Razorpay, the Indian omnichannel payments platform, has launched Vulcan — a transformer-based AI foundation model for payments, built with NVIDIA and AWS. The company says early deployments lifted transaction success rates by 8–10% and detected 8 times more international card fraud without increasing alerts.
Every one of those figures is company-reported, not independently audited. The model arrives as India’s e-commerce market races toward a projected $350 billion by 2030 — a scale where a few percentage points of failed payments add up to billions in lost revenue.
India’s digital commerce is on track to hit a $350 billion market value by 2030 — a number that has tripled in six years and now sits inside a widening band of forecasts, from a cautious $250 billion to an aspirational $345 billion, depending on who is counting and what they include. The growth is not the problem. The problem is that too many of those rupees still vanish at the last moment, when a UPI app times out, an OTP arrives late, or a card’s route turns weak. Consumers reach for cash instead. Merchants lose the sale.
Razorpay is betting that a single, continuously learning AI model — trained on 3 trillion data points drawn from 4 billion past transactions — can fix that. It has rolled out foundation model Vulcan with NVIDIA’s GPUs and Amazon SageMaker, claiming an 8–10% uplift in payment success rates, an eightfold jump in international card fraud detection, and a fivefold increase in catching disputed transactions. None of these numbers have been reviewed by an external auditor. The real question is not what Razorpay says it can do, but whether the numbers survive a second look.
The numbers Razorpay is asking the market to believe
Vulcan is not a chatbot. Razorpay’s Co-founder and CEO, Harshil Mathur, draws the distinction in one sentence: “Like LLMs are trained on text to understand language, Vulcan is trained on payments.” The model ingests roughly 3,000 signals per transaction — device type, issuer response times, channel behaviour, past fraud markers — and learns how money actually moves across India’s fragmented rails: UPI, domestic cards, net banking, wallets, and cash on delivery.
The output is four capabilities. Hyper-precision routing scores each payment path in real time, choosing the one with the highest predicted success. Network-level fraud detection spots compromised cards moving across multiple merchant sites before a single alert fires. RTO risk intelligence flags cash-on-delivery orders likely to be returned. Predictive checkout personalization learns which UPI app a shopper prefers and puts it first, lifting completion rates.
The interplay between these four systems is easier to see than explain — a breakdown of each capability makes the architecture visible.
Razorpay’s own testing, drawn from 51,000 businesses and 1.5 million shoppers, reports the same friction in Mumbai and a small town: failed transactions, drop‑offs, and OTP delays. The company says early live components pushed success rates up by 8–10%, boosted UPI app visibility by 40% — worth an extra 1–2 lakh completed purchases a month — and lifted Magic Checkout performance for merchants like Blinkit, Bachatt and redBus.
The figures are sharp. They are also entirely company-sourced. No third-party audit, no RBI or NPCI validation, no independent merchant benchmarking. That does not make them false. It makes them unverified — and for a model that Razorpay hopes will become the routing brain of a $350 billion market, verification is the point.
An AI model built for one country’s payments — and the global money waiting for it
India’s e‑commerce gross merchandise value hit $120–140 billion in 2024, up from roughly $14 billion a decade earlier. By 2030, forecasts range from $250 billion (Google‑Deloitte) to $345 billion (ANAROCK‑ETRetail), with BCG projecting $280–300 billion including services. Every projection carries a shared assumption: digital payments must work, everywhere, on the first try.
That reliability problem is what Razorpay’s Vulcan is designed to solve — and it is what draws the attention of global funds. Tiger Global, SoftBank, Sequoia/Peak XV, and strategic players like Walmart (Flipkart), Amazon, and Prosus (PayU) already carry deep India fintech exposure. Razorpay argues that an AI-driven lift in payment success rates would raise transaction volumes inside their portfolio companies.
Most Western processors — Stripe, Adyen, Checkout.com, Visa’s CyberSource — apply global risk engines that treat India as one more market on top of card rails. Razorpay’s model does something different: it trains a transformer architecture almost entirely on Indian transactions — UPI, domestic cards, net banking, wallets, and cash-on-delivery risk — and positions it as a shared intelligence layer across merchants. For Western payment gateways, that means success in India may increasingly depend on partnering with locally trained AI, not just porting a global model.
The next six to twelve months will show whether Razorpay’s numbers hold at scale. If the 8–10% uplift persists across a growing merchant base, Vulcan moves from promising pilot to infrastructure. If it stalls, the model looks more like a sales deck than a moat. The first regulatory signals from the RBI or NPCI on AI in payments will tip the balance either way.
Beyond the headline
The Money Trail
The commercial logic behind Vulcan is less about selling a standalone AI product and more about deepening Razorpay’s control over transaction flows. Improving success rates and cutting fraud losses increases processed volume and merchant dependency on its intelligence layer, strengthening the company’s position for large omnichannel contracts. The real financial upside lies in becoming the de facto routing and risk brain for India’s digital commerce, not in charging line‑item fees for model access.
The Reach
Razorpay’s India‑trained foundation model shifts the ground for global payment networks trying to grow in the country. As Vulcan internalises patterns unique to UPI, local issuers, and cash‑on‑delivery behaviour, international card schemes and cross‑border processors face a world where their generic risk engines may underperform. Gateways adopting Vulcan will quietly route more volume through paths that work best inside India’s ecosystem, privileging some networks, banks, and processors over others — and forcing Western players to adapt to Razorpay’s intelligence layer rather than the reverse.
What Isn’t Being Said
Public messaging around Vulcan emphasises reliability and fraud reduction but largely sidesteps questions of explainability, data governance, and competitive neutrality. A single shared model trained on billions of transactions effectively concentrates behavioural insight — and potential influence — over which routes, banks, and instruments succeed inside one private platform. It remains unclear how Razorpay will account for bias, access, or accountability when its model becomes embedded in everyday transaction decisions.
Where the money moves next
As Razorpay’s Vulcan begins processing live transactions, decisions made in the next few quarters will determine who captures the value from India’s payment reliability — and who gets left behind.
- Western e-commerce merchant operating in India
You should investigate integrating Razorpay’s Vulcan model or a similar AI‑driven solution. The company claims a 8–10% improvement in payment success rates and sharper fraud detection; even a fraction of that at your volume could recover significant revenue. Start by reviewing Razorpay’s Magic Checkout and Vulcan documentation for technical integration requirements and testing the impact on your conversion funnel with a small segment of traffic.
- US-based investor with APAC fintech exposure
Monitor Razorpay’s quarterly merchant adoption numbers for Vulcan and watch for disclosures from listed payment companies or funds like Prosus and PayU on AI‑driven payment performance. Sustained adoption at scale would signal that AI‑native routing is becoming a competitive differentiator, increasing the urgency to re‑evaluate positions in legacy processors and Indian fintechs that lag on in‑house AI capabilities.
- Western payment gateway or card network executive
Evaluate your existing routing and fraud engines in India against Razorpay’s locally trained model. If Vulcan’s uplift holds, your global risk engine may underperform on UPI and domestic card rails. Consider partnerships with Indian AI‑native processors or investing in a similar India‑specific foundation model tailored to the payment instruments that dominate local commerce.
- Western technology provider (AI/Cloud) targeting Indian fintech
Study Vulcan’s architecture — a transformer model on Amazon SageMaker with NVIDIA GPUs, ingesting thousands of signals per transaction — to identify where your own products or services plug in. The deal signals demand for accelerated computing and cloud‑native AI infrastructure in Indian fintech, and early movers who can offer the tooling to train and deploy country‑specific payment models will find a fast‑growing market.
Explainer
- Foundation model
- A large AI model trained on a broad dataset to learn general patterns, then adapted to many downstream tasks. In payments, a foundation model learns how money moves across instruments, gateways, and issuer behaviour. Unlike an LLM trained on text, Vulcan is trained on transaction data — routing paths, fraud signals, and checkout flows — so it can share insights across routing, fraud detection, and personalisation without separate models for each job.
- Transformer architecture
- A neural network design first used in language models that excels at spotting complex, long‑range relationships in data. It processes entire sequences at once rather than step by step. Razorpay adapted this to payment flows, letting Vulcan weigh thousands of signals per transaction — device, channel, issuer health — and decide the best route in real time, a leap from older models that examined fewer factors sequentially.
- Return-to-origin (RTO)
- In e‑commerce, an order that is returned by the customer or the courier, often in cash‑on‑delivery sales where no payment was captured upfront. RTO losses are a major cost for Indian merchants, sometimes reaching 15–25% of COD orders. Razorpay’s RTO risk intelligence uses purchase history and behavioural signals to flag orders likely to be returned, helping merchants tighten COD rules and reduce logistics and inventory waste.
- Magic Checkout
- Razorpay’s checkout product that pre‑selects the shopper’s most likely successful payment method, such as a preferred UPI app or saved card, and reduces the steps needed to complete a purchase. Powered by Vulcan’s personalisation signals, it aims to lift first‑attempt success rates by showing what works best for each user, not a generic list. Early reported gains include a 40% increase in shoppers seeing their preferred UPI app and tens of thousands more completed purchases per month.
RELATED STORIES
