A Global Bank Put Its AI Back Inside Its Own Walls

Fraud detection, a relationship-manager copilot, and risk models all worked in the lab and stalled at the same production gate: customer PII on shared infrastructure, GPU costs that made every invoice a surprise, cross-region latency that made real-time fraud scoring inconsistent, and a fragmented platform with no shared operational view. The bank moved the entire estate in-house on UPC AI Factory — operated by UnitedLayer, controlled by the bank — in twelve weeks. Year one: 42% lower TCO against the prior hyperscaler setup, 40% faster AI training, sub-100ms p95 inference latency, 100% of AI data and models inside the tenant boundary, and a fixed monthly cost finance could finally forecast.