Author: Emily Sands
Translated by: Peggy, BlockBeats
This article stems from Stripe's redefinition of the "AI economy infrastructure": In the past, payment infrastructure primarily served human users and software companies; now, it must also serve machine agents. Agents need to be able to read prices, complete payments, call wallets, and deploy services; AI companies also need to build new billing, risk control, and settlement mechanisms around token consumption.
The Machine Payments Protocol, Link, Stripe Projects, Metronome, Tempo, and streaming payments mentioned in the article essentially all point to the same trend: AI Agents are evolving from "tools" into new types of economic participants on the internet. They are both buyers and builders; they both create software and consume resources. Consequently, the traditional SaaS-era business model of per-seat pricing, post-service invoicing, and manual settlement is being replaced by new models featuring real-time metering, real-time settlement, and machine-readable interfaces.
This is not just Stripe's product narrative; it's also a broader industry judgment: if a significant portion of future software consumption is initiated by Agents, then payment, billing, wallets, risk control, and deployment processes all need to be redesigned. AI is changing not just production efficiency, but also the very organization of business itself.
The original text follows:
Agents are becoming buyers and builders. This will fundamentally change how business operates.
For the past few years, Stripe has been building economic infrastructure for AI. This happens on a few dimensions: We help the world's fastest-growing AI companies accelerate through payments, billing, checkout, fraud and risk prevention, and tax infrastructure; we deploy AI across our own payments flow to help businesses earn more; and we make Stripe natively embeddable in the AI tools developers use.
But in the last six months, the meaning of "economic infrastructure for AI" has shifted. Agents are emerging as a new kind of actor on the internet. And that has given us two new tasks: helping agents purchase and build on behalf of people and companies; and helping companies adapt to the new economies that token consumption enables. In other words, Stripe is empowering agents to act independently, and empowering companies to monetize token-powered products that agents create and consume.
Agents as buyers
Ecommerce was built for humans. Humans browse websites, click on pricing pages, enter card details, and complete checkout. Agents don't. They need a programmatic way to understand that a service costs money, how much it costs, and how to pay for it without a human clicking through a checkout.
That's why we created the Machine Payments Protocol in partnership with Tempo. It lets businesses programmatically accept payments from agents, without human intervention. There's no account creation, no checkout page, and no human-in-the-loop; just a machine-readable way for agents to buy services from businesses.
Once businesses can accept payments from agents, the next equally important question is: How can consumers safely authorize agents to spend money on their behalf? That's where our consumer wallet Link comes in.
More than 250 million people already use Link. And we're making Link work for agents, so that people can authorize agents to safely spend on their behalf. Humans remain in control, and agents gain the ability to spend on their users' behalf.
Helping agents buy things on the internet is one way to empower them. The other is helping them build.
Agents as builders
Today, vibe-coding is easy, but vibe-deploying isn't. Agents can generate an application in minutes, but deploying it to the internet still requires a lot of manual glue: creating accounts, provisioning services, managing credentials, connecting APIs, and switching back and forth between different consoles.
Stripe Projects lets developers—and their agents—register for, manage, and connect to the services needed to deploy an application, directly from the command line. The goal is simple: make vibe-deployment as easy as vibe-coding.
Token monetization
As agents begin to act as buyers and builders, businesses also need to adapt to a world where the cost of inference—the token cost of the product—changes dynamically. And that's changing software's economics.
In SaaS, serving one more user often costs almost nothing extra. But in AI, every prompt, every API call, every agent task has a real marginal cost. So companies can't just charge per seat. They need to meter usage in real time and charge customers for the resources they actually consume.
That's why usage-based billing is at the heart of AI monetization. Businesses can charge for what their customers actually value: usage, workflows, outcomes, or any billing unit that best fits their product.
Lovable is a great example. It started with simple subscriptions on Stripe so that it could commercialize quickly after launching. As the company evolved, so did its billing model. Today, Lovable charges for AI tokens consumed once customers exceed the usage included in their subscription plan. ElevenLabs has followed a similar path: it started with subscriptions on Stripe and later, as its product and its customers' usage evolved, added pay-as-you-go pricing.
Token theft
But usage-based billing only works if businesses can actually collect. And today, fraudsters aren't just stealing money or account credentials; increasingly, they're stealing tokens. They can sign up for an account, quickly consume a large volume of tokens, and disappear before the bill comes due, which collapses AI products' economics.
Token theft is one of the least-discussed problems in AI today. Stripe Radar, our fraud prevention product, evaluates new accounts in real time, predicts which free trials are likely to be abused, and identifies non-payment risk as usage accumulates.
Streaming payments
Abuse gets even more difficult when the customer itself is an agent. An agent can consume tokens at machine speed, which makes traditional billing models riskier—and sometimes unworkable.
Businesses can ask for upfront payment and cut off service when the prepaid amount runs out. This protects the business, but creates a worse experience for customers, and prevents good customers from spending more. Businesses can also let usage accumulate, then send an invoice later. That's better for customers, but then the bill may never get paid—and the tokens are already gone.
The better answer is obvious, but technically complex for businesses to build themselves: track usage and collect payment in real time. That's what Metronome and Tempo can do, together.
Metronome is a company we acquired that serves the most complex usage-based businesses; it can track usage in real time, as tokens are consumed. Tempo enables low-cost, high-frequency stablecoin payments that settle instantly. Together, AI companies can charge for tokens as they're used—without having to choose between hard limits and uncollectable bills.
We call this streaming payments: a new business model built for AI-native businesses, and for machine-speed software consumption.
This is what "economic infrastructure for AI" needs to be today.
It's no longer just payments for AI companies. It's commerce for agents, wallets for agents, deployment for agents, billing for tokens, fraud and abuse for tokens, and streaming payments for agents.
AI is changing business, and it's changing how businesses are built. So the infrastructure has to change, too.
We're building it at Stripe.
