.png)
The Payments Foundation Model Race: Stripe, Ant and Razorpay Are Rebuilding the Decision Layer
Meta title: Payments Foundation Models: The Race for the Decision Layer | Quantum Payments Meta description: Stripe, Ant International and Razorpay are replacing task-specific models with payments foundation models. What merchants must ask before they buy one - measurement, training data, portability and concentration risk. Suggested slug:payments-foundation-model-race-decision-layer Suggested URL:https://www.quantumpayments.io/post/payments-foundation-model-race-decision-layer Primary keyword: payments foundation model Secondary keywords: AI payment routing, payment success rate, AI fraud detection payments, AI payments, agentic payments, payment orchestration, unified commerce
Executive summary
A payments foundation model is a shared AI backbone trained on the structure of transactions,then adapted through task heads for routing,fraud,authentication,retries and checkout decisions. Stripe,Ant International and Razorpay are moving payments intelligence from isolated models to a central decision layer. For merchants,the strategic question is no longer which gateway has the best model,but what representation,data access and dependency they are buying.
What is a payments foundation model?
For the past decade,payment providers have generally built separate machine-learning models for separate jobs:fraud scoring,payment routing,authentication,retries and payment-method ordering. That approach created useful optimisation,but it also created duplicated data pipelines,slow model development and competing decisions across the same checkout.
A payments foundation model changes the architecture. It learns broad relationships within payment records before being adapted to particular decisions.
A backbone is the reusable core that converts payment data into a learned representation. A task head is the decision-specific layer that uses that representation to answer a particular question,such as whether to route through Acquirer A or Acquirer B.
Razorpay’s Vulcan illustrates the design. Its self-supervised Masked-Field Prediction task hides fields in structured payment records and asks the model to reconstruct them from the fields that remain. Because payment fields do not have a meaningful left-to-right order,Vulcan uses a Set Transformer rather than treating a transaction like a sentence.
The result is not an LLM for payments. It is a domain model that learns the structure of money movement.
The race is about ownership of the representation
Ant International’s 21 September launch makes the scale of this shift explicit. Its Antom 3-in-1 Transformer processes sequential,tabular and graph-based data simultaneously,runs on more than 10 billion parameters and processes 90 terabytes of data annually. It supports fraud detection,scam prevention and payment optimisation across Ant International’s wider financial stack.
The company’s second model,FalconTST,targets FX and cash-flow forecasting. Ant International cites 30-60% reductions in corporate FX hedging and allocation costs in some cases. Its security architecture combines the Antom model with Know Your Agent controls and AgentSafePay,which offers a fund-back guarantee against certain agent-specific risks.
This is significant because the model is not being positioned as a feature inside checkout. It is becoming the intelligence layer across payments,accounts,FX,treasury and agentic commerce. Ant International says its network reaches 150 million merchants,two billion consumer accounts and 210 markets,with more than 25 million transactions processed daily.
Razorpay’s Vulcan takes a more deliberately task-oriented route. One backbone serves four decisions across the same purchase journey:
Payment routing.
Fraud detection.
Cash-on-delivery delivery risk.
Checkout payment-method ordering.
Its technical account is important because it describes the controls behind the model rather than only its headline performance. Direct identifiers such as names,email addresses,phone numbers and bank account numbers are excluded before training data is written. Stable categorical references are hashed. Each application also has a decision-time cut-off so later outcomes cannot leak into the features used at the moment of decision.
Razorpay also separates route-influenced failures,including declines,timeouts and some late failures,from customer-side constraints such as insufficient funds,wrong PIN,limit exceeded and abandonment. That distinction is essential. A routing model cannot solve a wrong PIN,and counting it as a routing failure teaches the wrong lesson.
Stripe shows where the decision layer is heading
Stripe’s Payments Foundation Model is less a single launch announcement than an architecture embedded across its payment lifecycle. Stripe says its fraud training data expanded from roughly 800 million to more than 11 billion historical transactions,and that it is building multitask models that predict several outcomes while sharing representations across tasks.
The commercial implications are already visible. Stripe reports that Authorization Boost lifts acceptance rates by 3.8% on average,cuts processing costs by up to 3.3% and recovers about 20% of false declines. Across its lifecycle optimisation,Stripe cites as much as US$27 billion in annual incremental revenue,an average 32% fraud reduction while blocking fewer than 0.05% of legitimate transactions and a 51% reduction in dispute rates through deflection and pre-dispute resolution.
The important point is not any single percentage. It is the widening optimisation surface. Stripe tests checkout ordering,risk interventions,authentication paths,routing,ISO 8583 message formatting,retries,token use,clearing and disputes. It also applies network tokens selectively because tokenisation does not improve approval rates for every issuer or transaction pattern.
That is the practical meaning of a foundation model. The provider can carry learning from one payment stage into another,then measure the commercial effect across the lifecycle.
Visa’s TREASURE payment foundation model reportedly delivered a 111% improvement in abnormal-behaviour detection over previous production systems. Mastercard has also deployed a generative AI engine. The competitive boundary is moving from “who has a fraud model?” to “who controls the most useful payment representation and the decisions built on top of it?”
Task-specific models versus a payments foundation model
Dimension | Task-specific models | Payments foundation model |
Architecture | Separate models for fraud,routing,authentication and retries | Shared backbone with task-specific heads |
Data breadth | Usually optimised around one outcome | Learns relationships across payment methods,merchants,issuers,devices and time |
Measurement | One primary metric per model | Different metrics for each head,with shared representations |
Failure modes | Conflicting decisions,duplicated features and stale rules | Leakage,opaque trade-offs,model drift and errors that propagate across tasks |
Merchant risk | Switching one function may be possible | Greater dependency if one provider controls acceptance,risk and recovery decisions |
The strategic shift explains Bain & Company’s framing of the market moving from top of wallet to top of model. In an agentic payments environment,the agent may choose the merchant,method,credential and route. Brand preference still matters,but it must become machine-readable value.
The merchant question is now governance,not just performance
A foundation model can improve payment success rate and reduce fraud,but it also concentrates influence. If the same provider decides which traffic is accepted,which route is used,which payment method appears first and when a retry occurs,the merchant may be buying performance at the cost of strategic visibility.
That makes payment orchestration more than a routing feature. It becomes a control-plane decision. Merchants need to understand what data trains the shared model,what outcomes the provider can see and whether the merchant receives a decision layer of its own.
The same applies to unified commerce. A retailer operating online,in-store and through a marketplace should ask whether intelligence transfers across channels or whether the provider only optimises one surface. The value of an AI payments platform increases when checkout,terminals,subscriptions,inventory,accounting and reconciliation share usable outcomes.
Merchant checklist:10 questions to ask
What data is the model trained on,and is our transaction data used to train a shared model?
Can the provider see final authorisation outcomes,or only authentication and gateway responses?
Are improvements measured using out-of-time testing rather than random splits?
Are fraud results compared at a fixed alert volume,and are results reported by payment flow and merchant?
Do we receive a merchant-specific decision layer on top of the shared backbone?
Which decisions are covered:routing,fraud,retries,authentication,checkout ordering,delivery risk and refund risk?
How are route-influenced declines separated from customer-side failures?
Is tokenisation applied selectively based on issuer and payment behaviour,or universally?
What enrichment can we supply,including margin data,product metadata and risk preferences?
What does the contract say about model portability,data use on exit,access to final outcomes,concentration risk and classification of agent-initiated traffic?
A foundation model is only as good as the outcome data it is allowed to learn from. A single model in front of every decision may improve performance,but it also creates a single-vendor dependency that belongs in the risk register.
FAQ
What is a payments foundation model?
It is a large AI model trained to learn the structure of payment activity across merchants,methods,issuers,devices and time. A shared backbone can then support task-specific heads for routing,fraud,authentication,retries and other payment decisions.
How is it different from a traditional payment model?
A traditional model is usually built for one outcome,such as fraud or approval. A foundation model learns reusable representations that can support multiple decisions. Each task still requires its own data,labels,controls and measurement.
Does a foundation model guarantee a higher payment success rate?
No. Performance depends on the quality and breadth of the training data,the final outcomes available,the decision-time features,the route eligibility and the evaluation methodology. Out-of-time testing and controlled production experiments are essential.
Does a foundation model replace payment orchestration?
Not automatically. It can improve orchestration by ranking eligible routes or payment methods,but orchestration still requires merchant rules,contracts,rail access,failover controls and operational governance.
What is the main risk for merchants?
Concentration risk. If one provider controls routing,fraud,authentication,retries and checkout ordering,the merchant may have limited visibility into decisions and limited portability if it wants to change providers.
How do agentic payments change the model question?
Agentic payments introduce a new decision-maker. Providers must distinguish agent-initiated traffic,verify the agent’s authority and apply controls for prompt injection,misinterpreted intent and mandate breaches. Merchants should understand how that traffic is classified and measured.
Sources
Publishing and handover
Blog title: The Payments Foundation Model Race:Stripe,Ant and Razorpay Are Rebuilding the Decision Layer Blog URL:https://www.quantumpayments.io/post/payments-foundation-model-race-decision-layer Wix slug:payments-foundation-model-race-decision-layer Publishing instruction: Schedule in Wix for Monday 28 September 2026 at 7:30am AEST,at least one hour before go-live,with the scheduled post and assets confirmed no later than 6:30am AEST. Primary angle: Payments foundation models are becoming the decision layer behind routing,fraud,authentication,retries and checkout. The merchant buying question is shifting from model performance to representation ownership,training data,measurement,portability and concentration risk. Hero image CDN URL:https://cdn.marblism.com/DadBZvltvMb.webp Inline image one CDN URL:https://cdn.marblism.com/vyjiYKkQ4M-.webp Inline image two CDN URL:https://cdn.marblism.com/tMT55hu-BV5.webp Relevant tags: Payments foundation model,AI payments,payment orchestration,AI payment routing,AI fraud detection,payment success rate,agentic payments,unified commerce,payment technology,merchant operations Recommended organisation tags: Payments technology,Fintech,Artificial intelligence,Payment orchestration,Unified commerce,Merchant operations Hashtags: #PaymentsInfrastructure #PaymentIntelligence #ModelRisk #MerchantStrategy #FintechAI #PaymentOrchestration #AgenticCommerce
LinkedIn post 1
Schedule: Monday 28 September 2026 at 8:08am AEST Angle note for Sonny: Data-led strategic hook. This post explains the architectural shift and the competitive implications of owning the payment representation. It is not a merchant checklist. Relevant tags: @Stripe @Ant International @Razorpay @Bain & Company @AWS @NVIDIA
The payments industry is moving from task-specific models to a shared payments foundation model.
That changes the competitive question from:
“Which gateway has the best model?”
to:
“Whose payment representation am I buying?”
Razorpay’s Vulcan shows the architecture clearly. One backbone now supports four decisions across the purchase journey:
Routing.
Fraud detection.
Cash-on-delivery delivery risk.
Checkout payment-method ordering.
Stripe reports a 3.8% average acceptance lift through Authorization Boost and recovery of about 20% of false declines.
Ant International has put the scale of the race into focus with the Antom 3-in-1 Transformer,a payments foundation model running on more than 10 billion parameters and processing 90 terabytes of data annually.
The strategic frame is Bain & Company’s shift from “top of wallet” to “top of model”.
As AI agents increasingly influence what consumers buy and how they pay,the payment provider that owns the decision layer may control more value than the provider that simply moves the money.
For merchants,the diligence questions are now about training data,final outcome visibility,measurement,portability and concentration risk.
Read the full Quantum Payments analysis:
First comment: The most important distinction is between a model that predicts an outcome and a provider that controls the decision surface. Merchant-specific layers,selective tokenisation and access to final authorisation outcomes will become practical differentiators.
LinkedIn post 2
Schedule: Monday 28 September 2026 at 3:23pm AEST Angle note for Sonny: Operator and merchant-operational checklist. This post is designed for payments,risk and finance teams evaluating providers. It is deliberately practical and materially different from the morning’s strategic architecture post. Relevant tags: @Quantum Payments @Stripe @Ant International @Razorpay @Mastercard @Visa
Before signing for a payments foundation model,ask the provider these ten questions:
What data is the model trained on,and is our transaction data used to train shared models?
Can the provider see final authorisation outcomes,or only authentication results?
Are improvements measured with out-of-time testing rather than random splits?
Are fraud results compared at a fixed alert volume and reported by payment flow?
Do we get a merchant-specific decision layer on top of the shared backbone?
Which decisions are covered:routing,fraud,retries,authentication,checkout ordering,delivery risk and refund risk?
How are route-influenced declines separated from customer-side failures?
Is tokenisation applied selectively by issuer and payment pattern,or universally?
What enrichment can we provide,including margin data,product metadata and risk preferences?
What happens to our data,models and final outcomes if we leave,and how is agent-initiated traffic classified?
A foundation model can improve payment success rate and reduce fraud,but it can also increase single-vendor dependency.
The operational principle is simple:
A model is only as good as the outcomes it can learn from,and a provider should be able to explain how those outcomes are measured.
Read the full Quantum Payments analysis:
#MerchantOperations #PaymentsRisk #DataGovernance #PaymentArchitecture #AIProcurement #FintechStrategy
First comment: Add these questions to your next provider review alongside uptime,pricing and compliance. Model portability and data use on exit are not technical footnotes;they are commercial protections.