Razorpay Vulcan Isn't a Frontier Model. That's the Point.

Razorpay launched Vulcan today, a payments-specific AI foundation model built with NVIDIA and AWS, and the coverage around it has leaned on language borrowed from a different category entirely. This isn't a frontier model in the sense that word has meant every other time it's come up this year, and treating it like one actually undersells what Razorpay built and why it matters.

What Vulcan Actually Is

Razorpay describes Vulcan as a transformer-based model, proprietary from the ground up, with both its architecture and training data owned by the company. It's trained on roughly 3 trillion data points across 4 billion payments, reading about 3,000 signals per transaction. Razorpay is explicit that this is not a language model. It doesn't read or generate text. It's built to learn the movement of money, payment routing, fraud patterns, checkout friction, across a payments ecosystem that spans UPI, cards, net banking, wallets, and cash on delivery across hundreds of banks and gateways. NVIDIA supplied the accelerated computing, AWS the cloud infrastructure. Early components have already been running in production, with named customers including Blinkit, Bachatt, and redBus already seeing it in live payment flows ahead of today's full launch.

Why "Frontier Model" Is the Wrong Comparison

Every model that's earned that label in recent coverage, Fable 5, GPT-5.6, Claude Opus 5, Grok 4.6, Qwen's various releases, is competing on the same axis: general reasoning, coding, knowledge, multimodal understanding, benchmarked against each other across dozens of shared evaluation suites. Vulcan isn't attempting any of that, and Razorpay isn't claiming it is. Its actual competitive set is narrow, purpose-built financial infrastructure, the kind of system Visa, Mastercard, and PayPal already run internally for fraud scoring and authorization optimization, not a general assistant anyone would ask to write code or summarize a document.

That's not a knock. A payments model that's exceptional at one job, catching fraud and getting more transactions to actually complete, is more useful for that job than a general-purpose frontier model would be, the same way a purpose-built industrial tool beats a multitool for the one task it's designed around. The "frontier" framing that showed up in some coverage this week isn't a compliment being applied correctly, it's a category error that measures Vulcan against a scoreboard it was never trying to top.

India's First, Not the World's First

The claim that's held up consistently across coverage is narrower and more accurate: Vulcan is India's first transformer-based AI foundation model built specifically for payments. That version of the claim appears well-supported and hasn't been disputed by anyone covering the launch.

The global version doesn't hold. Stripe unveiled what it explicitly called the world's first AI foundation model for payments on May 7, 2025, at its Sessions conference, more than fifteen months before Vulcan, trained on tens of billions of transactions and capturing hundreds of signals per payment. Stripe reported its own early results at the time: an 80% reduction in card-testing attacks over two years using prior specialized models, then a 64% jump in attack detection for large enterprises once the foundation model itself went live. Worth noting Stripe leaned on the same infrastructure partner Razorpay did, its own launch doubled as the announcement of a deeper NVIDIA partnership. Razorpay being first in India is real and worth crediting on its own terms. Framed as a first globally, it isn't, and at least one outlet covering today's launch flagged that directly alongside the announcement.

The Numbers Are Claims Right Now, Not Verified Results

Every specific figure attached to Vulcan's early performance comes from Razorpay itself: payment success rates up 8 to 10%, 8 times more international card fraud caught, 5 times more fraudulent or disputed transactions identified without increasing alert volume, and 40% more shoppers seeing their preferred UPI app at checkout through Razorpay's Magic Checkout, which the company says adds 1 to 2 lakh completed purchases a month. None of that has independent audit or third-party verification attached to it yet. That doesn't make it false, Razorpay has real production deployment and named customers to point to, but it's the same standing every other vendor-reported benchmark on this publication gets until someone outside the company checks it.

Razorpay's AI Timeline

Vulcan reads differently once it's placed against Razorpay's own history, because this isn't a company trying AI for the first time and reaching straight for a foundation model. The pattern goes back at least three years.

Razorpay shipped an AI payment gateway optimizer in October 2023, a narrow tool that had 50 businesses on it at release. By 2024, that had scaled into a company-wide push, more than 60 new products with AI-driven features across the business, which leadership credited with part of a 24% revenue increase that year. The company then moved into agentic payments specifically, running early demonstrations with the National Payments Corporation of India and OpenAI, work that later extended into a partnership with Anthropic. That partnership became public and concrete in March 2026, when Razorpay launched Agent Studio at its FTX summit, built on Anthropic's Claude Agent SDK and described by the company as the first AI-native agent platform built specifically for payments. The stated result was merchant onboarding time cut from 33 minutes to just over 3. From there, Razorpay extended agentic commerce partnerships to Zepto, Swiggy, and Zomato, and separately partnered with Replit to embed UPI payments directly into that platform's app-building flow.

Read in order, the arc is coherent rather than opportunistic: a narrow optimizer, then broad AI tooling across the product line, then agentic experiments built on other labs' models, OpenAI first, then Anthropic, and now a foundation model Razorpay owns outright, architecture and training data both. Vulcan isn't Razorpay discovering AI on the way to an IPO. It's the point where a three-year pattern of leaning on other companies' models turned into building its own.

The underlying problem Razorpay is describing is real and specific. An internal study the company cited, covering 1.5 million shoppers and more than 51,000 businesses, found the same payment friction, failed transactions, delayed OTPs, quietly lapsed subscriptions, showing up identically from metro cities to small towns. That friction falls hardest on first-time and small-town digital shoppers, the people most likely to simply not try again after one failed payment. CEO Harshil Mathur framed the ambition plainly: "Every payment teaches the system something that makes the next payment better." Fixing that reliably, at the scale of India's payments ecosystem heading toward a projected $350 billion digital commerce market by 2030, is a genuinely worthwhile bet, and it's also happening as Razorpay heads toward an IPO, where a working, deployed AI infrastructure story is exactly the kind of thing worth having on the table before investors start asking questions.

Vulcan was never trying to be the smartest model in the room. It was built to be the one nobody notices, because the payment just went through.

Previously on TheQuery:

• Every Benchmark Claim Looks Convincing. Here's How to Actually Check One.
• Sarvam Code Launched With a New Three-Agent Architecture