Scroll to the content

Here is a merger nobody has proposed. Anthropic, the company that makes Claude, buys Cloudflare, the company that stands in front of roughly a fifth of the web and decides which requests get through. The combined firm is called Claudeflare. This is not negotiable, it is also not happening, and I intend to spend the next three thousand words on it anyway.

I come to this with credentials. Aardling runs on WordPress and has done for a long time. I moved a community forum to Discourse and lived to write about it. And at one point, for a client whose name I will protect, I installed a slider plugin, which is the kind of thing you do once and then think about at 3:17 in the morning for the rest of your life.

Here is how it would go.

The first in what I am optimistically calling a series on mergers nobody has proposed, examined with more care than they deserve.

Claudeflare, as the cloud sees it. Illustration: Claude, October 2026.
Contents
  1. Where the intelligence lives
  2. The internet gets a nervous system
  3. Your context, following you around
  4. The internet gets a wallet
  5. EmDash, which is one piece and not the whole puzzle
  6. Infrastructure that doesn’t page anyone
  7. The part where none of this happens
  8. What already exists
  9. In the end

Where the intelligence lives

The thing that makes this combination different from “Anthropic signs a partnership with a cloud provider” (which it has done, several times, with the companies you would expect) is geography.

When you talk to Claude today, your request leaves your device, crosses an ocean or two, arrives at a data centre, the model runs, and the answer comes back the same way. Depending on where you are, that round trip takes a few hundred milliseconds, or long enough that you start to wonder whether you have offended it. It is fast enough to be useful. It is not fast enough to be invisible.

Cloudflare has servers in more than three hundred cities. Put the model there, in all of them, and the request stops crossing oceans. It goes to the building down the road (or, in Antwerp’s case, probably to Brussels, which is still an improvement on Virginia). The number worth holding in your head is ten milliseconds, because that is roughly where human perception stops registering a delay as a delay. Below it, a tool stops feeling like a tool and starts feeling like a faculty. The pause goes away. You stop consulting the AI and start simply having it.

It is the difference between carrying a dictionary and knowing the words. Everything below is downstream of that one shift.

The internet gets a nervous system

The internet has been, since the day it was switched on, a remarkably stupid pipe. It moves bytes without any idea what they are. Routing decisions are made on the basis of destination addresses, the way a postman delivers a letter without opening it (a comparison that holds right up until you remember the postman also doesn’t know whether the envelope contains a birthday card or anthrax).

Claudeflare would change that, and here is the part I find hard to stop thinking about. Cloudflare already sees a fifth of the web’s traffic. Not a sample. Not a snapshot. The live thing, in flight, at every one of those three hundred cities at once. Put a model inside that and you have something that has never existed: an entity that is reading the internet in the present tense, while it loads.

A model that has read the internet is impressive. Claude has done that. So has everyone else’s model, which is why they all know the same jokes. But reading the internet is archaeology. You are looking at what the web was on the day someone froze it for training. A model that is reading the internet, now, continuously, is something else entirely. It would notice that a story is breaking before any newsroom has published it, because it would see the traffic shift. It would know a product had launched before the press release, a site had gone down before the status page, a scam was spreading before the first victim had finished typing their card number. It would be the first thing in history to hold an opinion about the whole web while the web was still happening.

Tim Berners-Lee spent the better part of two decades trying to get the web to be “semantic” through standards and metadata, a plan that required every publisher on earth to annotate their content correctly and consistently. This was a bet on human diligence, and history has declined to honour it. Claudeflare’s version of the semantic web requires nothing from publishers. The intelligence sits in the infrastructure, not in the content. The web becomes meaningful because the network understands it, not because everyone agreed to fill in the same form.

  1. The first thing you would do with a network like that is security, because Cloudflare already does security and this is just the better version. Today their edge blocks attacks by matching traffic against rules someone wrote after the last attack. That works well against the attacks that have already happened and poorly against the ones that haven’t, which is the category you were actually worried about. A network that reasons about traffic is the difference between a security guard checking names against a clipboard and a security guard who has been doing the job for twenty years and knows something is wrong before he can say what.
  2. The second thing you would do is provenance. A photograph that transits Claudeflare’s network could be signed on the way through: where it came from, what was done to it, and when. Not blocked. Not judged. Just labelled, the way a bottle of wine carries its vintage whether or not you intend to drink it. The infrastructure would not solve misinformation. It would make convincing misinformation considerably more expensive to produce, which is the only kind of solution anything has ever had.

And now the obvious objection, which I will raise myself so that nobody has to email me. Everything in the last five paragraphs, described from the other side of the glass, can be abused as a surveillance apparatus. A network that understands all its traffic is a network that could understand your traffic. What separates a nervous system from a wiretap is whether the thing reasons about the shape of the traffic or the contents of the envelope, and that is not a law of physics. It is an engineering decision, made in a meeting room, by people, probably on a Tuesday. I will come back to the Tuesday.

A network that reads itself. Illustration: Claude, October 2026.

Your context, following you around

In the Claudeflare world, Claude is not a website you log into. It is a persistent layer at the network edge, near wherever you happen to be. Your Cloudelet (a word I am coining here and will not be talked out of) holds the context you have chosen to share with it:

  • your professional background for work tools,
  • your preferences for the shops you use,
  • your communication style for anything that involves writing, t
  • he fact that you have already told four separate government portals your address this year and would prefer not to make it five.

This context travels with you. As you move, edge nodes hand it off the way phone masts hand off a call. It is always local, always fast, and always yours, in the specific sense that you decide what it discloses, to whom, and for what.

The Cloudelet, on its lead. Illustration: Claude, October 2026.

The immediate consequence is that you stop introducing yourself to every system you meet.

  • You do not re-enter your details at every clinic.
  • You do not re-upload a CV to a portal that will immediately ask you to retype the CV into fourteen separate fields.
  • You do not explain to the bank, the mutuality, the municipality and the energy supplier, in four slightly different vocabularies, that you have moved house.

Your Cloudelet speaks for you, with your permission, with exactly the information that interaction needs and nothing beyond it.

Note what the Cloudelet does not do. It does not break down the silos. The clinic still cannot see the bank, the bank still cannot see the municipality, and that is a legal and political problem that no amount of infrastructure solves on its own. What it does is make you the integration point. The systems no longer need to talk to each other if they can each talk to the thing that travels with you. That is a smaller promise than “one government database to rule them all”, and I trust it more for being smaller.

The deeper consequence is that the assistance becomes cumulative. Today every conversation with Claude starts from nothing. It is not that the model is poor; it is that the architecture has no memory, like a brilliant colleague who is concussed each morning. At the edge, with persistent context, the thing gets more useful the longer you use it. It learns that you write in British spelling, that you want the short answer first, that you have asked about the same tax question three Octobers running and would like it answered before the fourth. That is how every other relationship in your life works, and how almost nothing in software does.

It is also, not by accident, the half of the merger that makes the other half pay. A Cloudelet that knows what you value is the only kind of agent that can be trusted with a wallet.

The internet gets a wallet

Now the part I think matters most, and it starts with a number that has been sitting in a specification document, unused, since 1997.

HTTP has a status code 402. That’s nerdspeak for “Payment Required”. It was written into the HTTP/1.1 standard alongside the ones you know (200 OK, 404 Not Found, 500 Something Is On Fire) and then marked “reserved for future use”. It has stayed reserved for twenty-nine years. Entire companies have been founded, listed and liquidated in the time it has taken the web to decide what to do with a status code it already had.

How an AI agent pays for a web page under x402Four HTTP exchanges. No login, no subscription, no cookie banner (probably).Your Cloudelet(AI agent, reading on your behalf)Publisher’s site(EmDash, at the edge)1. GET /the-article-you-wanted2. 402 Payment Required (price: $0.0018)3. GET again, with a signed payment header attached4. 200 OK (the article, and the writer gets paid)Status code 402 has existed since HTTP/1.1 in 1997. It has been labelled “reserved for future use” for 29 years.
Four exchanges, one status code. Diagram: Claude, October 2026.

Here is what the future use turned out to be.

  1. A client (a browser, an app, an AI agent) asks for a page.
  2. The server says 402, and names a price.
  3. The client pays, in real time, and asks again with the payment attached.
  4. The server hands over the page.

Four exchanges. No login, no subscription, no “create an account to continue reading”, and no cookie banner, although I have learned not to promise that last one on the European side of the Atlantic.

This is the x402 standard. Cloudflare has been circling it for a while: in July 2025 it announced “pay per crawl”, which uses that same x402 code to let websites charge AI crawlers for access, and this April it shipped EmDash, its new CMS, with x402 support built in. Set a price, add a wallet address, done. The announcement calls it “a built-in business model for the AI era”, which is marketing copy, but it is marketing copy for something real.

The reason this matters is that the web’s existing business model has quietly stopped working, and most people haven’t noticed yet because the invoices still arrive. Advertising was always a compromise. What publishers wanted was readers who valued their work enough to pay for it. What they got, because nobody could make a payment of one cent cost less than one cent to process, was readers whose attention could be sold to someone else. We built an entire economy on the workaround.

The workaround assumed a human was looking at the page. That assumption is now false at an alarmingly fast growing share of the time. Your AI assistant goes to a site, extracts what it needs, and gives you a summary. It does not see the banner ad. It does not click the affiliate link. It does not scroll past the “you may also like” box with the eleven articles about ear wax. The publisher did the work, the writer did the work, the content was consumed, and the money, at the exact moment it was supposed to change hands, did not.

x402 is the fix, and it is a surprisingly polite one. The agent asks. The site says: that’ll be $0.0018. The agent, which is spending your money and has been told how much of it you are willing to spend on this sort of thing, pays. The writer gets paid. Nobody fills in a form. At Cloudflare’s scale, with Claude as the agent for a few hundred million people, this stops being a feature of a CMS and becomes a settlement layer for the information economy. The thing advertising was pretending to be, done directly.

Caveeat: it does not solve everything. There will be content that should be free, and content that should never have existed. (I’m looking at you, AI content spinners.) There is also the question of what anything should cost remains as hard as it has been since the first person tried to sell a second person a fish. But the plumbing exists. And for twenty-nine years the pipe was there and nobody had connected it to a tap.

EmDash, which is one piece and not the whole puzzle

I should say a word about EmDash, since Cloudflare announced it on that fateful April 1st and a few people have asked whether I was sure.

I am sure. It is a new content management system, written in TypeScript, running on Astro, serverless, MIT licensed, and pitched (with the confidence of engineers who have never had to explain WordPress to a client) as the spiritual successor to WordPress. The spiritual successor. Not a fork. A successor. As if WordPress had died and this was the anointed chosen one.

WordPress has not died. It runs something like 43 percent of the web. But WordPress also has a plugin problem that has resisted every non-architectural fix attempted in twenty-three years: by Patchstack’s count, 96 percent of WordPress security vulnerabilities come from plugins. A WordPress plugin is a PHP script that runs inside WordPress itself, with the same access to the database and the file system. Installing a WordPress plugin is, in a completely literal sense, handing a stranger the keys to your server and asking them to be nice. And this week we’ve all seen what happens when something PHP-based turns out to be rather unsafe.

EmDash makes that structurally impossible instead of culturally discouraged. Every plugin runs in its own sandbox and declares, in a manifest, exactly what it is allowed to do. A plugin that sends an email must say so. A plugin that talks to an outside server must name the server. If it doesn’t declare it, it cannot do it, and you can read the list before you install anything, which is more than WordPress.org manages with a review queue that was over eight hundred plugins deep when Cloudflare wrote the post.

In the Claudeflare scenario the model audits those manifests at the edge. A plugin that calls itself a contact form and asks for network access to seventeen hostnames gets flagged in milliseconds, by something that has read the code, rather than in two weeks, by a volunteer who has not.

And EmDash sites scale to zero. When nobody is reading your site, it is not running, and it costs you nothing, which is the first time in the history of web hosting that the bill has agreed with reality. The teenager in Medellín with the music blog gets the same network as a national broadcaster. Not a cheaper version. The same thing.

So that is EmDash: a good CMS that happens to have the wallet built in. It is version 0.1.0. It is one piece. The piece is interesting because of where it sits.

Infrastructure that doesn’t page anyone

A less glamorous consequence. Running a network like Cloudflare’s is a continuous act of human judgement: capacity planning, route changes, rule updates, the 3 a.m. call when a BGP route in Marseille falls over and takes Portugal with it.

With a model in the network, a larger share of that judgement happens without a pager. The network notices the failed route, reasons about the alternatives (current load, latency, what kind of traffic is being moved), and picks one, in milliseconds, without waking anyone. It can route traffic toward whichever edge nodes are currently running on wind rather than gas, because carbon intensity is just one more variable once something is there to weigh it.

None of this is magic, and automated decisions at the infrastructure layer can propagate a mistake faster than a human could manage with both hands. There are good reasons critical infrastructure is not fully automated, and “the automation is now AI” is not one of them going away. The point is not that the engineers vanish. It is that they stop spending their nights on problems a model can handle and start spending their days on the ones it can’t.

The part where none of this happens

Thought experiments like this owe the counterargument a hearing. So here is my own counterargument.

Cloudflare’s value is partly its neutrality. It serves everyone, including every one of Anthropic’s competitors. Whether Google would keep routing traffic through an Anthropic-owned edge is not a rhetorical question; it is a strategic one with an obvious answer and a long tail of lawyers.

Regulators in Washington and Brussels would have views. Combining a dominant piece of internet infrastructure with a frontier AI lab would attract the kind of scrutiny that makes Microsoft’s purchase of Activision, which took twenty-one months and three jurisdictions, look like a planning application for a garden shed. They would want to know who controls the pipes and the thing that thinks about what is in them, and they would be right to ask.

And there is culture, which is what sinks most technology mergers while everyone is busy arguing about synergies. Anthropic is a research organisation with a deliberate relationship to risk. Cloudflare ships things on a Tuesday and blogs about it on the Wednesday. These are not naturally adjacent temperaments. The middle managers of the merged entity would earn whatever they were paid.

REGULATOR
The regulator, in its natural habitat. Illustration: Claude, October 2026.

What already exists

The odd thing about writing this out is how little of it is fiction.

The merger is the only part I made up, and a deep partnership (Claude as a first-class model on Cloudflare’s edge, EmDash as the house CMS) would deliver most of the value without the lawyers. So the question was never really whether Claudeflare happens. The question is about the choices some people make on any given Tuesday.

Because the same stack that could give the web a nervous system and a wallet could also make surveillance cheaper, lock-in stickier, and misinformation easier to industrialise. Nothing in the technology prefers one outcome.

  • EmDash being MIT licensed instead of something tighter was a choice.
  • x402 being an open standard instead of a proprietary payment rail was a choice.
  • Cloudflare offering a free tier that is actually free was a choice.
  • Anthropic spending money on safety research that does not ship as a feature was a choice.

Each of those choices was made in a room, by people, and each could have gone the other way without anyone outside the room noticing for years.

In the end

I am a writer with a website. I have watched the business model that was supposed to pay for writing erode for twenty years, one ad-blocker and one summarising assistant at a time, and I have made my peace with the idea that the fix, when it came, would be ugly. The fix turns out to be a status code from 1997 and a line in a config file. I did not expect to find that moving. I do.

Status code 402 has been reserved for future use for twenty-nine years. I will be watching my server logs for the first one. I expect it to be my own agent, reading my own post, and I expect to be charged $0.0018 for the privilege.

I’d pay it.

Gepubliceerd door Stijn Vogels

Sinds 2003 deelt Stijn Vogels zijn mening over mondiale merkwaardigheden. Gewapend met een geschiedenisdiploma en een scherpe pen zoekt hij naar zin in de chaos. Misschien vind je iets wat je aan het denken zet. Misschien ook niet. Maar dat is het risico van nieuwsgierigheid.