Every online payment in India runs a small gauntlet: the right bank, the right network, a fraud check, an OTP that has to arrive on time. When any link breaks, the sale is lost, and for a first-time shopper a single failure is often enough to send them back to cash. Razorpay’s answer is Vulcan, which it calls India’s first transformer-based AI foundation model built specifically for payments, launched this week with hardware and cloud muscle from NVIDIA and AWS.
The pitch rests on scale. Vulcan was trained on nearly 3 trillion data points drawn from around 4 billion payments, and weighs roughly 3,000 signals on every transaction, from the merchant and the card to the issuing bank and the gateway. The idea is to replace the usual patchwork, one model for routing, another for fraud, another for checkout, with a single system that sees the whole picture. Razorpay founder Harshil Mathur frames it by analogy: where a language model learns from text, Vulcan learns from how money moves. It is not a chatbot, and Razorpay is clear about that.
What Razorpay says it does
The early numbers Razorpay is putting forward are eye-catching. It reports an 8 to 10% lift in payment success rates in live deployments with customers including Blinkit, Bachatt, and redBus, eight times more international card fraud caught, and five times more fraudulent or disputed transactions flagged without raising the overall number of alerts. On its Magic Checkout product, it says 40% more shoppers were shown their preferred UPI app, worth an extra one to two lakh purchases a month.
Worth a caveat, though. These are Razorpay’s own early-deployment results, not independent benchmarks, and the company has not published the baselines those multiples are measured against or a technical paper anyone outside can scrutinise. An “8x” improvement means little without knowing 8x of what. The direction is plausible and the data advantage is real, but the specific figures deserve the usual pinch of salt that greets any vendor’s launch-day metrics.
The quieter question is the data
There is a bigger issue the announcement glides past. A model this powerful is powerful because it learns across merchants, spotting a stolen card the moment it surfaces at an unrelated seller. That cross-merchant intelligence is the whole point, and also the sensitive part. Razorpay has not said whether a merchant can keep its transactions out of Vulcan’s training, what happens to what the model has already learned if a merchant leaves, or how the arrangement squares with India’s tightening data-protection rules. For a system trained on other people’s payment data, those are not footnotes.
None of this is unique to Razorpay. Stripe launched a strikingly similar payments foundation model in 2025, so the concept is already proving out globally, and Razorpay is applying it to the messy particulars of UPI, cards, and cash on delivery. The timing is not accidental either: the company filed confidentially for an IPO in June, and a proprietary AI layer is exactly the kind of moat public-market investors like to see.
So is Vulcan a genuine leap or a well-timed rebrand of machine learning Razorpay already ran? Probably somewhere in between. The unified-model approach is a real architectural step, and the data behind it is genuinely deep. Whether it delivers for merchants, and treats their data fairly while doing so, is what the next year will show.
This website uses cookies.