Cloudflare's new Monetization Gateway will let sites charge any agent per request in stablecoin over x402. The plumbing is real, the demand mostly isn't. The fight that matters is who controls access to what agents read.
For about six weeks now, Supercomputing News has been running on rails designed to sell its articles to machines. Not to the crawlers that have always taken our work for free, but to AI agents that pay for it, per request, in a stablecoin, over a payment protocol called x402. We went live on Base in mid-May, before Cloudflare generalized its year-old Pay Per Crawl experiment into the product it announced this week. I'm putting that up front because it's the reason I can write the rest of this honestly. We have skin in the game, and the game is more complicated than either its boosters or its skeptics will admit.
On July 1, Cloudflare opened the waitlist for its Monetization Gateway, which generalizes the company's earlier Pay Per Crawl feature from crawlers into per-request charging for any caller: an AI agent, an API client, an MCP tool asking for a web search. The mechanism settles at Cloudflare's edge, across more than 330 cities, in stablecoins, using x402. It's a waitlist, not a general release. Cloudflare has named no launch partners and disclosed no adoption numbers, and that restraint is worth keeping in mind, because the narrative around this technology has run well ahead of the evidence.
The category shift underneath the announcement is real regardless of how Cloudflare's product performs. Charging a machine, per request, to read something is no longer a protocol experiment run by early adopters; it is becoming a stock capability of the infrastructure that fronts much of the web. Whether the demand for that infrastructure exists yet is a separate question, and the honest answer is: mostly not.
Cloudflare is one mover in a landscape that's still forming. Coinbase built x402 itself and operates a hosted facilitator that verifies and settles payments, so a seller needs no blockchain infrastructure of its own. Google shipped AP2, the Agent Payments Protocol, with more than sixty partners including Mastercard, PayPal, and Coinbase, and with an extension that makes stablecoins a first-class way for an agent to pay. Akamai has partnered with TollBit and Skyfire under the blunt banner "No Free Crawls." That's the micropayment camp, which treats each machine request as a metered transaction settled in crypto at the edge.
A second camp solves an overlapping problem with a different economic model. TollBit installs a tollbooth that authenticates and meters AI traffic. ProRata, from Bill Gross, estimates how much a given AI answer relied on a particular article and pays a share accordingly. A coalition of publishers and technologists is pushing RSL, Really Simple Licensing, as an open standard, so that terms get set in the open before any single vendor locks them in. Per-request micropayment and per-use licensing are not the same thing, and conflating them obscures what's being negotiated. The value in this shift, as with a lot of the current AI stack, is accumulating in a middle layer, the metering and settlement and licensing rail, rather than in the content at one end or the model at the other. It's a pattern investors have started pricing aggressively elsewhere in the stack.
Here's where running the thing myself makes me a skeptic rather than an evangelist. The x402 token ecosystem carries a headline valuation in the six-to-seven-billion-dollar range, and the x402 Foundation, now housed at the Linux Foundation, counts Google, Visa, AWS, and Circle among its members. The commerce underneath is harder to find. In March, CoinDesk, citing analysis from Artemis, reported real on-chain volume of roughly twenty-eight thousand dollars a day, and read close to half of observed transactions as self-dealing, wash trading, or infrastructure testing rather than genuine economic activity; one February spike of 3.8 million transactions was largely testing. Transaction counts have climbed sharply since, roughly fivefold in June, and the same question follows the new volume: how much of it is demand, and how much is the ecosystem exercising itself. That gap between capitalization and cash flow will be familiar to anyone who has followed the wider argument that AI capital spending is outrunning AI revenue; the agent-payments corner of it is the same shape in miniature. Ron Shultz, a bill-pay veteran at ACI Worldwide, put the demand question about as sharply as anyone has: go find me ten people who want to use a stablecoin to pay a bill, he said, and explain to me why. I run this rail, and I don't have a good answer for him yet, other than SCN's rails aren't built for people - they're built for agents. The plumbing works, but the traffic is a trickle.
The distance between the loud narrative and the quiet ledger is the story. It's also why I think the more consequential question is who ends up controlling the terms when agent payments do scale, rather than whether they scale on the timeline the valuations imply.
That question lands hardest on the people who own technical and scientific knowledge. When agents become the primary readers of the scientific and engineering literature, and for some research workflows they arguably already are, access to that literature gets repriced. Roughly half of the world's research articles already sit behind paywalls whose licenses restrict automated access; for biomedical literature the paywalled, AI-unavailable share runs to about two-thirds. The licensing market that governs the rest is expanding fast, with the number of publishers active in AI licensing expected to nearly double this year, while the paywalled literature itself sits concentrated with a handful of large publishers, among them Elsevier, Springer Nature, and Wiley. Elsevier has been extending its LeapSpace platform with agentic capabilities across more than twenty million full-text articles and a thousand-plus licensing partners, which is a fairly literal description of the publisher-as-gatekeeper model taking shape.
This is where Supercomputing News covers a standing question we've committed to tracking: whether the AI buildout produces capacity that science and research can actually reach, or whether commercial AI captures the resource. We usually ask that about compute. Anthropic's move to lock up 3.5 gigawatts of Google TPU capacity through 2031 was, in our reading of it, commercial AI pre-purchasing the kind of infrastructure that scientific computing will eventually need. The knowledge layer is that same question wearing different clothes. If machine access to the scientific record gets metered behind stablecoin toll layers whose terms are set by commercial infrastructure firms, the resource being captured isn't compute. It's the literature itself.
An April 2026 report from the Open Markets Institute, covered by Nieman Lab the following month, frames the publisher's predicament as a double bind: the same large technology companies that are building commercial AI products and draining publisher traffic are the ones now dictating what the replacement revenue looks like. The report's phrase for it is that Big Tech is "occupying both sides of the value chain simultaneously." Publishers can neither afford to block the agents nor afford to give the content away, and the take rate is set by whoever controls the rail. One of Nieman Lab's own predictions for 2026 was blunter: publishers will see no meaningful AI licensing revenue. The terms, in other words, are being locked in during a period when almost no money is changing hands.
The rail furthest along for global machine access to knowledge is denominated in a United States dollar stablecoin, USDC on Base, and it settles on infrastructure operated by American firms: Cloudflare, Coinbase, Circle. A researcher in São Paulo or Jakarta or Nairobi whose agentic tools need to read the current literature would, under the arrangement now being built, pay for that access in a dollar-pegged instrument cleared through American rails, on terms shaped by companies headquartered in one jurisdiction. I'm not forecasting a winner here. The substrate is being poured, and the demand it's meant to carry has not yet arrived.
I don't know whether x402 becomes the rail or a footnote. Running it for six weeks has taught me mostly that the technology is further along than the market for it, and that the fight worth watching has little to do with transaction fees. The quieter one, over who sets the terms of access to the knowledge agents are increasingly the ones reading, will decide whether the people doing science get a seat at that table while it's still being built.
HTTP 402 has sat mostly unused in the specification for decades, reserved with the label "Payment Required" and almost never implemented. x402, the protocol Coinbase published in May 2025, finally gives it a job.
The exchange is short. A client requests a protected resource. The server answers with a 402 status and a PAYMENT-REQUIRED header describing what payment it wants and where to send it. The client constructs a signed payment payload and returns it in a PAYMENT-SIGNATURE header. A facilitator verifies the signature and settles the payment on-chain. The client then replays its original request with proof of payment attached, and the server returns the resource.
A few properties make this workable for machines. Settlement is sub-second. There's no account and no login, because the payment itself is the credential, so an agent that has never been seen before can transact on first contact. And the value transfer is small enough to make sense at web scale: it settles in USDC on Base using EIP-3009, an authorization mechanism that lets a payment be made on the client's behalf without the client holding an account with the server. Payment merges into the request path rather than sitting beside it as a separate checkout step.
None of which means the rail is hardened. Formal security analysis published this spring demonstrated five practical attacks on x402's authorization, replay, and binding protections, with outcomes ranging from unpaid service to paid-but-denied requests. Working plumbing and finished plumbing are different things.
