Razorpay Vulcan: India's First AI Payments Foundation Model

India's First AI Foundation Model for Payments

On 18 August 2026, Razorpay announced Vulcan — India's first transformer-based AI foundation model built exclusively for the payments domain. Trained on nearly 4 billion transactions and 3 trillion data points, the model analyses approximately 3,000 signals per transaction to make real-time decisions on routing, fraud detection, and checkout personalisation. Razorpay built Vulcan using NVIDIA's accelerated computing infrastructure and AWS, with Amazon SageMaker supporting both training and production deployment. Early results shared at launch show an 8–10 per cent improvement in payment success rates and an eightfold improvement in fraud detection accuracy against Razorpay's previous baseline.

What Makes Vulcan Different From Previous Payment Intelligence

Most payments intelligence in India has historically run as a collection of separate, purpose-built systems: one model for fraud, another for routing, a third for checkout optimisation. Vulcan changes that architecture. It is a unified foundation model trained on the full payments lifecycle, which means it can learn the correlations between signals that separate models never see — the relationship between a routing decision, a downstream fraud event, and a customer dropping off at checkout, for instance.

The transformer architecture Vulcan uses is the same family underlying large language models from Anthropic and OpenAI, but trained on payments data rather than natural language. Just as a language model learns patterns across billions of text sequences, Vulcan learns patterns across billions of payment sequences: the signals that predict success, failure, fraud, and abandonment from the moment a checkout initiates.

Four Capabilities Delivered by a Single Model

Vulcan's architecture delivers four integrated capabilities across the payments stack, each relying on the shared representations the model builds across all four tasks simultaneously.

Hyper-Precision Routing

Rather than sending each payment down a static preferred gateway, Vulcan models the real-time probability of success for each available route given the specific attributes of that transaction. The routing decision is made dynamically, reducing failed payment attempts before they reach the customer as an error screen.

Network-Level Fraud Detection

Vulcan's training across Razorpay's full merchant network means it can detect fraud patterns visible only at scale. A stolen card appearing across unrelated merchants within a short window is a classic example; more subtle patterns involve correlated signals across merchant categories that no single merchant's model can accumulate enough data to learn. Vulcan's eightfold fraud improvement is attributed primarily to this network-level signal access, which is only possible because Razorpay processes payments for millions of merchants simultaneously.

RTO Risk Intelligence

Return to Origin events — where a customer does not accept a Cash on Delivery order — are a structural cost unique to Indian e-commerce. Vulcan's RTO risk module flags high-risk COD orders before checkout completes, giving merchants the option to require prepayment, offer incentives, or decline COD on orders with a low predicted acceptance rate. This capability required India-specific training data that a globally trained model would not accumulate.

Predictive Checkout Personalisation

For each customer and transaction context, Vulcan recommends the payment method most likely to result in a completed payment, based on learned preferences and success history. The personalisation reduces the friction of selecting from a long payment method list and improves overall checkout conversion.

The Market Context: A $350 Billion Target

Razorpay has positioned Vulcan as foundational infrastructure for India's digital economy at its next scale. The company estimates that India's digital e-commerce market will reach $350 billion by 2030. At that volume, the failure modes of today's payments infrastructure — failed transactions, fraud losses, COD returns — compound significantly unless they are addressed at the foundation layer. Companies already deploying early Vulcan capabilities include Blinkit, Bachatt, and redBus.

What This Means for Software Teams Building in India

Vulcan's launch has practical implications for product teams building payments, lending, or commerce products on Razorpay's infrastructure. Capabilities that previously required separate integrations — a fraud vendor, a routing optimiser, a personalisation layer — are increasingly available as platform intelligence from the payments processor itself. For engineering teams evaluating payment infrastructure choices, foundation-model-powered intelligence is now a feature to ask about, not assume away.

More broadly, Vulcan demonstrates that domain-specific foundation models trained on proprietary operational data can outperform ensembles of specialist models across multiple dimensions simultaneously. This is a pattern Indian software teams should expect to see replicated across fintech verticals: lending, insurance, wealth management, and cross-border payments each have analogous datasets large enough to support similar approaches.

The Bottom Line

On 18 August 2026, Razorpay launched Vulcan, India's first transformer-based AI foundation model for payments, trained on 4 billion transactions and 3 trillion data points with infrastructure from NVIDIA and AWS. The model integrates hyper-precision routing, network-level fraud detection delivering an eightfold accuracy improvement, RTO risk intelligence for Cash on Delivery orders, and predictive checkout personalisation into a single unified system. Early deployments are live with Blinkit, Bachatt, and redBus. Razorpay's $350 billion e-commerce market projection for 2030 frames Vulcan as foundational infrastructure for a payments ecosystem growing faster than current technology was designed to handle.