Razorpay launched Vulcan on Tuesday, describing it as India’s first transformer-based AI foundation model built specifically for payments.
The model draws on approximately 3 trillion data points from 4 billion payments processed on Razorpay’s network and evaluates roughly 3,000 signals for each transaction. It was built with NVIDIA accelerated computing, primarily H100 GPUs, and AWS cloud infrastructure that includes Amazon SageMaker. Razorpay said the model was trained and is hosted entirely in India.
What early testing showed for success rates and fraud
Early components have already operated on live traffic. The company reported an 8–10 percent rise in payment success rates, detection of eight times more international card fraud, and identification of five times more fraudulent or disputed transactions without an increase in alerts. On Razorpay Magic Checkout, 40 percent more shoppers were shown their preferred UPI application, supporting an additional 100,000 to 200,000 completed purchases each month, according to company figures.
Vulcan is running in beta across more than 51,000 businesses, including Blinkit, Bachatt and redBus. In an interview with Hindustan Times, co-founder and CEO Harshil Mathur said testing covering more than 1.5 million transactions and over 50,000 merchants produced “an 8 to 10% improvement in payment success rates, and a 5x reduction in fraud.”
Why Razorpay built the model from scratch
The architecture uses transformers adapted to payment patterns and transaction graphs rather than natural language. Mathur said open-source large language models were unsuitable because they “cannot model payment behaviour or transaction graphs” and that “there was no pre-existing foundational model for Indian payments that we could fine-tune.” The single model is designed to support routing decisions, fraud detection, risk assessment and checkout personalisation, improving with every payment it processes.
An internal study of 1.5 million shoppers and more than 51,000 businesses found payment friction affected consumers in both metropolitan and smaller markets. Razorpay said the system addresses failed transactions, drop-offs, delays and fraud across UPI, cards, net banking, wallets and cash-on-delivery routes that involve hundreds of banks and instruments.
NVIDIA, AWS and data kept inside India
Pahal Patangia, NVIDIA’s Head of Global Industry Business Development and Payments, said the collaboration converts complex payments data into “real-time contextual intelligence.” Kiran Jagannath, Head of FSI and Conglomerates at AWS India and South Asia, described the model as consolidating “billions of transaction insights into a single, continuously learning intelligence layer.”
Razorpay stated that personally identifiable information is removed and that processing remains on Indian infrastructure to meet RBI data-localisation rules and the Digital Personal Data Protection Act. Mathur said the company expects to release additional product capabilities and a technical account of the architecture and training process in the coming months.
India’s payment networks already handle tens of billions of UPI transactions each month. Razorpay presents Vulcan as a reliability layer for that scale. Independent market forecasts for Indian e-commerce by 2030 generally range between $200 billion and $300 billion.
Sources
