Inside the Race to Build the AI Payments Stack

The race to build the infrastructure for AI agents that can autonomously pay for services, data, and goods is already underway. Instead of a single standard, several major technology and financial companies are building different pieces of what is becoming the AI payments stack.

Coinbase, Stripe, Google, and Visa are approaching the problem from distinct angles. Coinbase focuses on web-native crypto payments, Stripe is creating machine-payment coordination protocols and settlement rails, Google is defining how agents prove authorization to spend money, and Visa is building merchant acceptance infrastructure so businesses can safely accept agent‑initiated purchases. Together, these efforts reveal how AI commerce may evolve—and why the future will likely rely on a mix of crypto rails and traditional payment networks.

The Four Competing Approaches to AI Agent Payments

Coinbase: x402 and Payments Built Into HTTP

Coinbase’s x402 protocol embeds payments directly into the core mechanics of the web. It revives the rarely used HTTP status code 402 “Payment Required” so servers can demand payment as part of a standard web request.

The flow works like this:

• A client or AI agent requests a protected API endpoint or resource.
• The server responds with an HTTP 402 Payment Required message and includes structured payment requirements.
• The client generates a signed payment using its wallet (typically with stablecoins).
• The request is retried with proof of payment.
• Once verified, the server returns the requested resource.

Because the protocol runs directly on HTTP, it allows programmatic payments without accounts, sessions, or manual checkout flows. This design makes x402 particularly suited to machine‑to‑machine commerce such as API metering, data access, and autonomous software services.

Settlement typically occurs using stablecoins on blockchain networks, enabling low‑value transactions and automated micropayments.

Stripe and Tempo: Machine Payments Protocol

Stripe, working with Paradigm-backed startup Tempo, introduced the Machine Payments Protocol (MPP) to standardize how machines coordinate and execute payments autonomously.

MPP acts as a programmatic payment coordination layer. According to Stripe, the protocol enables agents and services to request and settle payments for microtransactions, recurring billing, and escrow arrangements.

In many deployments today, the protocol runs on Tempo, a purpose‑built blockchain designed for high‑speed stablecoin transactions. The network launched alongside the protocol to support machine‑to‑machine payments at large scale.

The basic payment interaction resembles web service calls:

• An agent requests a service or resource.
• The service responds with payment requirements.
• The agent authorizes the payment programmatically.
• The resource is delivered after payment confirmation.

Tempo’s architecture targets high throughput and rapid settlement suitable for automated commerce between software agents.

In practice, Stripe appears to be positioning MPP as a protocol layer that can route payments across multiple rails, including stablecoins and traditional payment methods integrated into Stripe’s ecosystem.

Google: Agent Payments Protocol (AP2)

Google’s Agent Payments Protocol (AP2) tackles a different problem: proving that an AI agent actually has permission to spend money.

The protocol is designed as an extension to Google’s Agent2Agent (A2A) and Model Context Protocol (MCP) frameworks, which enable agents to communicate with each other and external tools.

Instead of acting as a payment rail, AP2 functions as a secure authorization and interoperability framework.

Its core idea is the use of cryptographically verifiable “mandates.” These digital credentials represent signed instructions from a user authorizing an agent to perform specific purchases.

For example, a user might authorize an agent to:

• Book flights under a certain price
• Buy a product within a defined budget
• Pay for API usage or compute resources

Merchants and payment providers can verify these mandates to confirm that the agent’s transaction reflects the user’s intent. The system also provides an auditable trail for accountability in case of errors or fraud.

Later updates to AP2 added support for autonomous “human not present” transactions, allowing agents to execute payments based on pre‑authorized instructions without requiring a user at checkout.

Visa: Intelligent Commerce Connect

Visa is addressing the other end of the transaction: merchant acceptance.

Its platform, Intelligent Commerce Connect, provides a single integration that allows businesses to accept payments initiated by AI agents. The system handles authentication, credential tokenization, spend controls, and payment initiation.

Instead of introducing a new payment rail, Visa positions the platform as a network‑agnostic hub that works with multiple payment systems and agent protocols.

Through the Visa Acceptance Platform, merchants can:

• Accept agent‑initiated payments across card networks
• Authenticate AI agents acting on behalf of customers
• Apply tokenized credentials and fraud protections
• Enforce spending limits and permissions

The platform can integrate with several agent commerce protocols, enabling merchants to participate in AI‑driven shopping without redesigning checkout systems from scratch.

Which Layer of the AI Payments Stack Each Company Controls

Taken together, these efforts reveal that the competition is not about a single protocol dominating the entire market. Each company is targeting a different control point in the AI payments stack.

Coinbase:

• Web‑native payments for APIs and digital services
• Stablecoin settlement and wallet infrastructure
• Micropayments embedded directly in HTTP requests

Stripe / Tempo:

• Protocol for coordinating machine payments
• A specialized settlement network optimized for agent commerce
• Integration with both crypto and traditional payment methods

Google:

• Agent authorization and identity
• Interoperability between agents, merchants, and payment providers
• Verifiable proof of user intent

Visa:

• Merchant acceptance and payment credential orchestration
• Fraud protection, tokenization, and spend controls
• Integration with existing global card infrastructure

Each company is effectively trying to become indispensable in one layer of AI‑driven commerce.

What This Competition Means for the Future of AI Commerce

The emergence of multiple protocols suggests that AI payments will not converge around a single system. Instead, the ecosystem is forming a layered architecture similar to the early internet.

Different payment rails will likely dominate different types of transactions.

Crypto‑native rails—such as those used by x402 or Tempo—are well suited for:

• machine‑to‑machine micropayments
• API usage metering
• cross‑border payments
• autonomous software services

Meanwhile, traditional payment networks will likely remain dominant for:

• consumer retail purchases
• regulated credential issuance
• chargebacks and dispute handling
• large merchant ecosystems

Visa’s platform already reflects this hybrid model by enabling agent‑initiated payments across multiple networks and protocols through a single integration.

The Most Likely Architecture: Hybrid Payments for AI Agents

The emerging architecture of AI commerce may look something like this:

• Authorization layer: protocols like Google’s AP2 proving user intent and agent authority
• Payment coordination layer: systems like Stripe’s MPP orchestrating transactions
• Settlement rails: stablecoin networks, blockchains, or card payment systems
• Acceptance layer: platforms like Visa’s Intelligent Commerce Connect enabling merchants to accept agent payments

In this model, an AI agent could dynamically choose the most efficient rail—crypto for machine‑to‑machine payments, or card networks for traditional retail purchases.

Rather than replacing existing infrastructure, the new protocols are gradually connecting autonomous AI agents to the global financial system. The result is not one winner, but a stack of interoperable technologies that together enable machines to participate in commerce.