Razorpay trained a payments model. The interesting part is who it already sits inside.

Vulcan, built with NVIDIA and AWS on what the company says is four billion transactions and three trillion data points, is a transformer sitting on top of India's most unforgiving consumer behaviour: pay, fail, abandon. Razorpay claims an 8–10% lift in success rates in tests with 51,000 businesses, and a sharper fraud catch than the rules engines it replaces. Blinkit and redBus are already live.

That is not a research paper. Success rate is the quiet P&L of Indian digital commerce. A two-point miss on UPI at a q-commerce checkout is not a "friction metric." It is a warm basket going cold while a rider waits downstairs. For a travel aggregator it is a seat that does not get locked. The merchants who feel this first are not banks. They are operators whose unit economics already assume the payment just works.

A two-point miss on UPI at a q-commerce checkout is not a friction metric. It is a warm basket going cold.

What Vulcan actually does is less cinematic than the launch copy. Payments in India are a mesh of UPI handles, cards, net banking, wallets, EMI, and a long tail of issuer quirks. Routing is still, at too many gateways, a decision tree dressed up as intelligence. A foundation model that watches every decline — issuer down, PIN retry, VPA typo, velocity spike — and updates the next hop is closer to infrastructure than to a chatbot with a treasury.

The competitive question is whether this stays a Razorpay product or becomes table stakes. PhonePe, PayU, Cashfree, and the bank stacks all sit on similar firehoses. NPCI sees the whole UPI graph. The winner will not be the lab with the prettiest demo. It will be whoever can improve authorisation without inflating fraud, and do it at the 200-millisecond patience of an Indian checkout.

Razorpay's advantage is distribution: it already sits in the merchant SDK. A model that never leaves the request path is more valuable than a dashboard that tells you, after the fact, that you lost 4% of carts to "technical decline." If Vulcan holds the claimed lift in production through festive, it will not just sell more Razorpay. It will raise the floor for everyone else.