Razorpay Builds Vulcan, an India-Made AI Foundation Model for Payments

Most of the conversation around foundation models in India has centred on language — sovereign large language models that can read, write and reason in Indian tongues. Razorpay has taken the same underlying idea in a very different direction. On 18 August 2026, the payments company unveiled Vulcan, which it describes as India's first transformer-based AI foundation model built specifically for payments.

Instead of predicting the next word in a sentence, Vulcan is trained to understand the patterns hidden inside how Indians pay — and to use them to make each transaction more likely to succeed, harder to defraud and smoother at checkout.

A foundation model, but for transactions

Vulcan borrows the architecture that powers today's large language models — the transformer — but repurposes it for the messy, high-volume world of digital payments. According to Razorpay, the model was trained on close to 3 trillion data points drawn from roughly 4 billion payments, and it weighs around 3,000 signals per transaction when making a decision.

Those decisions are practical rather than conversational: which payment route is most likely to go through, whether a transaction looks fraudulent, how risky a given attempt is, and how to reduce the friction that causes shoppers to abandon a checkout page. Razorpay says the model has lifted payment success rates by up to 10%, detected several times more international card fraud than its earlier systems, and improved the visibility of UPI apps at the point of payment.

The company is positioning Vulcan as proprietary and built from the ground up — both the model architecture and the training data are its own — and says it was built, trained and hosted in India. It was developed with NVIDIA's accelerated computing and Amazon Web Services' cloud infrastructure.

Why payments are a natural fit for AI

Payments are, in a sense, an ideal problem for a foundation model. Every transaction is a small bundle of structured signals — instrument type, issuing bank, network, time of day, device, merchant category and dozens more — and the outcome (success, failure, fraud) is known almost immediately. That creates an enormous, self-labelling stream of data to learn from.

India's payments landscape is also unusually complex. A single checkout can route through UPI, cards, net banking, wallets or EMI, across a long tail of banks whose systems vary in reliability. Small improvements in routing and success rates translate into large absolute gains when they are applied across billions of transactions. A model that can predict which path is most likely to clear, and reroute intelligently when one fails, addresses a genuine pain point for merchants who lose real revenue to failed payments.

The sovereignty angle

Vulcan lands squarely inside India's broader push for home-grown AI. The country has spent the past two years funding sovereign language models, building GPU capacity and encouraging domestic firms to train models on Indian data rather than depend on imported systems. A payments model trained on Indian transaction behaviour, owned by an Indian company and hosted domestically fits that narrative, even though it sits outside the government-funded IndiaAI programme.

There are open questions worth watching. Payments data is sensitive, and a model trained on billions of real transactions invites scrutiny over privacy, data governance and how those signals are used. Razorpay will also have to demonstrate that Vulcan's claimed gains hold up consistently across the diversity of India's banking rails, not just in aggregate. And rival payment aggregators are unlikely to cede the ground; expect competing claims of proprietary AI to follow.

The bigger signal

Whatever the eventual competitive dynamics, Vulcan is a notable data point in how India's AI story is maturing. The frontier is no longer only about chatbots and Indic language models. It is increasingly about vertical, domain-specific models trained on proprietary datasets — in payments, logistics, healthcare and beyond — where a company's own data becomes the moat.

For India's fintech sector, which already leads the world in real-time payment volumes, turning that transaction data into a purpose-built AI model is a logical next step. Vulcan is an early example of what that looks like in practice.

Sources

• Deccan Herald — "Razorpay launches Vulcan, AI foundation model built for payments": https://www.deccanherald.com/business/razorpay-launches-vulcan-ai-foundation-model-built-for-payments-4115864
• Business Standard — "Razorpay builds proprietary AI model Vulcan to boost payment success rates": https://www.business-standard.com/industry/news/razorpay-builds-proprietary-ai-model-vulcan-to-boost-payment-success-rates-126081801078_1.html
• Inc42 — "Razorpay Launches AI Foundation Model Vulcan To Expedite Digital Payments": https://inc42.com/buzz/razorpay-launches-ai-foundation-model-vulcan-to-expedite-digital-payments/
• MediaNama — "Razorpay launches Vulcan, an AI foundation model for payments": https://www.medianama.com/2026/08/223-razorpay-vulcan-ai-foundation-model-payments/
• AWS / Amazon Press Center — "Razorpay Launches Vulcan, India's First AI Payments Foundation Model, Fueled by NVIDIA and AWS": https://press.aboutamazon.com/aws-international/2026/8/razorpay-launches-vulcan-indias-first-ai-payments-foundation-model-fueled-by-nvidia-and-aws-re-architecting-payments-for-a-350-bn-e-comm-future-by-2030