Leave request
Presentation
Burger logoBurger logo
Cases Private LLM
FINTECH · EU · NDA

A European fintech company cut its LLM costs 55% without sending a single prompt outside its perimeter.

An anonymized production case for Private LLM Deployment & Operations: the workload moved from a public API to a private, OpenAI-compatible deployment operated by DigitalCare.

55%
lower cost per request
2x
faster median response
100%
of prompts inside the client’s infrastructure

Want the same numbers for your workload? Get a cost assessment: anonymized, like this one, if you need it.

Background

The client is a European fintech company in the mid-market scale bracket, processing sensitive customer documents and support conversations with LLMs at production volume.

The client’s name is withheld under NDA; the workload, stack, and results below are from the production engagement.

Problem

Two pains, stated bluntly.

Public API spend scaled linearly with usage: at high tens of millions of tokens per day the monthly bill was on a trajectory toward the mid five figures.

Prompts containing customer PII were routed through a third-party API outside the client’s perimeter, blocking a GDPR data-residency requirement the business needed to close.

Solution

Methodology over adjectives.

Evaluation first. Open models were tested against the incumbent API on the client’s real prompts: 500 test generations, blind-scored by the client’s own reviewers.

Model. Qwen3-32B was selected: quality parity on the client’s task mix at a hardware footprint of a single GPU node. Model names don’t identify the client.

The stack
kubernetesgpu nodesvllmopenai-compatible gatewayprometheus / grafana

The one-line migration. Client code unchanged except base_url: the endpoint is OpenAI-compatible, so SDKs, tooling, and agents kept working on day one.

Private LLM deployment · productionINSIDE THE PERIMETER
Client application trafficOpenAI SDK
Private OpenAI-compatible gateway · keys · quotas
Kubernetes · GPU nodes
vllm-replica-a
qwen3-32b
vllm-replica-b
qwen3-32b
Monitoring planeprometheus · grafana · metatron
Results

The bill at the client’s volume.

  • Cost per request down 55% at high tens of millions of tokens per day.
  • Median response time improved 2x against the public API baseline at ~40 requests/sec peak: the model sits next to the application, with no third-party rate-limit queuing.
  • 100% of prompts and completions inside the client’s infrastructure: the compliance blocker closed.
Monthly bill at the client’s rounded volume (estimates)
ArchitectureCost / 1M tokensMonthly bill
Public API (baseline)~$10.00~$27,000
Private LLM, DigitalCare-operated~$4.20~$11,300
Private LLM + API spillover for peaks~$4.90~$13,200

Quality parity: on the blind-scored evaluation from the Solution section, the selected model matched the incumbent API on the client’s task mix; reviewers preferred neither at a meaningful margin. Figures include GPU infrastructure and managed operations, rounded.

Cost / 1M tokens$10.00Public API$4.20Private LLM

Estimates at the client’s rounded volume; includes infrastructure and managed operations.

What’s next

The engagement continues: more workloads are moving onto the private endpoint, and a smaller model is being evaluated for latency-sensitive paths on the same cluster.

Show us your API bill.

We’ll model self-hosted economics for your real workload and tell you honestly which side of the breakeven you’re on.