BC Card puts a price on enterprise AI inference architecture

Red Hat has published unusually concrete operating results from BC Card’s move away from commercial large-language-model APIs toward an in-house inference platform. The Korean payment provider says the new system reduced GPU inference costs by as much as 70% and improved inference speed by two to three times.

Those are company-reported figures in an awards announcement, not an independent benchmark. They are still useful because the announcement identifies the architecture behind them: Red Hat OpenShift AI, Neural Magic’s vLLM compression technology, FP8 dynamic quantization and a multi-model design that BC Card calls an “SLM Swarm.”

What changed

BC Card established an AI Transformation team in early 2025 and has since expanded its portfolio to more than 10 commercial AI services. Red Hat says growing GPU demand and the cost of commercial model APIs pushed the company toward an enterprise platform it could operate directly.

The resulting agent-based system runs on OpenShift AI and uses compression and lower-precision inference to make GPU capacity go further. BC Card also reports that AI-driven automation saves an estimated 102,000 operational hours per year.

The deployment is not only an internal infrastructure story. Red Hat says BC Card has published 38 models and five datasets on Hugging Face, where they receive roughly 100,000 downloads each month. That gives the company a public distribution path alongside the private serving platform.

Who should care

Platform teams evaluating managed model APIs against self-hosted inference now have a deployment pattern with named optimization techniques and reported business outcomes. The important detail is that the savings are attributed to a system, not merely a model swap: orchestration on OpenShift AI, vLLM-based compression, FP8 quantization and multiple smaller models are presented as one operating design.

The announcement also provides a second infrastructure reference. KEPCO KDN says it used Red Hat Enterprise Linux, OpenShift and Red Hat Developer Hub to replace manual SCADA deployment processes with CI/CD, cutting deployment time from days to minutes. That project targets power-grid control rather than AI inference, but it reinforces the same platform argument: operational standardization is the deliverable.

What to do

Teams considering a similar inference build should treat BC Card’s percentages as a case-study claim to validate, not a portable forecast. A useful evaluation would measure the current API baseline, model-quality changes after compression and FP8 quantization, accelerator utilization, latency under representative load and the labor required to run the platform.

The practical takeaway is narrower and stronger than the award language: cost comparisons between hosted APIs and self-managed inference need to include the full serving architecture. BC Card’s reported result suggests that precision, compression and model-selection policy can matter as much as the underlying GPU.