Beyond the CPU Ceiling: NCBA Group’s Shift to an AI Factory Model

Industrializing Mobile Banking Intelligence

Customer: NCBA Group

Industry: Financial Services (Digital Banking)

Region: East Africa

Impact: 81% reduction in data prep time; 2.7x acceleration in behavioral scoring.

The Challenge:  The Infrastructure Ceiling

As a leader in African mobile banking, NCBA Group manages data-intensive workloads for a customer base exceeding 60 million users. Their legacy infrastructure, built on serial, CPU-bound Oracle databases, reached a physical performance wall.

The CPU-era data pipelines were fundamentally incapable of processing transactional volumes at the speed required for modern digital banking. This created an Impedance Incompatibility between their data storage and their operational ambitions, resulting in:

• The 6 AM Deadline Failure: Critical nightly ETL jobs prepared data for noon instead of the 6 AM operational deadline. This forced loan approval teams to operate on stale data, creating significant credit exposure and business risk.

• Operational Blind Spots: High-complexity reports and heavy JOIN queries frequently failed to complete on the legacy architecture, leaving the business without the visibility required for critical decision-making.

• Innovation Stagnation: The engineering team was overloaded with basic data logistics, leaving limited bandwidth for new initiatives such as AI models.

The Solution: The AI Production Layer

NCBA required a fundamental shift from a "prototype" mindset to an AI Factory model. They bypassed further expansion of their legacy 22-server Hadoop cluster —and implemented SCAILIUM as their AI Production Layer.

SCAILIUM provided a GPU-native dataflow engine that unified their disparate transactional sources into a continuous production environment.

The Industrial Backbone Specs:

• Infrastructure Consolidation: Consolidated a 22-server Hadoop environment into two SCAILIUM nodes.

• Compute Power: Each node equipped with 4x A100 80GB GPUs.

• Physics-Aligned Architecture: Leveraged multi-level GPU parallelization to eliminate the "Serialization Tax" inherent in CPU-only pipelines.

The Results: Intelligence at Silicon Speed

By moving ingestion, transformation, and curation directly onto the GPU, NCBA eliminated silicon starvation and achieved industrial-scale performance.

The Bottom Line

"It’s all about performance - being able to provide a positive experience while reducing time to insight," says Leslie Chemwolo, Head of Data Engineering and Infrastructure.

By establishing this high-velocity dataflow, NCBA is currently implementing SCAILIUM’s MCP (Model Context Protocol). This will bridge the gap from high-speed reporting to a fully industrial AI Production environment.