Razorpay’s New AI Model Detects 8x More International Card Fraud

The Model That Reads Money Like Language

Bengaluru-based Razorpay launched Vulcan on August 18, a transformer-based foundation model built specifically for the payments ecosystem — the first of its kind in India. Unlike general-purpose LLMs that process language, Vulcan was trained to identify patterns in how payments flow across Razorpay's network: routing decisions, fraud signals, risk markers, checkout behavior. The model ingests nearly three trillion data points drawn from four billion transactions across seven million-plus merchants. Each transaction contributes roughly 3,000 signals the model learns from in real time.

The launch arrives as Razorpay prepares for a domestic IPO. The fintech unicorn confidentially filed draft papers in June, targeting a raise at a valuation, SiliconIndia reported. Investors including Y Combinator, Tiger Global, Ribbit Capital, Lightspeed, and GIC have put significant capital into the company since its 2014 founding by Harshil Mathur and Shashank Kumar.

In March, Razorpay rolled out Agent Studio, an AI platform letting businesses deploy agents for payment management, revenue recovery, and financial operations. Vulcan now sits beneath that layer as a common intelligence engine, replacing the fragmented machine-learning models that previously handled routing, fraud, and risk as separate tasks.

Razorpay positions Vulcan as a direct response to payment friction its own study documented across 1.5 million shoppers and 51,000 businesses. Failed transactions, delays, and drop-offs persisted from metros to small towns. The company estimates India's digital e-commerce market will reach a substantial size by 2030, and it frames Vulcan as part of the infrastructure that growth will require. Stripe introduced its own Payments Foundation Model in May 2025, trained on tens of billions of transactions using hundreds of signals. Razorpay's entry marks the first Indian counterpart, built on a domestic transaction corpus and tuned for UPI's particular failure modes.

The Problem It Solves: Siloed Models, Shared Friction

Razorpay's ecosystem spans UPI, cards, net banking, wallets and Cash on Delivery across hundreds of banks and gateways; a single purchase can hit half a dozen potential routes before it either settles or errors out. The company said traditional systems handle this with separate specialized models for routing, fraud, risk and checkout, each reading its own slice of signals. Vulcan replaces that stack with a shared intelligence layer that scores every route in real time and picks the healthiest one before the payment is attempted.

Early deployment numbers put the gains in concrete terms: an 8-10% improvement in payment success rates, 40% more shoppers seeing their preferred UPI app at checkout, and an additional 1-2 lakh purchases completed every month on Magic Checkout alone. The company said those figures come from live traffic, not lab tests, and that customers including Blinkit, Bachatt and redBus are already running Vulcan components in production.

The routing logic drives the UPI app preference metric. On Magic Checkout, the model now surfaces the shopper's most likely UPI app 40% more often, cutting the extra taps and app-switching that cause drop-offs. For merchants, the company projects reduced lost sales, fewer OTP drop-offs and lower return-to-origin rates on Cash on Delivery orders.

The same 3,000 signals per transaction that tell the model which bank gateway is healthy right now also flag anomalous patterns. Early components detected eight times more international card fraud and identified five times more fraudulent or disputed transactions without increasing alert volume. The network effect is the differentiator: a siloed fraud model sees only the transactions flowing through its own merchant. Vulcan, because it trains across seven million-plus merchants, can identify a compromised card at the first point of reuse — before any individual seller's system would have enough data to flag the pattern. The Reserve Bank of India's April 2026 shift to risk-based authentication, which allows platforms to determine security checks per transaction's risk profile, creates the regulatory frame in which Vulcan's real-time risk scoring now operates.

Architecture: One Transformer, Three Trillion Points

Vulcan is not a collection of models stitched together — it is a single transformer-based foundation model. The architecture mirrors the family that underpins GPT-style large language models: a multi-head attention mechanism that evaluates the relevance of every element in a sequence to every other element simultaneously, rather than processing them sequentially. In language, that sequence is tokens; in Vulcan, it is the 3,000 signals Razorpay extracts per transaction — merchant transaction history, payment instrument type, issuing bank real-time performance, shopper behavioral patterns, time of day, transaction amount, device fingerprint, and hundreds of contextual markers. The attention layers learn which signals matter for which decision, and how they interact, without hand-crafted feature engineering for each task.

Training at this scale demanded serious compute. NVIDIA H100 GPU clusters handled the three-trillion-point training run, while Amazon SageMaker provides the deployment substrate for real-time inference at enterprise transaction volumes. Razorpay CEO Harshil Mathur confirmed the model was built, trained, and hosted entirely in India on the company's private infrastructure, a design choice driven by RBI data localisation norms and the Digital Personal Data Protection Act. Before any transaction data reaches Vulcan, personally identifiable information is stripped from both individual and merchant records.

India's payments complexity is the reason a generic foundation model could not be fine-tuned. Over 100 payment instruments operate simultaneously, including UPI, credit, debit, net banking, wallets, and cash on delivery, routed through hundreds of banks and gateways, with nuanced regulatory workflows like dynamic OTPs and tokenisation, and distinct behavioral variations across Tier 1, Tier 2, and Tier 3 demographics. Stripe's payments foundation model, trained on tens of billions of transactions, solves for a different topology. Vulcan's differentiation is its orientation toward this specific ecosystem breadth, which no single processor's data could fully represent outside Razorpay's vantage point.

The self-learning architecture means the model updates continuously. Every payment teaches the system something that makes the next payment better — Mathur's framing of the launch as a starting point rather than a product. That continuous learning loop, combined with the breadth of signals and the unified architecture, is what lets Vulcan adapt to new use cases such as authentication, lending, and real-time underwriting without retraining the base model. The transformer backbone stays fixed; new heads attach for new tasks, trained on the same shared representations.

Regulation: RBI Tightens, Data Stays Home

The Reserve Bank of India has spent the past two years tightening the screws on every layer of the payments stack. In April 2026 the regulator cancelled Paytm Payments Bank's licence, citing compliance and governance lapses, and signalled winding-up proceedings that disrupted wallet and merchant settlement flows across the ecosystem. That same month the RBI proposed a one-hour delay for UPI and IMPS transfers above ₹10,000, introduced draft prepaid payment instrument guidelines with a ₹2 lakh monthly cap and ₹25,000 peer-transfer limit, and updated the e-mandate framework with stricter additional-factor authentication and pre-debit notification rules. Entrackr reported these moves collectively add friction for PhonePe, Google Pay, Paytm, and subscription-heavy platforms such as Razorpay and Cashfree Payments, which could see higher failure rates and re-authentication burdens.

Fraud prevention has become the regulatory north star. The Ministry of Finance has directed all commercial banks to implement MuleHunter.AI, a specialised real-time fraud detection tool, and the RBI announced that stricter guidelines against digital payment frauds take effect in 2027 as a one-year pilot. Mathur said Vulcan's claim of stopping 8x more international card fraud "directly tries to solve this urgent regulatory mandate." The RBI's 2022 Digital Lending Guidelines, which mandate direct loan disbursal, fee transparency, consent-based data collection, and regulated-entity accountability for lending service providers, set a precedent that payments foundation models would need to meet similar standards. Expected credit-loss rules kicking in April 2027 add another compliance layer.

On the hardware side, the Odisha government's semiconductor subsidy hike, flagged in Moneycontrol's MC Tech3 briefing alongside Razorpay's AI launch, reflects a broader Centre push to make India a global chip hub. ORF documented the Government of India's identified focus areas for semiconductor manufacturing, and the Odisha incentive expansion targeting three additional projects aims to reduce import dependence for the GPU clusters that models like Vulcan demand. Razorpay developed Vulcan in partnership with Nvidia and AWS, and Mathur acknowledged the training compute requirement was significant. The company is funding development internally, but the subsidy environment lowers the long-term cost curve for domestic AI infrastructure that fintechs will need as they scale foundation models beyond payments into real-time underwriting and credit evaluation.

The regulatory timeline creates urgency. With the 2027 fraud-guideline pilot approaching, the DPDP Act in force, and RBI supervision shifting from entity-level to structural, shaping how fintechs structure payments, partnerships, and lending models, Vulcan arrives as both a compliance instrument and a competitive moat. Customer protection has moved to centrestage, Entrackr noted, "beyond the viability of institutions that form part of the RBI universe of regulation."

What Comes Next: API, Verticals, Agents

Razorpay's immediate next step is exposing Vulcan as a programmable API so merchants can embed the model's routing and fraud logic directly into their own checkout flows. The company's own roadmap for 2024-2025, published alongside the SPRINT/26 "100+ Launches One Blueprint" initiative, positions the API layer as the primary distribution vector — moving Vulcan from an internal optimization engine to a merchant-facing primitive.

The international footprint already in place gives the API a ready addressable base. Razorpay today processes payments in over 100 currencies, supports card acquiring across 180 countries, and runs a MoneySaver Export Account that lets businesses open virtual accounts in 200-plus jurisdictions while cutting cross-border transfer costs by up to 50 percent. Those rails mean a merchant in Singapore or the UAE can call the Vulcan API without negotiating new banking relationships — the settlement, compliance, and currency-conversion plumbing is already live. The company's "Without an India Entity" positioning, highlighted in its own materials, signals that the API is being designed for global developers from day one, not as an afterthought bolted onto a domestic core.

Vertical expansion follows the same pattern. The gateway today counts roughly 70,000 e-commerce merchants, 21,000 education institutions, 2,800 BFSI clients, and 13,500 SaaS businesses. Vulcan's transformer architecture is built to ingest heterogeneous label sets without retraining from scratch.

Agentic workflows represent the third growth axis. Razorpay's "Agentic Payments" framework, which turns chat conversations into checkouts, and its "Agent Studio" for delegating operational tasks to autonomous agents both rely on real-time risk scoring to authorize or block machine-initiated transactions. The "Payments for AI Builders" program, which ships one-click payment nodes for n8n, Replit, and Vercel workflows, extends that logic into the developer toolchain.

Competitive pressure accelerates the timeline. Razorpay's SPRINT/26 cadence, featuring 100-plus product launches in a single cycle, is explicitly designed to compress the feedback loop between model improvement and merchant adoption. If the API launches on schedule, the first external developers will stress-test Vulcan on live traffic before the end of 2025, and the resulting telemetry will feed the next training run.

The model that reads money like language has just learned its first sentence. The next one will be written by every merchant, shopper, and agent that calls its API — each transaction a token, each pattern a grammar, the whole corpus growing toward a fluency no single player could have authored alone.

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.