Private AI · Client result
From €5,600 to €13 a month.
A client of mine paid per call for machine translation. The bill had reached €5,600 a month and was still climbing, so the team rationed the work to keep it down.
I moved translation and three other AI workloads onto infrastructure the client controls. The translation bill is now €13 a month.
01 / Sound familiar?
Where paying per call starts to hurt.
The bill grows with every call.
Translation, classification, embeddings: every item you process is another line on the invoice. More usage means more cost, with no ceiling.
The team rations the work.
Material gets skipped, batched or delayed because of the bill or a rate limit. That lost work has a value, even if no invoice shows it.
Some data should not leave.
Your contracts limit who may process your customers’ data, or it belongs to a client or patient. Every API call makes another copy outside.
Nobody knows if moving pays off.
Running models yourself is not cheaper in general. It depends on the workload, and sometimes the right answer is to stay on the API.
02 / The result
What changed for the client.
A defence technology company processes large volumes of multilingual material. Every item runs through several models for translation, classification, embeddings and image analysis. I built a private inference gateway that now runs all of them on Hetzner servers with NVIDIA RTX PRO 6000 GPUs.
Cost no longer rises with every call.
Hosting is a fixed monthly amount, not a price per call. Capacity grows by adding servers.
Rationing ended.
The company now processes the material it used to skip because of the price.
Sensitive data stays inside.
The models run on infrastructure the company controls. Requests go to an outside API only through the fallback.
First year: €56,800 private vs €313,000 to €798,000 on APIs
Private: hosting €36,000, build €10,000 and 180 hours of upkeep. The API range prices the client’s quietest and busiest weeks at public list rates with caching, converted from US dollars at the ECB rate of 28 October 2026.
03 / 30-second check
Would this work for your workload?
Pick the AI workload whose bill worries you most. Seven quick answers give you a verdict and a rough cost comparison. Your answers stay in your browser.
04 / Free review
Want me to check your setup for free? Sign up for a spot.
I run these reviews alongside client work, so spots are limited.
- We go through three questions for your workload: can a model you run do the job, is there a reason to move, and can you carry it.
- You leave knowing whether private inference is worth pursuing, and what a proper test would involve.
- If the answer is to stay on the API, that is what I tell you.
Nando Theessen
You talk to me directly, and I do the review myself.
Thanks. One step left.
Pick a time in the calendar tab that just opened to confirm your spot. If it did not open, use the button below.
Prefer to write? nando@devwire.technology
Benchmarking with your data, implementation and ongoing operation are paid work, scoped once we both know the workload fits.
05 / FAQ
Before you sign up
If your question is missing, ask it in the review or by email.
When should I stay on the API?
When no model you can run does the job well enough, when usage comes in short bursts, or when nobody can own the service. Without high, steady volume, rationing or a restriction on the data, the API is usually the better deal.
Which workloads does this fit?
Bounded tasks your applications call over and over: translation, classification, document extraction, embeddings for search, and image or speech processing. General assistants for staff, coding tools and autonomous agents are a different decision with different trade-offs.
How rough is the cost comparison?
Rough. It sets your current API spend against my client’s actual first-year cost. Your own servers, models and peak decide your real number, and a load test at real input sizes settles it.
What should I bring to the review?
A short description of the task with an example input and output, your volumes, what it costs you today, and any limits on where the data may go.