Razorpay’s launch of Vulcan, a transformer-based AI foundation model trained on 3 trillion data points from 4 billion payments, signals a larger shift: emerging-market payment platforms are now building proprietary AI rather than renting it from the US. For European SMBs, the lesson isn’t technical—it’s about cost, latency, and strategic autonomy in your payment infrastructure.
There’s a pattern in emerging-market fintech that most European SMB leaders have missed. It’s not about who has the fastest GPU or the most capital. It’s about who owns the intelligence layer underneath their payment network.
Last month, Razorpay announced Vulcan—India’s first transformer-based AI foundation model built specifically for payments. Not a general-purpose LLM. Not ChatGPT bolted onto a payments dashboard. A single unified model trained on 3 trillion data points across 4 billion payments, running in production, improving with every transaction.
Here’s why this matters to you: the moment a payment platform stops outsourcing its fraud detection, routing, and checkout logic to third-party APIs and instead builds a proprietary foundation model, the economics of the entire stack change. Speed improves. Cost per transaction drops. Latency becomes predictable. And most critically, the business owns the asset—not a SaaS vendor in Silicon Valley.
What Vulcan Actually Does (And Why It’s Architecturally Different)
Razorpay processes roughly 4 billion payments a year. Until now, like most payment processors globally, the company ran separate machine-learning models for different problems: one for routing, one for fraud, one for risk scoring, one for checkout personalization. Each model saw only its own slice of the network.
Vulcan inverts that logic. A single foundation model trains across the entire transaction graph at once. It learns to recognize patterns in how money actually moves through UPI, cards, net banking, wallets, and cash-on-delivery—across hundreds of banks and gateways simultaneously. The result: when a stolen card hits the network, Vulcan can flag it the moment it appears at multiple unrelated merchants, before any siloed fraud model would have enough data to fire an alert.
Early results in production: 5x more fraudulent transactions caught without increasing false-positive alerts to merchants. 8–10% improvement in payment success rates. 40% better UPI app selection at checkout, driving an extra 1–2 lakh completed purchases per month.
Razorpay isn’t charging merchants separately for Vulcan. It’s bundled. That positioning is important—the company sees this as infrastructure, not a feature. It deepens Razorpay’s role as the connective layer across India’s entire payments ecosystem.
Why This Accelerates Regional AI Autonomy
The move matters because it demonstrates a playbook that emerging-market fintechs are repeating: build domain-specific foundation models, train them on your own data, deploy them in your own region, and own the compounding returns.
India’s payments complexity is extreme. Your typical European checkout works across maybe 5–10 payment methods at low latency. India has to route through dozens of methods, manage high-latency branches, handle cash-on-delivery risk, and account for variable internet reliability. Generic LLMs trained on English-language financial documents don’t understand this. A model trained on 3 trillion real Indian payment signals does.
That’s a direct cost advantage. Razorpay doesn’t need to pay API calls to a US fraud-detection vendor or rent inference time on a third-party LLM. The model runs on their own NVIDIA-powered infrastructure and AWS. One bill, one master, one model that learns.
Compare this to how many European payment processors still work: they layer Stripe’s APIs, third-party KYC vendors, maybe some OpenAI calls, a bit of custom ML. Elegant architecturally, but each step adds latency and monthly fees. And none of it belongs to you.
The Emerging-Market Template (And Why European SMBs Should Care)
Imagine a 50-person B2B payments firm in Germany handling 500 million euros in cross-border transactions annually. Today, they probably use Stripe Connect, some bolt-on fraud API, maybe a Sentry AI tool for compliance. All cloud-managed. All third-party. If Stripe raises prices, or changes rate limits, or deprioritizes their use case, the firm has no recourse.
What Razorpay is proving is that a regional platform with enough transaction volume can build its own foundation model in 18–24 months (Razorpay spent years on this) and drop both latency and cost significantly. For that German firm: if they reached 5–10 billion transactions a year, they could argue for building their own proprietary routing and fraud detection layer instead of renting it.
That’s the threshold. Below it, SaaS is efficient. At that scale and above, ownership wins.
The other lesson: India’s regulatory environment actually helped here. The DPDP Act (India’s data protection law) and RBI guidelines push fintechs to keep data in-country and maintain explainability over AI decisions. That means Razorpay had to build its own model anyway—it couldn’t outsource model training to a US cloud vendor without friction. Regulatory pressure became competitive advantage.
Europe’s AI Act and GDPR work similarly. If you’re subject to strict explainability rules or data residency mandates, building your own domain-specific model becomes cheaper and faster than negotiating compliance with a third-party SaaS vendor.
The Caution: Ownership Isn’t Risk-Free
Building your own payment foundation model means owning the model risk. If Vulcan misbehaves—makes routing decisions that systematically disadvantage certain merchant segments, or fraud detection that flags valid transactions at higher rates for one demographic—Razorpay owns that liability, not NVIDIA or AWS.
That’s why Razorpay hasn’t published independent benchmarks on Vulcan’s performance. The company reports early results (8–10% success-rate improvement, 5x fraud catch rate), but hasn’t shared methodology, baseline comparisons, or sample size. That’s a red flag if you’re evaluating a similar play for your own business.
Also, data concentration. Vulcan gets smarter with every transaction on the Razorpay network. That creates a compounding moat—but also a compounding regulatory target. If Razorpay’s data practices come under scrutiny, or a regulator demands to audit the model’s training data, the company has nowhere to hide. Everything lives in their own systems.
What to Actually Do If You Run a Mid-Market Payment or Lending Business
If your transaction volume is below 1 billion annually, stick with SaaS. Stripe, Adyen, whatever—the efficiency and compliance leverage outweigh the cost. You’re not at the scale where a proprietary model pencils out.
If you’re at 1–5 billion and growing, start the conversation with your AI/ML leadership: what would it take to build a custom fraud detection or routing model for our specific corridors or segments? Not the whole thing yet. One narrow, high-ROI model. Train it on your own data. Measure the latency and cost against your current vendor stack. That’s your benchmark.
If you’re above 5 billion transactions annually and you’re still outsourcing all your payment intelligence to third parties, you’re leaving 15–30% on the table in annual costs. Start building internal capability or acquire it. Razorpay shows the path; the window to own that infrastructure is now.
The Bigger Play
What Razorpay is really signaling is that the era of generic AI is ending for high-frequency, low-latency financial infrastructure. Just as open-weight AI from China is forcing European SMBs to rethink vendor strategy, domain-specific foundation models from emerging markets are forcing payment companies everywhere to ask: why are we renting intelligence from the US when we can own it?
That question doesn’t require you to rebuild your entire stack. But if you process payments or credit decisions, and you’re not thinking about what your proprietary model for that domain could look like, you’re betting your margin on someone else’s infrastructure. And in fintech, margins matter.
0 Comments