Razorpay Announces India’s First AI Payments Foundation Model

Razorpay has announced Vulcan, a transformer-based AI foundation model designed specifically for payments, trained on nearly 3 trillion data points spanning 4 billion payments on its network.

Built with NVIDIA GPUs and AWS cloud infrastructure, the model is intended to improve payment routing, fraud detection, risk assessment and checkout personalisation.

Razorpay said it has already tested early components of the model on live transactions across its network.

The company claims an 8–10% improvement in payment success rates, an eightfold increase in international card fraud detection, and five times more fraudulent or disputed transactions identified without increasing the number of alerts.

The model learns from roughly 3,000 signals per transaction. Razorpay said it developed the model as a proprietary system, with both its architecture and training data built in-house.

The company’s rationale for a shared model is that payment decisions often depend on overlapping signals. Routing, fraud, risk and checkout systems have traditionally been developed as separate machine-learning models.

Vulcan is intended to provide a common intelligence layer that can use those signals across multiple payment decisions.

One application is real-time routing. The model scores available payment routes before a transaction is attempted and selects the route it predicts is most likely to succeed.

Razorpay said the model is also being used to detect fraud patterns across merchants, identify risky cash-on-delivery orders and predict which payment method is most likely to work for an individual shopper.

Razorpay’s Magic Checkout has also seen a 40% increase in shoppers being shown their preferred UPI application, which the company said is helping complete an additional 100,000–200,000 purchases each month.

The model is not an LLM. Razorpay said it uses transformer architecture but is trained to model patterns in payment transactions rather than language.

The company described it as a foundation model because the same underlying representation of payment behaviour can be adapted to multiple downstream use cases rather than requiring a separate model for each task.

NVIDIA GPUs are being used for training and running the model at scale, while AWS infrastructure, including Amazon SageMaker, supports its development, training and deployment. Razorpay said customers including Blinkit, Bachatt and redBus have begun using capabilities derived from the model in live payment environments.

Harshil Mathur, CEO and co-founder of Razorpay, said the objective was to address reliability across India's expanding digital payment base. “Every payment teaches the system something that makes the next payment better,” he said.

Pahal Patangia, head of global industry business development and payments at NVIDIA, said the work combines payments data with “real-time contextual intelligence”, while AWS said the model consolidates billions of transaction insights into a single intelligence layer.

Razorpay said it eventually intends to extend the model across more payment decisions, including authentication, routing, fraud detection and lending.

In March, Razorpay launched Agent Studio, a marketplace and builder for AI agents designed to handle payment and financial operations. Initial agents included tools for recovering abandoned carts and failed subscriptions, responding to payment disputes and forecasting cash flow. 

Its agentic onboarding system, for example, was designed to reduce merchant onboarding from 30–45 minutes to about five minutes, while its integration agent can help businesses connect Razorpay to their technology stack in under 10 minutes.