Razorpay launches an AI model built specifically for payments, with Nvidia and AWS
Vulcan is trained on 4 billion transactions and 3 trillion data points, and Razorpay says it has already cut fraud rates and lifted payment success in production.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Razorpay launched Vulcan, a foundation model built specifically for payments, developed with Nvidia and AWS and trained on roughly 3 trillion data points across 4 billion transactions, according to Inc42.
- Razorpay says the model has produced an 8 to 10% improvement in payment success rates and an eightfold increase in international card fraud detection.
- Inc42 notes Stripe previously launched a similar payments-specific foundation model trained on tens of billions of transactions, making Razorpay's move a response to an existing category rather than a first.
Razorpay, the Bengaluru-based payments company, launched an AI foundation model called Vulcan on August 18, built specifically for payments processing in partnership with Nvidia and AWS. The company says the model is trained on roughly 3 trillion data points drawn from 4 billion transactions, with around 3,000 signals extracted per transaction, according to Inc42's reporting.
"Every payment teaches the system something that makes the next payment better," said Harshil Mathur, Razorpay's chief executive and co-founder, according to Inc42. Razorpay reports an 8 to 10% improvement in payment success rates and an eightfold increase in international card fraud detection since deploying the model, alongside a fivefold rise in identifying fraudulent or disputed transactions.
Inc42's reporting on the launch notes that Stripe previously introduced a comparable foundation model built specifically for payments, trained on tens of billions of transactions — meaning Razorpay's Vulcan responds to an approach Stripe had already taken, rather than introducing a new category of product.
A payments company the size of Razorpay building its own foundation model, rather than layering a general-purpose LLM onto its fraud and routing systems, is a bet that payments data has enough distinct structure that a purpose-built model beats a general one — and if Razorpay's own performance numbers hold up outside its own case study, that argument gets harder for competitors processing similar transaction volumes to ignore.
Will Razorpay publish independently audited fraud-detection results validating the performance improvements it has reported for Vulcan?
Still open. When the paper finds out, it will say so here and on the open questions page — including if it got this wrong.