Razorpay launches Vulcan AI payments model with AWS, NVIDIA

Razorpay launched Razorpay Vulcan, a transformer-based AI payments foundation model created with Amazon Web Services and NVIDIA to raise payment success rates and strengthen fraud detection across India. The model was trained on the company’s transaction data and uses NVIDIA GPUs and Amazon SageMaker to operate at scale.

Razorpay reported the model was trained on roughly 3 trillion data points covering about 4 billion payments and evaluates roughly 3,000 signals per transaction. Early components of Vulcan have been run on live traffic across Razorpay’s network to test routing, fraud and risk decisions before full deployment.

The company reported preliminary results from those live tests: an 8–10% increase in payment success rates; eight times more detection of international card fraud; five times more identification of fraudulent or disputed transactions without raising more alerts; and a 40% rise in shoppers shown their preferred UPI app in Razorpay Magic Checkout, which the firm said has helped complete about 100,000–200,000 additional purchases per month.

Razorpay developed the model to address routing failures that can cause a payment to fail even when the payer’s account and instrument are fine. The firm said industry practice has been to run separate models for routing, fraud, risk and checkout that do not share signals. Vulcan applies a single, shared intelligence layer that scores routes and other decisions in real time before a payment attempt.

The model uses transformer architecture, a class of neural networks commonly used in large language models, but Razorpay described Vulcan as specialized for payments data rather than for text. The company presented the model as a foundation model because it is designed to generalize across multiple payment tasks without being retrained for each specific use case.

Razorpay has started rolling elements of Vulcan into live environments with customers including Blinkit, Bachatt and redBus. The model’s routing capability directs payments down the path deemed most likely to succeed, and its network-level fraud detection looks for suspicious activity visible across multiple merchants. Other features include flagging risky cash-on-delivery orders before checkout and recommending the payment method most likely to work for each customer.

Harshil Mathur, Razorpay’s CEO and founder, described the product as built for customers who are still deciding whether to trust digital payments over cash and said each processed payment helps the system improve. Pahal Patangia, head of global industry business development and payments at NVIDIA, highlighted NVIDIA’s role in providing accelerated computing for training and live inference. Kiran Jagannath, head of financial services and conglomerates for AWS India and South Asia, pointed to Amazon SageMaker and AWS infrastructure as supporting development, training and deployment.

Razorpay said it plans to extend the model’s scope so more payment decisions, from authentication to lending, are handled by a single, continuously learning system. The company cited projections that India’s e-commerce market could reach about $350 billion by 2030 and positioned Vulcan as part of the technical infrastructure for handling increased transaction volumes.