Revolut Claims PRAGMA AI Outperforms Prior Systems After Training on 40 Billion Transactions

TL;DR: Revolut says its new PRAGMA AI model beats all earlier task‑specific systems after training on 40 billion real‑world transactions.

Context Revolut, a London‑based neobank, built PRAGMA to replace dozens of separate AI models that each handled fraud, credit scoring, customer lifetime value or product recommendations. The model processes structured financial events—such as a $14.99 subscription charge or a $150 holiday transfer—by encoding what happened, the specific details and the timestamp. Unlike text‑based AI, PRAGMA learns patterns across a user’s entire history, enabling it to predict fraud risk, creditworthiness or churn in milliseconds. Competitors like Mastercard (MA) and Visa (V) train AI on network‑level transaction data, while Plaid focuses on merchant name standardization; Revolut’s advantage lies in seeing the full arc of a customer’s financial life.

Key Facts Revolut states PRAGMA was trained on 207 billion data points drawn from 25 months of real customer activity, encompassing 40 billion transactions, app interactions and financial events from 25 million users across 111 countries. The model comes in three sizes: a 10 million‑parameter version for real‑time fraud checks, a 1 billion‑parameter version for accuracy‑focused decisions, and a mid‑tier option. All variants run on more than 200 NVIDIA H100 GPUs, supplied via Revolut’s cloud partner Nebius. Market data shows Mastercard (MA) trading flat at around $425 per share (market cap ≈ $425 billion), Visa (V) down 0.3% to about $260 per share (market cap ≈ $560 billion), and NVIDIA (NVDA) up 1.2% near $880 per share (market cap ≈ $2.2 trillion).

What It Means By consolidating multiple models into one foundation, Revolut reduces engineering overhead, shortens retraining cycles and speeds deployment across fraud, credit and retention use cases. The approach could pressure other fintechs and banks to reconsider fragmented AI stacks. Investors will watch whether Revolut’s performance claims translate into measurable improvements in fraud loss ratios or customer‑lifetime value metrics, and how rivals respond with their own unified models.