4 billion payments. 3 trillion data points. One foundation model trained on all of it.
Razorpay announced Vulcan on August 18, 2026 and published a technical explanation in September. Razorpay describes it as a transformer-based model for payment decisions, built with NVIDIA computing and AWS infrastructure. Its launch figures are company-reported; neither the announcement nor this article establishes that all training and hosting stayed within India.
Archive note: Assigned to the August 18 issue to match Vulcan’s launch. First published September 10; revised September 29. The later technical evidence remains labeled below.
Context and scope
- Payment fieldsBuild a shared representation
- Recent stateAdd route health where relevant
- Task headScore an eligible option
- EvaluationCompare against the incumbent
Conceptual sequence for understanding the system; implementation details vary by product and deployment.
Until now, the fintech industry addressed payment challenges through isolated, specialized subsystems: - A heuristic algorithm for gateway routing (e.g. switching between HDFC, ICICI, Axis, or SBI based on error code averages). - A separate scoring engine for fraud and risk detection (e.g. velocity limits, IP geolocation checks). - An independent recommendation system for checkout payment method sorting (e.g. showing UPI vs Credit Card).
Razorpay asked a foundational question: What if one model could understand how money moves?
Just as Large Language Models (LLMs) are trained on trillions of text tokens to understand human syntax and semantics, Vulcan is trained on financial transaction tokens to understand the physics of money flow.
Razorpay Vulcan AI Foundation ModelExecutive summary & Production Beta Outcomes
Razorpay reports early results for Vulcan across several payment tasks. These are company-reported comparisons against each task's incumbent system; they are not independent benchmarks or one combined experiment: - +8% to 10% improvement in payment success rates—the ultimate benchmark of whether a transaction completes smoothly. - 8x more international card fraud detected across cross-border e-commerce rails. - 5x more fraudulent or disputed transactions identified before chargebacks occur. - 1–2 lakh additional purchases completed every month through real-time checkout personalization. - Shared Representation: A reusable backbone learns patterns across payment records, while each application still needs its own inputs, labels, evaluation, and serving path.
Architectural Breakdown: How Vulcan Tokenizes Financial Flow
┌────────────────────────────────────────────────────────┐
│ 1. MULTIMODAL TRANSACTION TOKENIZATION │
│ UPI VPA · Bank BIN · Device Fingerprint · Merchant Cat │
└───────────────────────────┬────────────────────────────┘
│ Sub-Millisecond Feature Vector
┌───────────────────────────▼────────────────────────────┐
│ 2. VULCAN TRANSFORMER CORE (NVIDIA) │
│ Set Transformer · Masked-Field Prediction │
└───────────────────────────┬────────────────────────────┘
│ Unified Inference Vector
┌───────────────────────────▼────────────────────────────┐
│ 3. MULTI-TASK PREDICTIVE HEADS │
│ [Dynamic Route] ── [Fraud Risk] ── [Personalize] │
└───────────────────────────┬────────────────────────────┘
│ Direct Execution
┌───────────────────────────▼────────────────────────────┐
│ 4. TASK-SPECIFIC EVALUATION │
│ Compare routing, fraud, and checkout separately │
└────────────────────────────────────────────────────────┘1. Tokenizing the Anatomy of a Payment
Language models treat words as subword tokens. Razorpay describes Vulcan as a Set Transformer over structured payment fields, trained with masked-field prediction. Examples of potentially relevant fields include: - Temporal Tokens: Timestamp, day-of-week cadence, banking holiday states, and real-time network traffic congestion. - Entity Tokens: Card Issuer BIN, Acquiring Gateway, UPI Handle (VPA), Merchant Category Code (MCC), and device platform. - Behavioral Velocity Tokens: User transaction frequency, cross-merchant spending patterns, and OTP delivery latency windows.
Razorpay describes a separate recent-context encoder for live route health. That context is combined with the payment representation to score eligible routes. The model ranks routes already allowed by merchant configuration; it does not create new payment paths.
2. Unifying the Payments Stack
The shared backbone supports distinct decision layers. Razorpay has not published a universal under-15-millisecond inference guarantee or said that one pass replaces every checkout service:
| Dimension | Legacy Payments Architecture | Razorpay Vulcan Foundation Model |
|---|---|---|
| Model Structure | Separate task-specific systems | Shared backbone with task-specific decision heads |
| Training Data | Not stated in the public comparison | Razorpay reports 4 billion payments and nearly 3 trillion data points |
| Infrastructure | Not stated in the public comparison | NVIDIA computing and AWS services, according to Razorpay |
| Inference Latency | Depends on each service and payment flow | Low-latency routing is a design requirement; public per-request latency is not specified |
| Adaptability | Task-specific updates | Shared pretrained representation plus separately evaluated task heads |
3. Infrastructure and deployment questions
Razorpay describes Vulcan as built for India's payment environment in collaboration with NVIDIA and AWS. Deployment location, model quantization, and serving internals should be confirmed from technical documentation before being treated as implementation facts.
| Company-reported launch result | Comparison described by Razorpay | Reading caution |
|---|---|---|
| 8–10% payment-success improvement | Routing against the incumbent routing system | Depends on traffic and evaluation window |
| 8× international-card fraud detection | Signals at comparable alert volume | A detection ratio is not a universal fraud rate |
| 5× fraud or dispute identification | Same alert-volume framing | Later outcomes and false positives matter |
Risks and limits
Model Drift vs Sudden Banking Infrastructure Flips. When a bank changes availability, historical patterns can mislead routing. A production design should combine recent route-health signals with fallback rules and monitor drift.
False Positive Mitigation in High-Value Orders. Increasing fraud detection by 8x requires careful calibration to ensure legitimate high-ticket luxury or B2B transactions are not inadvertently blocked.
Deployment Boundaries. Public material does not establish that Vulcan weights run inside a mobile SDK. Model serving, network latency, and fallback behavior need separate measurement.
The Architecture Pattern for Fintech Builders
For engineers designing AI-native transaction systems:
1. Unify the Data Plane: Ingest payment metadata, banking telemetries, and user interactions into a single streaming vector bus (Kafka/Kinesis). 2. Train for Multi-Task Prediction: Structure the transformer with multiple output heads sharing the same foundational representation (routing, fraud, checkout ranking). 3. Strict Latency Budget: Measure model, network, and fallback latency against the checkout service-level objective. 4. Close the Feedback Loop: Automatically label transaction outcomes (settled, failed, chargeback) to continuously fine-tune model weights.
Conclusion
Vulcan shows a credible way to reuse learned payment representations across routing, fraud, delivery risk, and checkout ordering. Razorpay's launch results are promising company reports. The durable engineering test is whether each decision improves on later traffic, at fixed operating constraints, without hiding new failure modes.
September 2026 evidence update
Razorpay’s launch release reports the 4-billion-payment training scale and early task outcomes; its technical article explains decision-time features and out-of-time evaluation. These remain company-reported results. The article now avoids an unsupported claim that all training and hosting stayed in India.
Sources and further reading
- Razorpay Vulcan engineering article — backbone, task heads, routing context, and evaluation limits
- NVIDIA AI Models — GPU-accelerated training architectures and TensorRT inference in fintech
- AWS Cloud India FinTech Solutions — general platform context, not proof of Vulcan data residency or serving configuration