Cloudflare Is Building the Seller's Side of Agent Payments
A week ago I argued that most of the money AI agents spend won’t move over card rails. My example was a research agent paying a data service each time it pulls a record, many times a day, in small amounts, with no checkout and no cardholder. I said a wallet-to-wallet stablecoin transfer suits that spending better than a card, and then spent most of that piece on the payer’s problem: once agents are paying from wallets, someone needs to see what they spent, attributed to the right agent and totalled against the right budget.
What I skipped was the other end of my own example. For an agent to pay a data service per record, the data service has to be able to charge it. It needs a way to tell a piece of software what a request costs, take payment, confirm the payment cleared and let the request through, all without an account, a subscription or a card on file. I wrote as if that already existed. This week Cloudflare announced a tool for exactly that job. A closed beta from one company is a signal worth taking seriously, and not yet much more than that, so I want to go through what it covers and what it leaves open.
What Cloudflare announced
According to Fortune, Cloudflare announced on September 30 tools, not yet available to the broader public, that will let businesses charge AI agents in USDC, Circle’s dollar-pegged stablecoin. Fortune describes the initiative, called the Monetization Gateway, as letting an agent automatically pay a small fee each time it accesses an online service, instead of a human signing up for a monthly subscription or entering a credit card number. It is in closed beta for eligible US-based buyers and sellers.
Cloudflare will work as an intermediary between websites and the AI agents that visit them. When an agent requests information or a service, Cloudflare presents a price and grants access once the payment clears, automatically. Underneath, it uses Coinbase’s x402 system, which Fortune describes as a payment standard that lets websites charge software for access, and it settles USDC transactions on Base, Coinbase’s blockchain network. The x402 project’s own site is more concrete: if a request arrives without payment, the server responds with HTTP 402, the “Payment Required” status code, and the client pays and retries the request.
The launch builds on cloudflare.pay, which Cloudflare rolled out in August and which, per Fortune, lets AI agents prove they are acting on a user’s behalf and spend from a linked wallet. On timing, Cloudflare product manager Rohin Lohe told Fortune: “We’ll be ramping up features and users through the rest of the year, with plans to be widely available by early 2027.”
The article gives no figures on how many sellers are in the beta, what they charge, or how much money has moved through it.
Why the seller’s side was the missing piece
Most of the attention in agent payments has gone to the buyer. Fortune’s piece names Stripe, Tempo, MoonPay, Robinhood, Visa and Mastercard among the companies that have moved to let agents hold funds, shop or make payments. Those answer the question of how an agent pays. Far fewer answer how the thing being paid gets paid, and in agent-to-agent commerce I think that is the harder half, because the seller is the one who has to change how it makes money.
Think about how a data service or an API charges today. Usually a person signs up, picks a plan and puts a card on file, and the service bills monthly or by usage against that account. An agent that needs one record from a service it has never used before doesn’t fit that model. Signing up for a monthly plan to make a single request is poor economics for the buyer, and a seller with no way to charge for a single request can give it away, block it, or put it behind a signup flow the agent can’t complete. Card rails don’t solve this per request. A card payment carries fixed per-transaction costs and a full authorisation and dispute process, which is reasonable for a purchase and out of proportion for a very small charge made by software thousands of times.
A price returned with the response, paid in a stablecoin and checked before access is granted, is the shape that fits. The idea isn’t new, and x402 existed before this launch. What I think Cloudflare changes is distribution. Its business is already sitting between websites and the traffic that reaches them, which is exactly where a price would need to be quoted. If the product works the way Fortune describes, a seller whose traffic already runs through Cloudflare wouldn’t have to build payment handling into its own service to start charging agents. That is the difference between a standard some developers have implemented and an option for businesses that never planned to think about agent payments. I’m inferring this from what the product is, not from adoption data, which doesn’t exist yet.
For my argument, this moves “agents will pay each other in stablecoins” from a technical possibility towards a possible business model. A buyer with a funded wallet is only half a market until sellers can quote it a price and collect.
What it doesn’t show yet
It’s a closed beta, limited to the US, and the only timeline is a product manager’s statement about early 2027. I read dates in payments launch announcements as intentions. Nothing reported tells me that sellers will want to price per request, that agents will be given wallets funded to pay those prices, or that the volume will turn out to be meaningful. One infrastructure company shipping the mechanism is evidence that it expects the demand, not that the demand has arrived.
The cloudflare.pay detail made me adjust something I’d written. In the SoFi and Mastercard piece I described two paths for stablecoins: one underneath the card, and one where the card isn’t part of the transaction at all. Cloudflare’s design sits on the second path, with no card anywhere in it. But it doesn’t remove the person. The agent spends from a linked wallet and proves it is acting on a user’s behalf, so the person moves from the moment of payment to the setup before it: someone funds the wallet and authorises the agent, and after that the agent makes payments nobody approves one by one. I had been writing about agent-to-agent spending as if there were nobody on the buyer’s side. The more accurate version is that nobody is present at each payment, though somebody is accountable for all of them. If an agent overspends or pays for something useless, a person still answers for it, only later.
The reporting also says nothing about what happens when an agent pays and the service doesn’t deliver. On a card there is a process for that, however much merchants dislike it. For a very small per-request fee I doubt a formal dispute would be worth anyone’s time, and my guess is that the practical remedy will be the operator noticing and stopping payments to that service. That’s a guess, not something the launch describes. If it’s right, the buyer’s own record of what it paid, and for what, has to do the job a chargeback does on cards.
That brings the payer’s problem back in. Per-request pricing produces the kind of spend that is hardest to watch: individual payments too small for anyone to notice, adding up to totals someone should. A mechanism that makes those payments easy to make also makes the visibility problem I described last week bigger.
One standard, one stablecoin, one chain, for now
Cloudflare’s configuration is tidy: x402 for the request and payment, USDC for the money, Base for settlement. For a seller behind Cloudflare that’s one setup, but I’d be careful reading it as the shape of the whole market.
x402 itself isn’t tied to Base or to USDC. Its site describes the standard as blockchain-agnostic, supporting EVM-compatible chains, Solana and others, with stablecoin payments as the primary use case. Cloudflare’s choice of USDC on Base is a product decision made on top of a standard that allows many combinations, and the other companies on Fortune’s list aren’t building to one design either.
Let’s imagine a situation. An agent pays for thirty different services in a working day. Some sit behind Cloudflare and expect USDC on Base. Some implement x402 themselves and settle on a different chain or in a different stablecoin. Some are only reachable through a card-based agent programme. Whoever operates that agent has to hold balances where each seller wants them, choose the right way to pay each one, and afterwards reconcile all of it into one picture of what was spent. That is a routing and reconciliation problem, the one I spend my working days on in card payments, where no single provider covers everything a merchant needs. My expectation, and it’s no more than that, is that agent payments will fragment in a similar way across chains, stablecoins and providers, because that tends to happen whenever more than one company can build the rail.
Where I’d hedge is on depth. In card payments, each provider tends to bring its own integration, response codes and reporting, and much of the orchestration work sits there. If x402 becomes the common way sellers quote a price and accept payment, the request-and-pay step could look the same across sellers even when settlement differs underneath. That would push the orchestration problem for agents lower down, into which chain and stablecoin a balance sits in and how funds move between them, rather than integrating with every seller separately. That is a meaningfully easier problem than the one card payments have. I don’t know which way it will go, and one closed beta can’t tell me.
Where this leaves my view
I still think most agent-to-agent spending won’t run on cards, and this launch is consistent with that. It’s the first time I’ve seen a company with Cloudflare’s position in front of websites put out the seller’s side of it, which matters more to the argument than another wallet product would. It is also a beta, in one country, on one chain, with no numbers attached.
What I’d watch is whether it opens up on the schedule Cloudflare has given, whether it stays tied to USDC on Base as it widens, and whether sellers who have always billed by subscription decide per-request pricing is worth offering. The last one matters most to me. Cloudflare can make charging an agent easy, but whether a business wants to sell to agents by the request is a decision each seller makes on its own economics, and nothing in this announcement settles that.