When a customer sends KES 50,000 via M-Pesa at 2 AM from a location they have never used before, and the transaction is completed in under two seconds, there is a good chance that a machine learning model has already analysed the transaction in real time, checking the amount against the customer's typical patterns, the location against their known locations, the recipient against their transaction history, and decided that the transaction is legitimate. This analysis happens in milliseconds, on servers in a data centre, before the customer even sees the confirmation message.
Machine learning (ML) is transforming Kenya's banking and financial technology sector from the inside. It is not a customer-facing feature that users see or interact with directly. Instead, it operates behind the scenes (in fraud detection systems, credit scoring engines, customer analytics platforms, and operational automation tools) making decisions that affect millions of transactions every day. And every one of these ML systems runs on servers in data centres, creating a growing and often overlooked demand for data centre infrastructure.
The ML Landscape in Kenyan Financial Services
Kenya's financial services sector is one of the most technologically advanced in Africa, and machine learning adoption reflects this. The sector's ML deployment can be categorised into several use cases, each with different infrastructure requirements.
Fraud Detection and Prevention
Fraud detection is the most mature and widely deployed ML application in Kenyan banking. Every bank and mobile money operator in Kenya uses ML models to analyse transactions in real time, looking for patterns that indicate fraud, money laundering, or account takeovers.
The ML models used for fraud detection are typically trained on historical transaction data, millions or billions of past transactions labelled as legitimate or fraudulent. The model learns patterns: a customer who normally sends KES 1,000-5,000 to family members suddenly sending KES 500,000 to an unknown account at 3 AM triggers a high-risk score. A new device logging into an account from an unusual location triggers additional authentication.
Real-time fraud detection requires ML models to score every transaction in milliseconds, the model must return a risk score before the transaction is approved or declined. This requirement for low-latency inference means the models must run on servers with fast processors, preferably in data centres close to the banking systems that generate the transactions. Hosting fraud detection models on cloud servers in South Africa or Europe adds 50-200ms of latency, which may be acceptable for batch processing but is problematic for real-time transaction scoring.
Credit Scoring and Lending
Credit scoring is the ML application with the most transformative potential for Kenya's financial inclusion. Traditional credit scoring relies on formal financial records (bank statements, loan repayment history, employment verification) that exclude millions of Kenyans who are unbanked or underbanked.
ML-based credit scoring uses alternative data sources to assess creditworthiness. M-Pesa transaction history (frequency of transactions, amounts, consistency of income flows), airtime purchase patterns, utility payment records, social media activity (with consent), and device data can all be used to build credit models that serve people who have no formal credit history.
Several Kenyan fintechs have built their businesses around this approach. Companies like Tala, Branch, and Silba have disbursed millions of small loans to previously unbanked Kenyans, using ML models that analyse mobile phone data to make lending decisions. These models are trained on large datasets (millions of past loans with known repayment outcomes) and are continuously refined as new data becomes available.
Customer Analytics and Personalisation
Banks use ML to segment their customer base and personalise products and services. By analysing transaction patterns, channel preferences, and product usage, ML models can predict which customers are likely to be interested in a new product, which are at risk of churning, and what products would best serve each customer's needs.
These models typically run in batch mode (processing data overnight or periodically rather than in real time) and therefore have less stringent latency requirements than fraud detection models. They can run on standard CPU servers and do not necessarily require GPU infrastructure. However, they do require access to large datasets and significant data processing capacity.
Data Centre Requirements for ML Workloads
Training Infrastructure
Training an ML model is the most computationally intensive phase. It involves processing large datasets (often terabytes of transaction history) through multiple iterations of the model to optimise its parameters. For simple models (like logistic regression for credit scoring), training can be done on standard CPU servers in hours. For complex models (like deep neural networks for fraud detection), training may require GPU servers and take days or weeks.
The data centre infrastructure required for training depends on the model complexity. For most Kenyan banking ML applications, a cluster of 4-8 GPU servers (using NVIDIA T4 or A100 GPUs) is sufficient. For frontier models (like large language models), the requirements are much larger, but these are not typically used in banking applications.
Training workloads are typically batch jobs, they run for hours or days and then stop. This means the infrastructure can be shared among multiple teams and multiple models. Some organisations train models in the cloud (using AWS, Azure, or Google Cloud GPU instances) and deploy them in local data centres for inference, combining the cloud's flexibility with local hosting's compliance advantages.
Inference Infrastructure
Inference (using a trained model to make predictions) is the ongoing, production workload. Every time a customer makes a transaction, the fraud detection model must score it. Every time a loan applicant submits an application, the credit scoring model must assess it. These inference requests happen continuously, 24 hours a day.
For most banking ML models, inference can run on standard CPU servers. A well-optimised credit scoring model can score a loan application in under 100 milliseconds on a modern CPU server. Real-time fraud detection, which must handle thousands of transactions per second, may benefit from GPU acceleration but can also be achieved with CPU-optimised models.
The key inference infrastructure requirement is not raw compute power but reliability and low latency. Inference servers must be available 24/7, must respond within tight latency budgets, and must be deployed close to the systems that generate the requests. This is why inference models for Kenyan banking are typically hosted in Kenyan data centres rather than in the cloud.
Data Infrastructure
ML models are only as good as the data they are trained on. Kenyan banks and fintechs must store and process large volumes of transaction data, customer data, and alternative data sources. This data must be stored securely (complying with the Data Protection Act), organised for efficient access, and available for training pipelines.
Data infrastructure for ML includes data lakes (large-scale storage for raw data), data warehouses (organised storage for structured data), feature stores (storage for pre-computed ML features), and data pipelines (automated workflows that move data from source systems to ML training environments). All of this infrastructure runs in data centres.
The ML Talent Challenge
Kenya produces many software developers, but the specific skills required for ML engineering, statistical modelling, feature engineering, model optimisation, MLOps (the practice of deploying and managing ML models in production), are in short supply. The Microsoft Africa Development Centre in Nairobi employs ML engineers working on global products, and some of these skills are transferring to the local ecosystem. Strathmore University, the University of Nairobi, and other institutions have introduced data science and ML programmes, but the pipeline is still thin relative to demand.
For banks and fintechs, this skills gap means that ML adoption requires either hiring from a limited talent pool (driving up salaries), training existing staff (which takes time), or partnering with specialised ML companies (which adds cost and dependency). Some organisations address the gap by using cloud-based ML platforms (AWS SageMaker, Azure Machine Learning, Google Vertex AI) that reduce the operational complexity of ML deployment.
The Regulatory Environment
The Central Bank of Kenya (CBK) has been cautiously supportive of technology adoption in financial services, including AI and ML. The CBK's guidance on technology in banking emphasises risk management, data protection, and the need for human oversight of automated decisions.
Key regulatory considerations for ML in Kenyan banking include:
Explainability: When an ML model declines a loan application, the customer has a right to know why. The CBK may require that ML-based decisions can be explained in terms that customers and regulators understand. This is straightforward for simple models (like decision trees or logistic regression) but challenging for complex models (like deep neural networks).
Data protection: ML models trained on personal data must comply with the Data Protection Act. This includes obtaining appropriate consent, ensuring data is used only for stated purposes, and implementing security measures to protect the training data.
Fairness and bias: ML models can inadvertently discriminate against certain groups if the training data reflects historical biases. Ensuring that credit scoring models do not discriminate based on gender, ethnicity, or other protected characteristics is both a regulatory requirement and an ethical imperative.
The Future
Machine learning in Kenyan banking and fintech is still in its early stages. The most mature applications (fraud detection, basic credit scoring) are well-established. The next wave will include more sophisticated applications: natural language processing for customer service (chatbots that understand Swahili and Sheng), computer vision for document verification (automating KYC processes), and predictive analytics for financial planning.
Each of these applications requires data centre infrastructure, for training, inference, and data storage. As ML becomes more deeply embedded in Kenya's financial services, the data centre infrastructure that supports it will become more critical, more specialised, and more valuable. The banks and fintechs that invest in ML today are also, indirectly, investing in the data centre infrastructure that makes it possible.