Summary
Razorpay has built an AI payments foundation model trained on four billion transactions, seeking to improve payment success rates, fraud detection and checkout personalisation as India's digital payments ecosystem scales. The model, built with NVIDIA and AWS, uses around 3,000 signals per transaction and analyzes payment flows across merchants, banks, instruments, and gateways. Early tests across 1.5 million transactions and more than 51,000 businesses showed an 8-10% improvement in payment success rates, an eight-fold increase in international card fraud detection, and a five-fold increase in identification of fraudulent or disputed transactions.
Key Facts
- AI payments foundation model trained on 4 billion transactions
- Built with NVIDIA and AWS; ~3,000 signals per transaction
- Tests: 1.5M transactions, 51,000+ businesses; some merchants route 50-60% of volume through it
- Results: 8-10% payment success rate improvement; 8x international card fraud detection; 5x fraudulent/disputed transaction identification
- At Blinkit: 1-2 percentage point improvement in payment success rates
- "First payment foundation model in India"; built within India, PII removed before training
- CEO Harshil Mathur: "The payment foundation model is one single foundational model that works for all of these things"
- Not currently charging merchants separately; monetization via volume and additional services (lending, marketing)
- Guardrails: maintaining existing false-positive rates for fraud detection
- Razorpay IPO continues via confidential route
Why It Matters
Razorpay's foundation model represents a step change in how payment companies apply AI: instead of bespoke models for fraud, success rate, or checkout, a single multidimensional foundation model learns from the entire payments ecosystem and extends to new use cases (authentication, routing, lending). That it is "the first in India" matters because India's payment architecture — UPI, cards, net banking, diverse banking ecosystem — is unique, so models built elsewhere can't simply be deployed. This could become a competitive moat as Indian payment volumes scale, with meaningful gains in success rates and fraud detection translating directly to revenue.