Razorpay's Vulcan AI flags a sovereign payments race

Razorpay has released production data for Vulcan, its AI payment model, reporting that it now processes close to four billion transactions annually, analyses roughly 3,000 signals per payment, and has delivered an 8–10% lift in transaction success rates alongside an eightfold improvement in international card-fraud detection. The numbers are notable in isolation. The wider significance, however, is what they signal about the architecture of global payments infrastructure over the next two years.

Vulcan does not operate in a vacuum. Razorpay is already collaborating with India's National Payments Corporation of India (NPCI) and Nvidia on a sovereign AI model for the Unified Payments Interface (UPI), the government-backed rail that processed more than 130 billion transactions in the 2023–24 fiscal year. That collaboration reframes the story from a single fintech's internal upgrade into an early data point in what is shaping up to be a jurisdiction-by-jurisdiction race to own the decision-making layer of national payment infrastructure.

The explainability gap

Vadim Drozd, CEO of payment infrastructure company FinteqHub, argues that the very accuracy gains Razorpay is celebrating contain a structural liability. A model that is five times more effective at catching disputed transactions will, by the same logic, silently block a meaningful share of legitimate payments, and at a scale of billions of transactions, even a fraction of a percent of errors translates into hundreds of thousands of real customers and businesses facing denials that no frontline support agent can explain. The only explanation, as Drozd puts it, "exists in the form of a vector of weights within the transformer."

Drozd draws a parallel with autopilot systems: the architecture is sound precisely because it is invisible to the user, but a bad outcome becomes catastrophic when no one on board can reconstruct why the system made the decision it did. Razorpay's claimed absence of an increase in false positives is a company-issued figure and has not been independently verified. Regulators have not yet developed a framework specifically for the explainability of automated payment refusals at this scale, though Drozd expects the first such requirements within 12–18 months, modelled on GDPR's right-to-explanation provisions.

A fragmented decision layer, not just fragmented rails

The deeper structural risk is cross-border. A sovereign AI payment model trained on the behavioural vocabulary of Indian UPI transactions will be poorly suited to European or Latin American traffic without substantial retraining. Drozd flags that the industry has already lived through one version of this fragmentation with stablecoins and cross-border regulatory divergence, where payment rails and jurisdictions pulled apart. What is different this time is that the fragmentation is happening at the decision-making layer itself, the algorithms that route, approve, and block payments in real time.

If every major economic bloc fields its own sovereign model within the 18–24 month window that industry observers are projecting, the result is likely to be a patchwork of domestically optimised AI systems that cannot readily communicate with each other. That is a significant problem given that cross-border payments remain among the fastest-growing and most profitable segments of the global payments market.

For investors and platform builders, the read-across is strategic. The first movers to embed explainability as a product feature, rather than a retroactive compliance retrofit, will hold a structural trust advantage. Conversion rates, which AI payment models are explicitly designed to improve, ultimately depend on consumer and merchant confidence in the system's decisions. Platforms that can demonstrate auditable, human-intelligible reasoning for a refusal will be better positioned to capture cross-border volume in a fragmented sovereign-AI landscape. The irony Drozd identifies is sharp: the technology deployed to remove friction from payments may introduce a new category of friction that is invisible, inexplicable, and disproportionately painful for the users it cannot explain itself to.