For a few years the loud fight over AI and knowledge has been about training. Did the model see your work? Was that fair use or theft? Who signed the big licensing check, and for how much?
That fight mattered. It still does. But it has always been a little oddly timed. The expensive moment in the new stack is often not the day a model was trained. It is the day an agent opens your page, your dataset, your filing, or your paper and uses it to answer something for someone else. In that moment, almost nobody gets paid.
On September 22, Firecrawl put a name on the gap. The company, which builds tools that turn websites into something agents can actually read, announced a $75 million Series B and launched Alexandria: a library meant to give agents one way to find sources, see what they hold, and pull from them. Live web, specialized indexes, official data providers, custom connectors. One interface.
Caleb Peffer, Firecrawl's CEO, said the quiet part in the press release: "People have licensed data for training, but almost no one gets paid when an AI agent actually uses what they know."
I have been thinking about that sentence more than about the round.
Training licenses were a settlement with the past. A model company needed a corpus. A publisher had one. They negotiated a lump sum, or a multi-year fee, and both sides could pretend the hard part was over. Publishers with real leverage, the New York Times or News Corp, could get real money. Most everyone else got a shrug and a robots.txt file.
Agents break that story. An agent does not "train" on your Wikipedia page once and move on. It comes back when it needs a fact the weights cannot be trusted for. It grounds an answer in something fetched five seconds ago because the model is stale, incomplete, or wrong for this task. The act that creates value is the read at runtime, not the scrape that happened before the weights were frozen.
Publishers have been inching toward this distinction for a while. By mid-2025, Digiday was already documenting the shift from one-time training deals toward what people call grounding or RAG licensing — pay per crawl, pay per query, pay when content shows up in an answer. The vocabulary is messy. The economic instinct is not. Recurring use is a better business than a single archive sale, if you can meter it.
Alexandria is interesting because it tries to sit in the middle of that meter for agents specifically.
Firecrawl already pays some providers under individual agreements. The clearest example is Wikimedia Enterprise. Since March 2026, Wikipedia requests through Firecrawl have been routed through Wikimedia's paid commercial API. The company says millions of those requests flow through every month. That is not charity branding. It is a recognition that Wikipedia is infrastructure, and infrastructure that gets hit by agents at machine speed needs a bill that is not "hope the donation banner works."
What Firecrawl says it will do next is open that model wider. Individuals, creators, organizations — not only big providers with enterprise sales teams — should be able to put knowledge into Alexandria and get paid when agents use it. The self-service system is not open yet. Rates are not public. How "use" will be measured is still vague. Those are not small details. They are the product.
Still, the bet is clear enough. The missing market is not another crawl of the open web. It is a checkout for knowledge that was never on the open web in a usable form, or was on it but never got paid when a machine read it.
That sounds obvious until you notice how little of the stack is built around it.
Search engines made a different bargain: show up in results, get a chance at a human click, monetize the human. Agents skip the click. They read, summarize, decide, and move on. Attribution may travel with the answer. Money usually does not. Training lawsuits and training licenses tried to price the original absorption of content into models. They barely touch the continuous, high-frequency consumption that agents are normalizing.
So you get a strange economy. The knowledge that makes an agent useful is often the knowledge someone else maintains. The company that owns the agent captures the product value. The model lab captures the token bill. The site that got hit for the grounding passage absorbs the load and, if it is lucky, a citation.
Alexandria is one attempt to insert a counter: pay the source when the agent draws on it. Firecrawl's own pitch for the funding is unusually blunt about where the money is supposed to go. "We're going to spend a good chunk of it buying knowledge from people," Peffer wrote. The library metaphor is doing real work here. The ancient Library of Alexandria gathered scrolls in one place and hoped the roof held. The modern version gathers access rights and hopes the pricing holds.
I do not know whether Firecrawl becomes that market, or whether some other company does, or whether the market fragments into a hundred bilateral deals the way publisher licensing already has. Exclusivity, take rates, and measurement fights are waiting as soon as real money moves. Firecrawl's internal claim that agents using Alexandria scored 21 percent higher on answer quality across 845 tasks is the kind of number you should treat as a company number until someone else reproduces it. The contributor waitlist is still a waitlist.
What feels durable is the diagnosis.
We spent years arguing about whether machines should be allowed to learn from the public web. The sharper problem now is that they operate on it, continuously, on behalf of users who will never visit the original page. If that is the default interface to knowledge, then "did you train on this?" is the wrong primary question. The primary question is closer to: when an agent uses what you know, does that use show up on anyone's ledger?
Firecrawl is answering yes, and trying to build the ledger. Wikimedia Enterprise is already on it. A lot of smaller sources are not. Expertise that never made it online is definitely not. The self-service dream — that a researcher, a local publisher, or someone with a dataset that never had a homepage could earn when agents query them — is either the important part of this story or the slide that never ships. We do not know which yet.
I suspect the companies that win here will look less like scrapers and more like clearinghouses. The hard problem is not fetching HTML. It is deciding what a fair unit of knowledge use even is, proving the use happened, and routing a small payment before the agent has already moved on to the next tool call. That is unglamorous work. It is also the kind of work that, if it works, becomes hard to remove.
The training fight asked whether the past was stolen. The more interesting fight is whether the present read has a price.
That fight mattered. It still does. But it has always been a little oddly timed. The expensive moment in the new stack is often not the day a model was trained. It is the day an agent opens your page, your dataset, your filing, or your paper and uses it to answer something for someone else. In that moment, almost nobody gets paid.
On September 22, Firecrawl put a name on the gap. The company, which builds tools that turn websites into something agents can actually read, announced a $75 million Series B and launched Alexandria: a library meant to give agents one way to find sources, see what they hold, and pull from them. Live web, specialized indexes, official data providers, custom connectors. One interface.
Caleb Peffer, Firecrawl's CEO, said the quiet part in the press release: "People have licensed data for training, but almost no one gets paid when an AI agent actually uses what they know."
I have been thinking about that sentence more than about the round.
Training licenses were a settlement with the past. A model company needed a corpus. A publisher had one. They negotiated a lump sum, or a multi-year fee, and both sides could pretend the hard part was over. Publishers with real leverage, the New York Times or News Corp, could get real money. Most everyone else got a shrug and a robots.txt file.
Agents break that story. An agent does not "train" on your Wikipedia page once and move on. It comes back when it needs a fact the weights cannot be trusted for. It grounds an answer in something fetched five seconds ago because the model is stale, incomplete, or wrong for this task. The act that creates value is the read at runtime, not the scrape that happened before the weights were frozen.
Publishers have been inching toward this distinction for a while. By mid-2025, Digiday was already documenting the shift from one-time training deals toward what people call grounding or RAG licensing — pay per crawl, pay per query, pay when content shows up in an answer. The vocabulary is messy. The economic instinct is not. Recurring use is a better business than a single archive sale, if you can meter it.
Alexandria is interesting because it tries to sit in the middle of that meter for agents specifically.
Firecrawl already pays some providers under individual agreements. The clearest example is Wikimedia Enterprise. Since March 2026, Wikipedia requests through Firecrawl have been routed through Wikimedia's paid commercial API. The company says millions of those requests flow through every month. That is not charity branding. It is a recognition that Wikipedia is infrastructure, and infrastructure that gets hit by agents at machine speed needs a bill that is not "hope the donation banner works."
What Firecrawl says it will do next is open that model wider. Individuals, creators, organizations — not only big providers with enterprise sales teams — should be able to put knowledge into Alexandria and get paid when agents use it. The self-service system is not open yet. Rates are not public. How "use" will be measured is still vague. Those are not small details. They are the product.
Still, the bet is clear enough. The missing market is not another crawl of the open web. It is a checkout for knowledge that was never on the open web in a usable form, or was on it but never got paid when a machine read it.
That sounds obvious until you notice how little of the stack is built around it.
Search engines made a different bargain: show up in results, get a chance at a human click, monetize the human. Agents skip the click. They read, summarize, decide, and move on. Attribution may travel with the answer. Money usually does not. Training lawsuits and training licenses tried to price the original absorption of content into models. They barely touch the continuous, high-frequency consumption that agents are normalizing.
So you get a strange economy. The knowledge that makes an agent useful is often the knowledge someone else maintains. The company that owns the agent captures the product value. The model lab captures the token bill. The site that got hit for the grounding passage absorbs the load and, if it is lucky, a citation.
Alexandria is one attempt to insert a counter: pay the source when the agent draws on it. Firecrawl's own pitch for the funding is unusually blunt about where the money is supposed to go. "We're going to spend a good chunk of it buying knowledge from people," Peffer wrote. The library metaphor is doing real work here. The ancient Library of Alexandria gathered scrolls in one place and hoped the roof held. The modern version gathers access rights and hopes the pricing holds.
I do not know whether Firecrawl becomes that market, or whether some other company does, or whether the market fragments into a hundred bilateral deals the way publisher licensing already has. Exclusivity, take rates, and measurement fights are waiting as soon as real money moves. Firecrawl's internal claim that agents using Alexandria scored 21 percent higher on answer quality across 845 tasks is the kind of number you should treat as a company number until someone else reproduces it. The contributor waitlist is still a waitlist.
What feels durable is the diagnosis.
We spent years arguing about whether machines should be allowed to learn from the public web. The sharper problem now is that they operate on it, continuously, on behalf of users who will never visit the original page. If that is the default interface to knowledge, then "did you train on this?" is the wrong primary question. The primary question is closer to: when an agent uses what you know, does that use show up on anyone's ledger?
Firecrawl is answering yes, and trying to build the ledger. Wikimedia Enterprise is already on it. A lot of smaller sources are not. Expertise that never made it online is definitely not. The self-service dream — that a researcher, a local publisher, or someone with a dataset that never had a homepage could earn when agents query them — is either the important part of this story or the slide that never ships. We do not know which yet.
I suspect the companies that win here will look less like scrapers and more like clearinghouses. The hard problem is not fetching HTML. It is deciding what a fair unit of knowledge use even is, proving the use happened, and routing a small payment before the agent has already moved on to the next tool call. That is unglamorous work. It is also the kind of work that, if it works, becomes hard to remove.
The training fight asked whether the past was stolen. The more interesting fight is whether the present read has a price.