NPCI and Nvidia announced a partnership on February 18 to build a sovereign AI foundation model designed for India's payment infrastructure. The collaboration uses Nvidia's Nemotron architecture to create a payments-native model trained on Indian financial data, a departure from general-purpose AI models that dominate the financial sector today.
The partnership extends NPCI's existing AI work. NPCI deployed its Financial Model for India, known as FiMI, as part of the UPI Help Assistant, which handles grievance redressal for the platform's 500 million active users. The new foundation model will expand beyond customer support into fraud detection, transaction monitoring, and dispute resolution at scale.
India's UPI processed 20.39 billion transactions in February 2026, with daily averages reaching 728 million. At these volumes, traditional rule-based fraud detection systems face capacity constraints. NPCI intends the model to operate in a Mixture of Experts configuration, where specialized sub-models handle different transaction categories and risk profiles simultaneously while maintaining the low-latency requirements of real-time payments.
The sovereign designation reflects a deliberate decision to keep the model's training data, inference capabilities, and operational control within India. Unlike third-party AI services on global cloud platforms, the NPCI-Nvidia model will be deployed on Indian infrastructure and trained exclusively on Indian payment data, aligning with RBI's data localization requirements for payment system operators.
Fraud on UPI has been a persistent concern as volumes scaled. NPCI completed the phase-out of UPI Collect, the pull payment mechanism, in February 2026 after tracing over 70 percent of fraud cases to collect requests. An AI-driven fraud detection layer operating at the switch level would add real-time pattern analysis to complement the structural removal of high-risk payment flows.
The foundation model approach allows NPCI to serve its entire participant base rather than leaving AI adoption to individual banks and payment service providers. India has 694 banks live on UPI, along with hundreds of prepaid payment instrument issuers. A centralized AI layer at the switch level could standardize fraud detection quality across participants of varying technical capability, from large private banks with dedicated data science teams to small cooperative banks operating with minimal technology staff.
NPCI has not disclosed a deployment timeline for the production model, though the UPI Help Assistant integration with FiMI provides a working baseline. The Mixture of Experts architecture is suited to high-throughput environments like UPI, which handles over 700 million transactions daily. The next regulatory milestone is the RBI's risk-based authentication framework taking effect on April 1, 2026, which creates a compliance incentive for banks to adopt advanced transaction-scoring capabilities.