Razorpay bets on AI foundation model Vulcan to tackle India’s payments challenge

Razorpay has launched Vulcan, which, it says, is India’s first AI foundation model built specifically for payments.

The launch comes as the fintech company looks to tackle one of the biggest challenges facing the country’s digital economy: making online payments more reliable, secure and successful.

The model, developed with NVIDIA and AWS, has been trained on nearly 3 trillion data points across 4 billion payments. It aims to improve payment routing, fraud detection, risk management, and checkout experiences through a single AI system rather than a collection of separate models.

The announcement comes at a time when India’s payments ecosystem is expanding at extraordinary speed.

According to data released by the Indian government, UPI accounted for 81% of all retail digital payments in FY2024-25, making it the world’s largest real-time retail payment system. Digital payment volumes reached more than 22,000 crore transactions during the year. Government data also shows that UPI processed more than 24,000 crore transactions in FY2025-26, reflecting continued growth in both adoption and transaction value.

All this growth has created a complex payments environment. A single online purchase can move through UPI, cards, net banking, wallets or cash on delivery, while interacting with hundreds of banks and payment gateways.

The Bengaluru-based company said Vulcan learns from around 3,000 signals associated with every transaction and then predicts the route most likely to succeed.

Unlike a conventional machine learning model that is trained for a specific task, a foundation model is designed to learn broader patterns that can be applied across multiple use cases. Razorpay also emphasised that the system is not a large language model. Instead of understanding text, it is trained to identify patterns in the movement of money.

Early results shared by the company suggest an 8 to 10% improvement in payment success rates, eight times more international card fraud being detected, and five times more fraudulent or disputed transactions being identified without increasing alert volumes.

Razorpay co-founder and chief executive officer Harshil Mathur said the technology was designed for consumers who still remain uncertain about digital transactions. “India’s appetite for digital payments is real, but it isn’t universal yet. For a large part of the country, going digital still comes down to one thing: does it work, every single time?” he said.

NVIDIA supplied the accelerated computing needed to train the model, while AWS provided the cloud infrastructure used for development and deployment.

The broader industry is moving in the direction of AI. Payment companies across the world increasingly using the tech to improve authorisation, personalise checkout, and strengthen fraud detection.

According to 2025 research by Mastercard, on an average, organisations lost $60 million globally to payment fraud in a year, and AI was helping institutions reduce false positives and improve customer retention. At the same time, the company noted that fraudsters are also using generative AI to develop more sophisticated scams.

Visa has also accelerated its AI initiatives over the past two years, introducing AI-driven fraud tools and frameworks for what it calls agentic commerce, in which AI systems can eventually search for products and complete purchases on behalf of users. The company has described AI as one of the defining forces shaping the future of payments alongside real-time transactions and digital identity.

Building AI systems for payments increasingly depends on access to large datasets, specialised computing infrastructure, and cloud platforms.