Razorpay has launched Vulcan, a transformer-driven AI foundation model designed specifically for digital payments. The model is being positioned as a payments-focused intelligence layer that can work across transaction routing, fraud detection, risk assessment and checkout personalisation.
Razorpay says Vulcan was developed using NVIDIA GPUs and AWS infrastructure, including Amazon SageMaker, and trained on about 3 trillion data points from 4 billion payments, with roughly 3,000 signals considered per transaction.
Payment companies have used machine learning for years, but these systems are typically built for individual tasks.
About Razorpay Vulcan
Vulcan is a transformer-driven foundation model for payments. It is not an LLM designed to generate text, answer questions or operate as a consumer chatbot.
Instead, the model is trained to recognise patterns in transaction and payment behaviour. Razorpay says its design and training data were developed in-house, while NVIDIA computing infrastructure was used for training and running the model, and AWS services supported development, training and deployment.
The model is created to provide a common intelligence layer across multiple payment decisions, including:
- real-time payment routing
- fraud detection
- risk assessment
- checkout recommendations
- payment authentication
- selected lending use cases
Razorpay is attempting to move away from separate machine-learning models for separate payment problems toward a model that can learn from a much broader set of transaction relationships.
[ALSO READ: Tech Mahindra and Fortinet Forge Partnership to Deliver Managed SASE Services ]
A transaction can be affected by the payment method, issuing bank, merchant, device, transaction amount, customer history, location, time, network status and other signals. Fraud decisions also depend on relationships between transactions rather than simply whether an individual transaction looks unusual.
According to Razorpay, an internal study of 1.5 million shoppers and more than 51,000 businesses found recurring problems around failed payments, delays, and checkout drop-offs across markets.
A conventional system might have one model determining fraud risk and another choosing a payment route. Vulcan is designed to learn patterns throughout the wider payment environment, allowing the same underlying intelligence to be applied to several decisions.
Razorpay Vulcan system has been trained on approximately 4 billion payments, generating roughly 3 trillion data points. The company says about 3,000 signals are analysed per transaction.
Training and operating a model that processes transaction behaviour at this scale entails considerable computing infrastructure. NVIDIA GPUs provide the training and inference hardware, while AWS services, including Amazon SageMaker, provide the cloud environment used for model development and deployment.
Early results
Razorpay mentions early deployments involving Blinkit, Bachat and redBus produced an 8–10% improvement in payment success rates. It also reports an eightfold increase in international card fraud detection and says fraudulent or disputed transactions were identified five times more successfully without an increase in alerts.
On checkout personalisation, Razorpay says 40% more shoppers were shown their preferred UPI application through Magic Checkout, contributing to an additional 1–2 lakh purchases per month.
These are Razorpay’s reported results from early deployments, not independent industry benchmarks.
The $350 billion e-commerce context
A July 2026 report from the Centre for Social and Economic Progress (CSEP) estimates India’s merchandise e-commerce market at roughly $120–130 billion in 2024 and projects it could reach $300–350 billion by 2030, depending on the definition and coverage used.
The Economic Survey has also previously pointed to the Indian e-commerce sector crossing the $350 billion mark by 2030.
At that scale, payment reliability becomes more than a checkout issue.
A payment failure is expensive even when there is no fraud involved. The customer may abandon the purchase, move to another platform or simply decide not to try again.
Likewise, overly aggressive fraud controls can create a different problem: a genuine customer gets rejected.
A better payment intelligence system therefore has to balance three things at once:
conversion, security and cost.
[ALSO READ: Razorpay Launches AI-Powered ‘Agent Studio’ to Automate Payment Operations ]
Razorpay’s broader AI strategy
The company has spent 2026 adding AI at several layers of its payments and financial infrastructure.
In May, Razorpay launched a payment command-line interface designed to let developers and AI builders manage payments directly from their coding environments. In June, RazorpayX launched an agentic connected banking platform permitting businesses to use AI agents for functions such as payouts and collections.
Earlier in the year, Razorpay also announced work with NPCI and OpenAI on agentic payments, exploring how AI agents could eventually participate in UPI-based commerce.
If Vulcan succeeds, its contribution will not be the fact that Razorpay has built a “foundation model.” The contribution will be whether the model can make payments more reliable, fraud detection more precise, and checkout decisions more effective at scale.
