Razorpay Unveils Vulcan, India's First AI Payments Foundation Model

The foundation model land grab has moved far beyond chatbots. On August 18, 2026, Bengaluru-based Razorpay — one of India’s largest payments providers — launched Vulcan, a transformer-based AI foundation model purpose-built for payments. Developed in partnership with NVIDIA and Amazon Web Services and trained on nearly 3 trillion data points spanning 4 billion transactions, Vulcan is being positioned as India’s first AI payments foundation model, and one of the first serious attempts anywhere to apply frontier-model architecture to the plumbing of money movement.

One Model to Rule Routing, Fraud, and Checkout

The core insight behind Vulcan is that the payments stack has historically been fragmented. Intelligent routing, fraud detection, risk scoring, and checkout personalization each ran as separate systems — separate models, separate rules engines, separate silos. A transaction that looks risky to the fraud layer might be exactly the one the routing layer should retry down a different bank rail, but the two systems could not talk to each other.

Vulcan collapses these functions into a single, unified intelligence layer. Because it is a transformer trained on the full corpus of Razorpay’s transaction history — successful payments, failed payments, chargebacks, disputes, and everything in between — the model learns the relationships between signals that legacy systems treated as independent. Razorpay co-founder and CEO Harshil Mathur describes it as a model that “understands the language of money,” and notes that the uncovered insight driving the project was precisely that success rates and fraud are deeply interconnected problems, not two separate ones.

The Early Numbers

Razorpay is backing the launch with early-deployment metrics that, if they hold at scale, are substantial for a market where payment failure is a daily friction point:

• 8–10% improvement in payment success rates — the ultimate measure of whether a payment simply works
• 8x more international card fraud detected and stopped
• 5x more fraudulent or disputed transactions identified overall

An 8–10% relative lift in success rates deserves context. India processes some of the highest transaction volumes in the world through UPI, cards, and net banking, but success rates on cards and certain bank rails have historically lagged global norms, with failures often landing in the 10–20% range depending on the method and issuer. Recovering even a few percentage points of failed transactions translates directly into recovered revenue for millions of merchants. At the scale of a $350 billion e-commerce market projected by 2030 — the figure Razorpay and AWS invoked at launch — a few points of success rate is billions of dollars in completed commerce.

Built, Trained, and Hosted in India

A notable part of the story is where Vulcan runs. The model was built, trained, and hosted in India, on infrastructure from NVIDIA and AWS. That detail matters on three fronts.

First, data residency. Indian financial regulators have grown increasingly assertive about where payments data lives and how it is processed. A domestically trained and hosted model sidesteps the cross-border data friction that has complicated other fintech infrastructure plays in the market.

Second, sovereign AI capability. Vulcan joins a small but growing list of large-scale models trained on Indian soil with Indian data, reflecting the broader push — from both government policy and large enterprises — toward sovereign AI infrastructure for critical sectors like finance.

Third, the NVIDIA–AWS alignment. For NVIDIA, payments is exactly the kind of vertical domain where accelerated computing demonstrates value beyond generic LLM training. For AWS, anchoring India’s first payments foundation model on its India region is a marquee reference win in one of its fastest-growing markets.

Vertical Foundation Models Are the Real Trend

Strip away the launch-week messaging and Vulcan is evidence of a structural shift in how AI gets deployed. The first wave of the foundation model era was horizontal: general-purpose LLMs from OpenAI, Anthropic, and Google competing on benchmarks and token pricing. The next wave is vertical: domain-specific models trained on proprietary, non-public data that no general chatbot will ever see.

Payments may be the most commercially compelling vertical of all. Transaction graphs, fraud patterns, issuer behaviors, and routing histories are proprietary, time-sensitive, and endlessly replenishing — a perfect training substrate. A general model cannot replicate that moat, because the data cannot be scraped from the web. Razorpay is not the first payments company to apply ML to fraud or routing — every serious processor has — but packaging it as a foundation model with a unified architecture across functions is a meaningful step up in ambition.

The competitive read is straightforward. Adyen and Stripe have invested heavily in ML-driven routing and fraud (Stripe Radar, Adyen’s single-stack risk engine). Razorpay’s Vulcan applies similar logic — one model, one data substrate, many functions — to the uniquely Indian problems of issuer heterogeneity, UPI-card mix, and a fraud landscape that has grown more aggressive as cross-border e-commerce has exploded.

What to Watch

Three open questions will determine whether Vulcan is a milestone or a press release. First, whether the 8–10% success lift persists as the model generalizes beyond early-traffic merchants, or whether it reflects cherry-picked deployment cohorts. Second, how the model handles adversarial drift — fraud rings adapt faster than almost any adversary a model faces, and a unified model is a single point of failure if it can be gamed. Third, whether Razorpay opens Vulcan’s capabilities beyond its own rails — as an API for banks and other processors — which would transform it from an internal advantage into platform infrastructure.

None of those caveats diminish what has been built. Training a transformer across 3 trillion payment data points and shipping it into live transaction flow is genuine engineering, and doing it entirely in India marks a coming-of-age moment for the country’s AI infrastructure story. The foundation model era is no longer just about who writes the best prose — it is about who builds the best models for the systems the economy actually runs on. With Vulcan, that frontier now includes the split second when you tap “Pay.”