing calculated on your numbers.
  • Model selection. By language, task type and licence, which deserves close reading: not everything called open allows unconditional commercial use.
  • Deployment. Inference server, quantisation where it pays off, automatic restart, service isolation and backups.
  • Private API. A clean entry point for your applications, with authentication, rate limits and activity logging.
  • Guardrails. Scope limits, control over what goes in and out, and explicit care with what ends up written in the logs.
  • Monitoring and upgrade plan. Knowing the service is alive, how long it takes, and when to change model without breaking what already works.
Experience

I have run this on my own infrastructure since 2024

The clinical platform I co-founded and operate runs on an open model self-hosted on our own GPU. It is not a proof of concept: healthcare professionals use it and clinical data never leaves controlled infrastructure.

  • Open model in production with 36 institutions using the platform.
  • No clinical data travels to third-party APIs, which is precisely what makes the project viable within its regulatory frame.
  • Real experience with the boring, decisive parts: automatic restart after a power cut, watching disk and logs, updates that don't break validated behaviour.
Questions

Frequently asked

What GPU do I need?

It depends on the model and concurrent usage, not on your company's size. For an internal assistant with a quantised mid-size model, a single professional GPU of 24 to 48 GB usually serves dozens of concurrent users. Sizing comes from measuring your case, not from a table on the internet.

Is it cheaper than the API?

Above a certain volume yes, because cost stops growing with usage. Below it, the API wins. We calculate the break-even with your numbers before you buy anything, and sometimes the conclusion is that it isn't worth it.

Can I do it without buying hardware?

Yes. A rented GPU at a European provider gives you the same data isolation with no upfront investment, and it's the sensible way to start. If volume grows, then you consider buying.

Who maintains it?

Delivered documented and monitored for your team to run. If you prefer me to run it, a monthly arrangement covers updates and incident response.

Does this make us GDPR compliant?

It removes the third-party transfer, usually the hardest part to justify, but it is not an automatic stamp. Legal basis, access control, minimisation, retention and log hygiene still apply. It does simplify the analysis a lot.

Related services

RAG to production Getting the model to answer on your documents, citing sources and inventing nothing.
AI system audit Measuring whether it answers well before your users find out it doesn't.
Case studies Healthcare, legal and emotional wellbeing: AI in production with sensitive data.

Can't your data leave the building?

Tell me what you want to solve and with what information. I'll tell you whether self-hosting pays off for you, what it would take and what it really costs.

Email me[email protected] LinkedInroger-guerrero-plana