Scroll to the content

Here is the second merger nobody has proposed. The Wikimedia Foundation, which runs Wikipedia, buys the Internet Archive, which runs the Wayback Machine. Neither of them is for sale. Neither of them has shareholders. One of them is a charity with an endowment and the other is a charity in a former church, so “buys” is doing a lot of work in that sentence, and I intend to let it.

I come to this with credentials. I have a history degree, which means I have spent years citing documents that other people kept for me. I have run Aardling since the first George W. Bush administration was still a plausible topic for a blog, which means roughly a third of the links in my own archive now go nowhere. And I once spent an evening in the Wayback Machine reading a version of my own site from 2006, which is the kind of thing you do once and then decline to discuss.

Here is how it would go.

The second in a series on mergers nobody has proposed, examined with more care than they deserve.
The first was Claudeflare, in which Anthropic bought Cloudflare.

Contents
  1. Who is buying whom, exactly
  2. The footnotes stop rotting
  3. A provenance layer for the age of answers
  4. The library gets a turnstile
  5. Two legal departments, one courtroom
  6. What a library does that a lab does not
  7. The part where it doesn’t happen
  8. The merger that already happened

Who is buying whom, exactly

Let me get the money out of the way, because the money is the joke.

The Wikimedia Foundation reported revenue of about 208 million dollars in 2025, sits on an endowment of about 169 million, and employs around 650 people. The Internet Archive reported revenue of about 23.7 million dollars for 2025 against 27.8 million in expenses, and finished the year with net assets of 2.8 million. One of these organisations could buy the other with the interest on its savings account. That is not a criticism of either of them. It is just the ratio, and the ratio is what makes the thought experiment work.

The two already behave like a couple who have never discussed it. Wikipedia is the world’s most-cited website that cites other websites, and the other websites keep dying. The Internet Archive is the only organisation that keeps copies of the dying websites at a scale that matters. A bot called InternetArchiveBot has been running between the two for years, and the Turn All References Blue project has repaired more than 30 million broken citations across hundreds of wikis by swapping dead links for Wayback snapshots. Nobody signed a contract for that. It is a volunteer arrangement between two charities, which is how most of the internet’s load-bearing infrastructure turns out to be held together when you look closely.

So the merger is less an acquisition than a formalisation. The encyclopedia buys its own footnotes.

The footnotes stop rotting

Start with the problem that every history teacher, every journalist and every person who has clicked a reference link on Wikipedia already knows. According to Pew, 38 percent of the web pages that existed in 2013 were gone a decade later. Pew’s sample found about a quarter of 5.4 million URLs dead on the live web, and the Wayback Machine had copies of about 72 percent of those. Brewster Kahle likes to cite the early-web figure that an average page lasts between 40 and 100 days, which is shorter than most houseplants.

A footnote is a promise. It says: do not take my word for it, go and look. On paper the promise held for centuries because paper is lazy and does not reorganise itself. On the web the promise has a half-life. Wikipedia is, at the moment, the largest collection of promises on earth, a great many of which point at a 404.

In the merged organisation, which I will call Wikiback for the length of this post and not a minute longer, the footnote changes. Every citation added to any Wikipedia article is archived at the moment of citation, and the snapshot is the citation. The live page becomes a courtesy. The archived copy is the record. You click the little number, you get the page as the editor saw it on the day they used it, with a second link to whatever the page has become since, if it still exists. Historians have a word for this. The word is “source”. The rest of the internet has been working without one.

This is mostly already possible. InternetArchiveBot does a version of it after the fact. The difference under one roof is that it stops being a repair job and becomes the default, which is the difference between a smoke detector and a fire brigade that arrives in 2031.

A provenance layer for the age of answers

Now the part where it gets large.

The reason anyone at Wikimedia would be thinking about this at all is that human visits to Wikipedia fell about eight percent year on year in 2025, after the foundation updated its bot detection and discovered that a chunk of what had looked like a spring surge was scrapers built to pass as people. Marshall Miller of the foundation put the cause plainly: search engines answer the question themselves now, and younger users go to video. The encyclopedia is being read more than ever and visited less than ever, which is the kind of sentence that used to describe the Bible.

The Internet Archive has the mirror-image problem. It has passed one trillion archived web pages, a number it celebrated in October 2025 with a party in the Funston Avenue church, and the thing a trillion pages is most useful for in 2026 is training a model that will then answer questions without mentioning where it read them.

Put the two together and you have the one thing the answer engines do not have and cannot build on their own: a public, non-commercial, timestamped record of what was written, by whom, when, and what it said at the time. Call it the provenance layer: the part of the internet that remembers, run by people who have no product to sell you.

Here is what that looks like in practice. A model says that a Belgian minister resigned on a Tuesday. You ask it how it knows. Today you get a confident paragraph and, if you are lucky, a link that may or may not still resolve. In the Wikiback world the model has been trained on, or is reading from, a corpus where every claim carries a pointer to a snapshot that cannot be edited after the fact, and the snapshot carries a pointer to the Wikipedia talk page where three people argued about whether the resignation counted. The argument is part of the record. The argument is, in fact, the most valuable part of the record, because it is the only part that a model cannot generate for you with a straight face.

This is not a new idea. It is the idea of a library, with a timestamp. The novelty is only that two organisations that each hold half of it would finally be the same organisation.

The library gets a turnstile

In January 2026, on Wikipedia’s 25th birthday, the foundation announced that Microsoft, Perplexity and Mistral had joined Meta and Amazon as paying customers of Wikimedia Enterprise, the commercial API that serves the encyclopedia in formats a company can plug into a product. Lane Becker, the foundation’s senior director of revenue, described it as a version of Wikipedia tailored for commercial use. Everyone else described it as Wikipedia finally sending AI companies an invoice.

I wrote about pay-per-crawl in the Claudeflare piece, where the question was whether Cloudflare could stand at the door of a fifth of the web and charge admission. The question here is quieter and, I think, more interesting. What happens when the two largest free things on the internet agree on a single price list?

The merged library charges the machines and not the people, and it charges them for the one thing the machines cannot make themselves: being able to prove it. A model company can scrape Wikipedia for nothing, as it always could. It can scrape the Wayback Machine, as it always could, and the Archive’s budget suggests it has been getting a bargain. What it cannot do without the licence is ship an answer with a verified pointer to the exact revision and the exact snapshot. The free tier is the text. The paid tier is the receipt.

That is a better business than the Enterprise API alone, because the receipt is what the model companies are being sued for not having. It is also a business that the Internet Archive, on 23 million dollars a year, could never build, and that Wikimedia, without the Archive, has nothing to sell.

The Internet Archive’s recent history is a list of court cases, and this is where the merger stops being charming.

In June 2020 four publishers, led by Hachette, sued the Archive over its lending of scanned books. The district court ruled against it in March 2023, and the Second Circuit affirmed that ruling on 4 September 2024, holding that scanning a book you own and lending the scan one copy at a time was not fair use. The Archive did not take it to the Supreme Court. More than 500,000 books came out of its lending library. Then the record labels came for the 78s. Universal, Sony and Concord asked for 621 million dollars over digitised shellac records, a sum roughly 26 times the Archive’s annual income, and the case ended in a confidential settlement in September 2025. In between, in October 2024, the site was knocked offline and 31 million accounts were breached.

Wikimedia’s recent history, by contrast, is a list of fundraising banners and one argument with a US attorney about tax status. Its lawyers mostly send letters. The Archive’s lawyers mostly receive them.

So the honest version of the merger is not that the big charity buys the small one. It is that the organisation with the money takes on the organisation with the lawsuits, and gets a trillion pages and a federal depository designation in the bargain. (Senator Alex Padilla designated the Archive a federal depository library in July 2025, which means a former church in San Francisco is now an official home for US government documents, a sentence I have read four times and still enjoy.)

Whether a foundation that has spent two decades being carefully boring about copyright would want that bargain is the question. Wikipedia’s whole posture is that it is the least litigable website on earth. The Archive’s posture is that someone has to find out where the line is, and it may as well be them. Merging those two postures produces either a library with teeth or an encyclopedia with a target on it. I have been to enough family dinners to know which one you get first.

What a library does that a lab does not

There is a lazier version of this piece in which I say the AI labs are the villains and the two charities are the heroes, and I would like to avoid it, because the labs are in fact the Archive’s best customers and the encyclopedia’s most reliable readers. The difference is not virtue. The difference is what each kind of institution is for.

A lab is for the next thing. It trains a model, ships it, and trains the next one on what the first one learned. It has no reason to keep the old corpus around once the weights are baked, and every commercial reason not to say exactly what was in it. A library is for the old thing, forever, with the receipt attached, and it has no idea what you will want it for. That is not a design flaw. That is the job description, and it is why libraries are older than every company currently training a model.

Here is the part that interests me most. The labs will keep getting better at answering. The merged library gets better at something else: it becomes the place where the answers can be checked, by anyone, without a login, with a timestamp, in the same way a court record or a parish register can be checked. In a world where text is cheap and confident, the expensive thing is the record of who said what first. We are going to need a lot of that, and I would rather it belonged to two charities than to a fifth company with a rocket.

The part where it doesn’t happen

Now the caveats, which I include so that the people who email me about the Claudeflare piece can save themselves the trouble.

  • First, there is nobody to sell. The Archive is a nonprofit run by its founder. Kahle has spent thirty years turning down the offers that would have made this question moot, and a man who keeps a trillion pages in a church is not going to hand them to San Francisco’s other charity because a blogger in Antwerp drew a diagram.
  • Second, the Wikimedia community would revolt, and the Wikimedia community revolts over the font on the donation banner. A foundation that takes on half a billion dollars of publishers’ grievances is a foundation whose volunteers will immediately ask whether their edits are now evidence. They would not be wrong to ask.
  • Third, the two organisations have different theories of truth. Wikipedia believes a thing is true if a reliable source says so and nobody has reverted it. The Archive believes a thing existed if it was online for long enough to be captured. Those are compatible, but they are not the same, and an organisation that holds both has to decide, every day, which one wins when a source turns out to have been lying at the time of capture. Wikipedia’s answer is a talk page. The Archive’s answer is to keep the lie, timestamped. The merged answer is both, which is correct and exhausting.
  • Fourth, the money. Wikimedia’s endowment could absorb the Archive’s budget, but the Archive’s storage grows faster than the endowment does, and a trillion pages is a number that only ever goes up, which is the exact opposite of a donation. A foundation that currently spends its money on servers and salaries would be adding a petabyte habit it cannot quit without becoming the thing it was set up to prevent.
  • And fifth: Wikipedia just got a competitor that does not care about any of this. Grokipedia launched in October 2025 with the explicit pitch that the encyclopedia is biased and a model can do it better. Personally, I don’t think a model can do it better. But I also do not think the response to “a machine will write the encyclopedia” is “two charities should merge”. The response is to make the citations so good that the machine has to use them. Which, I notice, is what this whole post has been about.

The merger that already happened

I keep a folder of old drafts on this site. There are 285 of them, going back to 2005, and when I open one I usually find that the link I was writing about is dead, the article I was quoting has moved, and the only reason I can reconstruct what I meant is that somebody in a former church in San Francisco kept a copy without being asked. Then I look up the thing I was arguing about, and the encyclopedia tells me how it turned out, and the footnote on the encyclopedia goes, as often as not, to that same copy.

Nobody merged these two things. They just grew into each other, bot by bot, link by link, because a few hundred people on each side decided that remembering was worth doing at a loss. The internet’s memory is a volunteer arrangement between two charities, and it has outlasted every company that promised to replace it. It will outlast the next one too, provided the invoices get paid.

So no, Wikipedia is not buying the Internet Archive. The Internet Archive is not for sale, and the Wikimedia Foundation does not buy things, it sends banners. But if you want to know what the merged organisation would look like, you can find out today. Open any Wikipedia article. Scroll to the references. Click one of the ones that goes to web.archive.org.

That is the merger. It has been running since 2015. It just never got a press release. And it would like, at some point, to be paid.

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.