Razorpay Launches Vulcan AI Model Trained on 4 Billion Payments to Cut Fraud and Boost Success Rates

Razorpay launches AI model trained on four billion payments

Razorpay has launched an AI foundation model to enhance online payment success. This advanced system improves fraud detection and payment routing capabilities significantly. The model learns from billions of transactions, processing thousands of signals per payment. Beta tests show improved success rates and better fraud identification for merchants. Payments firm Razorpay has launched an artificial intelligence (AI) model trained on four billion transactions to improve online payment success and detect fraud, bringing routing, risk checks, and checkout personalisation into one system.Developed along with Nvidia and Amazon Web Services, the model has learnt from three trillion data points and studies roughly 3,000 signals for every transaction, the company's cofounder and chief executive

Razorpay Unveils Vulcan: India's First AI Model to Power a $350B E-Commerce Future

India's digital payments story has been one of growth and innovation - yet a quieter gap remains: a payment that doesn't go through, an OTP that arrives late, a subscription that silently lapses, a card compromised by fraud. For millions of first-time or small-town shoppers, one such experience is often enough to send them back to cash. Today, Razorpay, India's Omnichannel Payments Platform for Businesses, announced the launch of the Razorpay Vulcan, India's First Transformer-based AI Foundation Model built for payments, designed to make every digital payment in India more reliable, safer, and predictable. Built with NVIDIA and AWS technology, it combines Razorpay's payments data powered by NVIDIA's accelerated computing and AWS's cloud infrastructure - groundwork that India's e-commerce market needs as it heads toward a projected $350 bn by 2030. An internal Razorpay study across 1.5 million shoppers and 51,000+ businesses found the same payment friction - failed transactions, drop-offs, delays - surfacing identically from a metro high street to a small-town market. That's what convinced Razorpay to build a single shared model rather than keep refining each one separately. The Impact so far: Ahead of today's full launch, early components of the Razorpay Foundation Model have been running across 3 trillion data points on the company's network - testing routing, fraud, and risk decisions on live transactions. Customers including Blinkit, Bachatt, and redBus, among others, have already started seeing the benefits of these capabilities in live payment environments. * 8-10% improvement in payment success rates * ⁠⁠8x more international card fraud detected and stopped * ⁠⁠5x more fraudulent or disputed transactions identified, without increasing the number of alerts * ⁠⁠40% more shoppers see their preferred UPI app on Razorpay Magic Checkout, helping complete 1-2 lakh more purchases every month Why Razorpay: Most companies see only one slice of a payment. Razorpay sees payments moving across merchants, instruments, issuers, and gateways at once - the breadth needed to understand India's payments ecosystem as a whole. Why this matters for India: India's payments landscape is unlike any other: a single purchase can be processed via UPI, cards, net banking, wallets, or Cash on Delivery across hundreds of banks and gateways. Take Meera, buying running shoes for ₹2,400 at 9 pm: she taps "Pay" and sees "Payment unsuccessful. Please try again." Her card, bank, and money are all fine - her payment simply had several possible routes, and one was briefly the wrong choice at that moment. This pattern, repeated across millions of Indians, led Razorpay to build a model that scores every route in real time and picks the healthiest one before a payment is attempted. What's been missing till now: The industry has tackled this with separate, specialised models - one each for routing, fraud, risk, and checkout - that don't talk to each other, even though the same signals matter to all of them in one way or the other. It's like several doctors examining a patient, each reading only their own test results. The Solution: One shared intelligence layer for every payment The Razorpay AI Payments Foundation Model learns from the entire payments ecosystem's data points at once, and keeps improving with every transaction it processes - instead of solving one narrow problem at a time. Built on transformer technology - the same family of AI architecture behind many of today's AI systems like LLMs - it is adapted specifically for the patterns hidden inside Indian payments data. Not another ML model, and not an LLM either: A traditional ML model is built for one job; solve a new problem, and you start over. A foundation model learns how payments move, so that understanding extends to new use cases without re-training. And while the term comes from LLMs, this isn't one - LLMs understand text; this model understands the language of the movement of money. By the numbers: * Trained on approximately 3 trillion data points across 4 bn payments * Learns from roughly 3,000 signals per transaction * Built entirely as a proprietary, ground-up model - both the architecture and the training data belong to Razorpay * Generic LLMs understand text. This model understands the complex movement of money at a massive scale Powering training and live decisions with NVIDIA and AWS: Training a model on 3 trillion data points across 4 bn payments needs serious computational muscle. NVIDIA's GPUs powered the training and running of the model at scale; AWS's cloud infrastructure, including Amazon SageMaker, supported development, training, and deployment. * For businesses, this means fewer lost sales, reduced OTP drop-offs, lower fraud losses, and fewer RTO returns * For consumers, it means payments that simply work, every time Harshil Mathur, CEO & Founder of Razorpay, said, "India's appetite for digital payments is real, but it isn't universal yet - for a large part of the country, going digital still comes down to one thing: does it work, every single time? That's the customer we built this for: the one still deciding whether to trust a screen over cash in hand. An AI-led payments foundation model doesn't just solve today's problem and stop there. Every payment teaches the system something that makes the next payment better. That's what makes this feel less like a product launch, and more like the starting point for how payments in India keep getting better on their own, for years to come." Pahal Patangia, Head of Global Industry Business Development and Payments, NVIDIA said, "India's rapidly evolving digital economy is creating an opportunity to make payments more intelligent, reliable, and secure. NVIDIA's work with Razorpay in partnership with AWS on AI payments foundation models has opened up a new frontier, turning complex payments data into real-time contextual intelligence. This has a proprietary and purpose-built semantic AI layer that can help advance the next generation of digital financial services." Kiran Jagannath, Head of FSI and Conglomerates, AWS India and South Asia, said, "Razorpay is reimagining payments intelligence at India scale with an AI Foundation Model - built on Amazon SageMaker - that consolidates billions of transaction insights into a single, continuously learning intelligence layer, replacing fragmented ML models with unified AI that delivers higher payment success rates, rapid iteration, and enterprise-grade security for mission-critical payment flows. As India's digital economy grows, we are excited to power the AI infrastructure behind payments that simply work for every Indian." What Razorpay's AI Payments Foundational Model can do ● Hyper-Precision Routing: Sends each payment down the path most likely to succeed, in real time ● Network-Level Fraud Detection: Spots fraud visible only across merchants, flagging a stolen card the moment it's used across unrelated sellers ● RTO Risk Intelligence: Flags risky Cash on Delivery orders before checkout ● Predictive Checkout Personalisation: Recommends the payment method most likely to work for each customer. Looking ahead: Razorpay sees this as the starting point, not the destination - with the goal that every payment decision, from authentication to routing to fraud to lending, is eventually powered by one continuously learning model. With India's digital e-commerce market projected to reach $350 bn by 2030, Razorpay sees the AI Payments Foundation Model as part of the groundwork that growth will need.

What does Razorpay's AI model for payments, Vulcan train on?

Razorpay has launched Vulcan, which it calls "India's first transformer-based AI foundation model for payments." The model works on payment routing, fraud, risk and personalisation, and Razorpay built it with Nvidia and Amazon Web Services (AWS). According to Entrackr, customers including Blinkit, Bachatt and redBus have started using some of its capabilities in live payments. What is a foundation model? One large model trained on broad data that many tasks draw on, instead of a separate narrow model for each task. Founder Harshil Mathur likened Vulcan to a large language model (LLM). "Like LLMs are trained on text to understand language, Vulcan is trained on payments to understand how money moves," he wrote. What do we know about it? * Training scale: "4 billion payments. 3 trillion data points. One model trained on all of it," Mathur wrote. The model uses around 3,000 signals per transaction, and Razorpay built, trained and hosted it in India. * What it draws on: Signals across merchants, payment instruments, issuers and gateways, which the model uses to pick the route most likely to spot fraud patterns across merchants. * Ownership: Razorpay developed Vulcan from the ground up, with its architecture and training data proprietary to the company. * Origin: Before building Vulcan, Razorpay ran an internal study of 1.5 million shoppers and more than 51,000 businesses that found the same payment friction across metros and smaller markets. * Business model: Razorpay is not charging merchants for Vulcan. What does Razorpay claim it delivers? * "8-10% improvement in payment success rates." * "8x more international card fraud detected." * "5x more fraudulent or disputed transactions identified," which Razorpay says it achieved without increasing the number of alerts. * Through Magic Checkout, 40% more shoppers shown their preferred UPI app, completing an additional 1-2 lakh purchases a month. These are Razorpay's own beta results. It has not published the methodology, baseline or sample period behind them. Who benefits? Razorpay pitches Vulcan as built for consumers. Mathur said it aims to make payments dependable for people "still deciding whether to trust digital transactions over cash," and the product page calls it "powered for a billion Indians." The benefits Razorpay lists go to merchants: fewer failed payments, lower fraud losses, fewer undelivered orders and higher checkout conversion. For consumers, Razorpay says only that they "could benefit from more reliable payment experiences." What do we not know? Razorpay has not disclosed three things that matter for a model trained on live payments: * Whether merchants can opt out. Razorpay has not said whether a merchant can keep its transactions out of Vulcan's training, or what happens to what the model has already learned if a merchant leaves. * Whose data it trains on. The company describes Vulcan's signals as spanning merchants, payment instruments, issuers and gateways. It has not said whether the 3 trillion data points also include data that identifies individual consumers. That distinction decides whether the DPDP Act applies to the training at all. * The legal basis for training. Razorpay owns Vulcan and its training data, but for much of the transaction data it handles, it is a processor acting for merchants, not the owner. It has not said what allows it to use that data to train a model of its own. Why does the legal basis matter? Across several products, Razorpay acts as a data processor for merchants, with the merchant classified as the data fiduciary for end-customer data. Its own Agent Studio privacy policy sets this out under Section 8(2) of the Digital Personal Data Protection Act, 2023 (DPDP Act). A data processor may use data only on the fiduciary's instructions and for the fiduciary's purposes. Training a model Razorpay owns is a separate purpose from processing a merchant's individual transactions. Under the DPDP Act, consent must be free, specific and informed. Whether the consent Razorpay's merchants and their customers gave covers training Vulcan on that data is not something the company has made public. What does Razorpay plan next? Razorpay says it will expand Vulcan to authentication and lending, using one AI layer across more of the payment journey. Lending is where it matters most. A model trained on transaction patterns to score payments, then extended to score creditworthiness, raises separate questions on fairness and explainability, including whether a declined borrower can be told why. The expansion comes as Razorpay heads toward a public listing. It secured shareholder approval in May for a Rs 2,700 crore fresh issue. Vulcan improves with every transaction and Razorpay owns it outright, which makes the data feeding it a commercial asset ahead of that listing. How does this fit Razorpay's AI push? Razorpay launched Agent Studio in March, which drew questions on AI-driven dark patterns and price discrimination in payments. It also built a voice AI agent with Sarvam AI. On liability there, Razorpay said "the introduction of agentic shopping does not rewrite the rules of commercial liability," placing commercial disputes with the merchant and payment security with itself. A misrouted payment or a wrongly flagged transaction falls closer to payment security, which the company owns, it has not said who absorbs that loss. MediaNama has sent the folllowing questions to Razorpay.

How Razorpay is Using AI to Reduce Payment Failures and Fraud

Fraud detection: Razorpay reports 8× more international-card fraud detected and stopped. Razorpay now uses AI across several parts of the payment journey, from checkout and payment routing to fraud checks and authentication. Its latest move is Vulcan, a transformer-based AI foundation model built for payments. Razorpay trained Vulcan on about 3 trillion data points from 4 billion payments. The model can study around 3,000 signals for each transaction. These signals can include payment behaviour, customer patterns, merchant activity, device details, payment routes and risk indicators. The scale gives Razorpay a large base of payment data for its AI systems. Instead of treating every payment as a separate event, the model can study wider patterns across its payment network.

Razorpay Launches Vulcan: AI Model Trained on 4 Billion Payments

Razorpay launched Vulcan, an AI model trained on more than 4 billion payment transactions. It is built to spot fraud by checking thousands of signals during each transaction. Razorpay launched Vulcan, a new AI model made for digital payments. According to reports, the AI model is trained on nearly 3 trillion data points spanning 4 billion payments on its network. Notably, the model is built on NVIDIA GPUs and AWS cloud infrastructure to improve payment routing, fraud detection, risk assessment, and checkout personalization. Vulcan checks thousands of signals . This gives it a wider view of what is happening with a payment. During the launch, the company claimed that it tested early components of the model on live transactions across its network and, impressively, reported an 8-10% improvement in payment success. Sometimes, a single payment may not look suspicious on its own, but things can change when several signals are checked together. Vulcan is built to find these patterns by analyzing and user behavior to spot signs of fraud. The model can then help payment systems flag risky transactions faster. The model learns from huge amounts of payment data. Thus, it recognizes patterns linked to suspicious activity. The aim is not only to catch fraud but also to avoid wrongly stopping genuine payments. Vulcan shows how quickly AI is becoming part of India's digital payment system. Razorpay can use the model to study payment behavior at a scale that would be difficult to handle manually. For businesses, better fraud detection can mean fewer losses. For customers, it could mean safer online payments. Vulcan is still in its early stages, but its focus is clear: use AI to make digital transactions more secure.