Bengaluru-based Razorpay launched Vulcan on 18 August 2026, described as India\u2019s first transformer-based AI foundation model built specifically for digital payments. Trained on nearly 3 trillion data points across 4 billion transactions, Vulcan processes around 3,000 signals per transaction to improve payment success rates, detect fraud, and personalise checkout experiences in real time. Built in partnership with NVIDIA and Amazon Web Services (AWS), the model is already live across Razorpay\u2019s network and has shown measurable improvements for merchants including Blinkit, Bachat, and redBus.
What Is Razorpay Vulcan?
Vulcan is a foundation model designed to understand patterns in the movement of money. Unlike a large language model (LLM) that processes text, Vulcan is trained on payments data and uses transformer architecture to analyse transaction behaviour at massive scale.
Razorpay CEO and co-founder Harshil Mathur explained that traditional payment systems rely on separate machine-learning models for each function, such as one model for routing payments, another for detecting fraud, and yet another for risk assessment. Vulcan replaces this fragmented approach with a single intelligence layer that handles routing, fraud detection, risk assessment, and checkout personalisation simultaneously.
The model is proprietary to Razorpay, with both its architecture and training data owned by the company. It was developed using NVIDIA GPUs for training and deployment, while AWS infrastructure, including Amazon SageMaker, supported its development, training, and deployment lifecycle. Razorpay said no payment data was shared with third parties during the process.
How Vulcan Works: The Technology Behind the Model
Training Scale and Data Signals
Vulcan was trained on approximately 3 trillion data points drawn from 4 billion payments processed through Razorpay\u2019s network. The model analyses around 3,000 signals per transaction, drawing from data across merchants, payment instruments, issuers, and gateways.
These signals include information about which bank is processing a transaction, what time the transaction occurs, which payment method is being used, and historical patterns of success or failure for similar transactions. By learning from this vast dataset, Vulcan identifies patterns such as a particular bank performing poorly for certain cards at specific times and uses those signals to route payments through paths most likely to succeed.
Real-Time Decision Making
One of Vulcan\u2019s most significant technical capabilities is its speed. The model processes more than 3,000 signals per transaction and makes automated decisions within 29 milliseconds, fast enough to influence the outcome of a payment before it is submitted to the banking network.
This real-time processing means that when a customer initiates a payment, Vulcan instantly evaluates the best route, assesses fraud risk, and personalises the checkout experience, all within the window of a single transaction attempt.
Key Capabilities of Vulcan
Payment Routing
Vulcan\u2019s routing capability evaluates payment routes in real time and selects the one most likely to succeed before a transaction is attempted. The model identifies patterns such as a particular bank underperforming for certain card types or experiencing higher failure rates at specific times of day. It can also recommend which saved payment instrument is most likely to work for a particular customer, reducing failed transactions and improving overall success rates.
Fraud Detection
The model operates as a network-level fraud detection system, capable of spotting suspicious patterns across Razorpay\u2019s entire merchant base rather than in isolation for a single business. It identifies and stops 8 times more international card fraud than previous systems and detects 5 times more fraudulent or disputed transactions without increasing the number of alerts generated.
Vulcan also flags potentially risky Cash on Delivery (COD) orders, helping merchants avoid losses from undelivered or fraudulent orders. The ability to identify fraud across the network, rather than at the individual merchant level, gives the model a broader view of emerging fraud patterns.
Checkout Personalisation
Through Razorpay\u2019s Magic Checkout product, Vulcan personalises the checkout experience for each shopper. The model has enabled 40% more shoppers to see their preferred UPI application at checkout, helping to complete an additional 1 to 2 lakh purchases every month. This personalisation reduces friction at the final step of a transaction, where many customers abandon their purchases.
Performance Metrics and Results
Razorpay reported the following results from early deployments of Vulcan across its payments network:
| Metric | Result |
|---|---|
| Payment success rate improvement | 8 to 10% |
| International card fraud detected | 8 times more |
| Fraudulent or disputed transactions identified | 5 times more |
| Shoppers seeing preferred UPI app at checkout | 40% more |
| Additional monthly purchases completed | 1 to 2 lakh |
These results were measured across approximately 1.5 million shoppers and more than 51,000 businesses that participated in the pilot phase. Customers including quick-commerce platform Blinkit, savings app Bachat, and travel platform redBus have been among the early users of Vulcan\u2019s capabilities in live payment environments.
Razorpay said the model was developed after an internal study found that payment friction, including failed transactions, drop-offs, and delays, affected consumers identically across metropolitan and smaller markets. The company positioned Vulcan as a solution aimed at making digital payments more dependable for users who are still deciding whether to trust digital transactions over cash.
Industry Context: India\u2019s Digital Payments Landscape
India\u2019s digital payments ecosystem has grown at an unprecedented pace. Unified Payments Interface (UPI), operated by the National Payments Corporation of India (NPCI), processed 23.66 billion transactions worth ₹29.88 lakh crore in July 2026 alone, according to NPCI data. On average, UPI processed 763 million transactions a day during the month, with average daily transaction value at ₹96,383 crore.
This explosive growth has brought new challenges. Banks reported 13,516 cases of card and internet frauds amounting to ₹520 crore in FY25. The Reserve Bank of India (RBI) is developing its own Digital Payments Intelligence Platform (DPIP) to address these concerns, using artificial intelligence to detect and prevent payment frauds in real time. The RBI\u2019s prototype is being developed by the RBI Innovation Hub and is being implemented across several banks.
NPCI has also initiated a pilot for an AI model capable of tracking fraudulently obtained money in real time as it moves across bank accounts, a technology designed to intercept digital fraud before funds disappear into networks of mule accounts.
In this environment, Razorpay\u2019s Vulcan represents a private-sector response to the same challenge the regulators are tackling: making digital payments safer and more reliable as volumes continue to surge. The company cited India\u2019s digital e-commerce market, which it projects will reach $350 billion by 2030, as the broader market opportunity for the technology.
About Razorpay
Razorpay was founded in 2014 by Harshil Mathur and Shashank Kumar, both alumni of the Indian Institute of Technology (IIT) Roorkee. The company started as a payment gateway for startups and small businesses and has grown into India\u2019s only full-stack financial solutions company, offering payments, banking, and lending solutions. Razorpay powers online payments for 76 of India\u2019s top 100 startup unicorns and millions of businesses, with a valuation of $7.5 billion as of its last funding round. The company is headquartered in Bengaluru and has more than 3,300 employees.
The Way Forward: Future Applications
Razorpay said it plans to extend Vulcan beyond its current four functions. The company is exploring applications in payment authentication, lending decisions, and marketing, depending on how the model performs in additional use cases.
Mathur said the foundation model approach allows the company to keep applying new use cases to the same underlying intelligence layer, rather than building separate models from scratch for each new application. Razorpay said it plans to publish details of its technical approach in the coming weeks, which could offer deeper insight into how the model processes and learns from payment data.
Pahal Patangia, Head of Global Industry Business Development and Payments at NVIDIA, said the collaboration represents a new frontier in turning complex payments data into real-time contextual intelligence. Kiran Jagannath, Head of FSI and Conglomerates at AWS India and South Asia, described the model as consolidating billions of transaction insights into a single, continuously learning intelligence layer.
As India\u2019s digital economy expands, models like Vulcan could play a significant role in shaping how the country\u2019s payments infrastructure evolves, bridging the gap between growing transaction volumes and the need for reliable, fraud-resistant systems.
Key Takeaways
- Razorpay Vulcan was launched on 18 August 2026 as India\u2019s first transformer-based AI foundation model built specifically for digital payments.
- The model was trained on 3 trillion data points across 4 billion transactions and analyses 3,000 signals per transaction within 29 milliseconds.
- Vulcan achieved an 8 to 10% improvement in payment success rates, detected 8 times more international card fraud, and identified 5 times more fraudulent or disputed transactions.
- The model was built in partnership with NVIDIA (compute and architecture) and AWS (Amazon SageMaker for development, training, and deployment).
- Razorpay, founded in 2014 by Harshil Mathur and Shashank Kumar, is valued at $7.5 billion and powers payments for 76 of India\u2019s top 100 startup unicorns.
- The RBI is separately developing its Digital Payments Intelligence Platform (DPIP) using AI to detect and prevent payment frauds in real time, with the prototype being built by the RBI Innovation Hub.
