Selling access is not stewardship

On July 16, the president of Wikimedia Enterprise published the Foundation's fullest public defense of its commercial program to date. I read it with real interest, especially after my conversations with Lane in 2024 and 2025, which ended with my questions unanswered. The timing — eight days after my book on this history went to press — is hard to ignore.

The post deserves a response on the merits, because it is candid in ways the Foundation has rarely been, and because its candor concedes more than it defends.

Start with what the post gets right. Trillion-dollar companies should not consume donor-funded bandwidth for free. Charging them for high-volume, high-reliability access is fair, and the license has always permitted it. The 30 percent revenue cap is real governance. The content, as Mr. Becker says, remains free for human readers. If the question were merely “may a nonprofit sell an API,” there would be nothing to write. The question is what the program is for, and here the post is more honest than five years of Foundation messaging before it:

“Wikimedia Enterprise's commercial strategy isn't licensing-based. It's access-based.”

Hold on to that sentence. Wikipedia's volunteers wrote an encyclopedia under two conditions, printed at the bottom of every page since 2001: credit the work, and share what you build from it. Attribution and share-alike are the entire bargain. In March 2024, the Foundation's own staff admitted in writing that many AI companies “may not be compliant with the letter of the Creative Commons rules or the spirit of the licenses.” The remedy it built is a program whose president now states, as a first principle, that the strategy is not about the license. He is right. That is precisely the problem.

The infrastructure argument, audited

The post's central claim is financial: “while our content is free, our infrastructure is not,” and donors should not carry commercial crawlers' load. True. Now put numbers on it, from the Foundation's own audited financial statements for the fiscal year ending June 30, 2025.

Internet hosting — the specific line the post's whole cost argument rests on — was $3.5 million last fiscal year. That is 1.8 cents of every expense dollar. The same year, the Foundation grew revenue to $208.6 million, held $296.6 million in net assets, and sat beside a legally separate endowment of $169.4 million. Salaries ran thirty-three times the hosting bill. The crawler strain on the servers is real; I don't doubt the engineers feel it daily. But an institution of this size does not need a commercial product to survive its bandwidth bill, and it is telling that the post never prints the number. The cost of “free” was covered, sixty times over, before the first Enterprise contract was signed.

Enterprise itself brought in $8.3 million — 4 percent of revenue, from thirteen paying customers. Worth having. Not, on the Foundation's own books, the thing keeping Wikipedia alive.

“It's working” — for whom?

The post declares victory: five years in, Enterprise has repaid its investment and contributes to sustainability. Measured as a business, true. Measured as protection — the word in the title — the largest case sits unmentioned. OpenAI's chatbot leans on Wikipedia more than any other source; the analytics firm Profound found Wikipedia in nearly half of ChatGPT's top-ten citations. OpenAI is not an Enterprise customer. It pays nothing, attributes nothing, and appears nowhere in a post about getting AI companies to pay their share. A program that collects from the willing while the heaviest user rides free is not protecting Wikipedia from AI. It is billing the polite.

The rate limits, and what they reveal

The post announces news: the Foundation has “begun rate-limiting large-scale reusers” of the free public APIs, so that for high-volume users Enterprise becomes “a must-do rather than a nice-to-have.” After years of insisting it could not confront extractors, the Foundation has found an enforcement instrument it is willing to use: the throttle. The license, the terms volunteers actually wrote for, stays in the drawer.

In 2017 the movement adopted a strategic direction that stated, in its own words: “By 2030, Wikimedia will become the essential infrastructure of the ecosystem of free knowledge.” One of its two pillars was “Knowledge as a service: a platform that serves open knowledge to the world across interfaces and communities.” The normalized schemas and realtime feeds Enterprise sells are that pillar, built and then fenced. Now the free path is being narrowed to make the fenced path mandatory.

There is a name for degrading an open resource to sell the reliable version, and it is not stewardship.

The two missing words

Run a search on the post. In roughly 1,400 words about protecting a copyleft commons from AI extraction, attribution appears zero times. Share-alike: zero. Creative Commons: zero. Compliance: zero. The word “licensing” appears exactly once, in the sentence disclaiming it as the strategy. The volunteers' terms are not argued against or deferred to a future post. They are unspoken. Every instrument the Foundation has engineered — schemas, billing, service tiers, now traffic shaping — routes around the one promise the whole commons was built on.

The community saw this coming before Enterprise had its first AI customer. On December 10, 2022, ten days after ChatGPT launched, a Wikimedia mailing-list thread asked whether the new AI was trained on Wikipedia. The first reply came within the hour, linking the GPT-3 paper: “I would argue it's violating the copyright but I'm aware the foundation isn't very interested in defending it.” Four years later, the president of Wikimedia Enterprise has confirmed that prediction as official strategy, in one sentence: not licensing-based. Access-based.

Mr. Becker's post ends by telling companies: pay for the value you derive so we can keep the commons free forever. Mine ends where the book ends. The volunteers' bargain, standing since 2001, trades use for credit and openness. Meter the commons if you must. A steward's first duty is to the promise, and the promise holds two words that this program, by its own president's account, was not built to defend.