Three names landed in the same coalition this week, and the price of news changed shape. The New York Times. The Economist. Agence France-Presse. All three joined SPUR, a coalition of publishers pushing for paid use of their content in AI training and generation. No timestamped press release. No member roster. No legal filing. Just a headline that reads like a settlement and behaves like a promise. I have spent ten years watching crypto projects announce partnerships that never shipped. This one smells the same: a claim of coordination where the only verified fact is the announcement itself. The chart didn't move on any publisher's stock, because the market already knows the difference between a deal and a demand. The event isn't the deal. The event is the signal. And the signal is weaker than the headline.
Here is the thing I keep circling back to, as an editor who has stood on both sides of this fence. The story is being told as a copyright story. It isn't. It is a data-infrastructure story wearing a copyright costume. And until somebody builds the plumbing, the billing is theater.
What SPUR actually is, and what it isn't
SPUR is a coalition of content owners seeking collective leverage over the AI companies that train on their material. The acronym's own expansion is unclear in public reporting, which tells you how thin the sourcing runs. Think of the mechanism as ASCAP for the training corpus: pool the rights, set a price, negotiate as one bloc. The strategy is a century old. Music publishers did it in 1914. Patent pools did it through two world wars. Now journalism wants in.
The timing matters more than the names. The dominant AI labs have already signed a scatter of individual content deals โ News Corp, Axel Springer, the Associated Press, Le Monde, and others. Each reportedly lands in the tens of millions of dollars per year. Those deals set an anchor. They also set a ceiling. And they expose the structural weakness SPUR exists to fix. When the buyer can pick off suppliers one at a time, there is no market price. There is only whatever the weakest supplier will accept.
That is the entire thesis. Everything else is decoration.
So let me strip the decoration. A coalition is not a contract. A demand is not a payment rail. And a press mention is not a settlement. I have audited enough on-chain treasuries to know the difference between a commitment and a deposit. SPUR, right now, is a commitment. The deposit has not cleared.
The economics of a supply-side cartel, priced honestly
Start with the number everyone skips past. If the leading lab's deal with a major publisher runs around twenty-five million dollars a year, then compare that to what the same lab spends on compute. Tens of billions annually, across clusters, power contracts, and silicon. Twenty-five million is a rounding error inside a data center's electricity bill. Content licensing, at current prices, does not threaten any AI company's financial model. It barely registers as a line item.
That single fact reframes the whole negotiation. The coalition's demand is not a financial threat. It is a reputational and legal threat. The publishers know this. The labs know this. The only people who seem not to know it are the readers being sold the headline that content owners are about to extract real value from AI.
I want to be precise, because precision is the only thing that survives a sideways market. There are three variables that decide whether a supply-side cartel works, and I can score each one against the public evidence.
Variable one is exclusivity. Does SPUR forbid its members from signing solo deals? Nobody has shown me the term sheet. Without an exclusivity clause, the coalition is a suggestion, not a bloc. The music industry learned this the hard way: publishers who defect early get paid first and paid best. Variable two is dependency. How badly does a given AI lab need a given corpus? The dominant lab can crawl the open web, license public-domain archives, generate synthetic text, or simply train on user data. Its dependency on any single publisher is soft. Variable three is the billing mechanism. And this is where the whole edifice wobbles.
Because you cannot bill for usage you cannot measure.
The provenance problem: you can't invoice what you can't trace
This is the section the copyright lawyers will not write, because it is not a legal problem. It is an engineering problem, and the engineering is not done.
Suppose SPUR wins tomorrow. Suppose every AI lab agrees in principle that publishers should be paid per use. Now answer the operational question: how many times was a specific Economist article used to train a specific model, and how many times did that model reproduce its structure in a generated answer? There is no production-grade system that answers this at the precision a billing system requires. There are research directions. None of them are ready for an invoice.
Chasing the ghost in the smart contract code is a familiar exercise for me, but here the ghost is in the model weights, and it is far harder to catch. Let me walk through what actually exists.
Training-data provenance means tagging the origin of every document that enters a training run. In practice this is a logging problem that most labs treat as optional. Pipelines deduplicate, filter, and shuffle. They do not, as a rule, preserve a per-document ledger of influence. Even the labs that claim to track sources are tracking where a file came from, not how much it shaped the model.
Attribution is the harder layer. Influence functions try to estimate how much a single training example changed a model's output. They are expensive, approximate, and famously unstable on large networks. Watermarking embeds detectable signals in generated text, but it marks outputs, not inputs, and it can be stripped by paraphrasing. Retrieval-based attribution works when the model fetches a document at inference time โ that use case is genuinely traceable โ but it says nothing about the training corpus.
Put those three together and you get the real constraint. The technical stack can support retrieval-time billing. It cannot yet support training-time billing at the granularity publishers want. Which means any near-term deal will be priced on a proxy โ a flat annual fee, a seat count, a crawl volume estimate โ not on measured use. And a proxy price is exactly what a coalition is supposed to escape. The moment you price on a proxy, you are back to a bilateral negotiation with extra steps.
I have run this kind of audit before. In 2024 I traced the inflow patterns of the first spot Bitcoin ETFs and found that a third of early money came from micro-cap funds that had been active in DeFi. The lesson I carried out of that work: you can only make a claim as strong as your data lineage. If your data lineage is broken, your claim is a rumor with a spreadsheet attached. SPUR's claim has a spreadsheet. The lineage is the missing piece.
Scanning the block for the missing brick is the right metaphor here, and the missing brick is measurement.
The crypto crowd is about to get this wrong โ again
Now the part I have to write carefully, because my own industry is circling this carcass.
The moment a story about content ownership and AI money appears, a certain segment of crypto Twitter reaches for the same hammer. On-chain content provenance. Tokenized copyright. Creator-economy protocols that register IP on a ledger and pay out royalties in a stablecoin. It is a beautiful story. It is also, for this specific problem, mostly the wrong tool.
Here is why. Blockchains are excellent at two things: proving that a specific transaction happened, and enforcing rules without a trusted middleman. They are terrible at measuring off-chain reality. The hard part of AI content billing is not recording a payment. It is determining whether a use occurred and how large it was. A ledger does not know how many tokens of a New York Times investigation entered a training run. It only knows what an oracle tells it. And the oracle for 'how much did this model learn from this article' does not exist.
So when someone pitches you an on-chain solution to AI copyright, ask them one question: what feeds the oracle? If the answer is 'we estimate it,' you have not built a provenance system. You have built a spreadsheet with a token attached.
This is not to say the blockchain tooling is worthless. On-chain registries are genuinely useful for timestamping ownership claims โ proving who declared a work first, in a way no central authority can quietly rewrite. That is a real, if modest, contribution. What it is not is the measurement layer that makes usage-based billing possible. Follow the scholar, not the token. The scholar here is the engineer who can actually measure model influence. The token is the thing being sold to you before the measurement exists.
I have watched this exact pattern in the AI-agent scam space. In 2025 I deployed a counter-agent against a hundred suspected scam bots and found a coordinated network of fifteen projects using AI to mimic legitimate influencers. The tell was always the same: they sold the narrative of a capability before the capability existed. The provenance token pitch has the same fingerprint. Beautiful claim. Empty vault. Beneath the surface, the nest was empty.
The selective-licensing fork that quietly disarms the coalition
The coalition's real enemy is not the dominant lab. It is the lab's signing pen.
Watch the pattern. When a cartel forms, a rational buyer does not fight it head-on. The buyer splits it. Sign the strongest member first, at a premium, on a solo deal. The second-strongest sees the check and wants one too. Within a quarter, the bloc is a queue of individual sellers, each racing to be picked. This is not a conspiracy. It is basic game theory, and the AI labs are running it in plain sight. They have already signed News Corp, Axel Springer, the AP, Le Monde โ one at a time, each on its own terms.
A coalition's power is inversely proportional to how many of its members have already signed alone. And several of the loudest names in this fight have signed alone.
That is why the New York Times' position is the most interesting one on the board. The Times is doing two things at once: it has joined the collective push, and it is litigating against the dominant lab and its partner. Negotiation and litigation in parallel is not indecision. It is leverage engineering. The lawsuit raises the cost of the status quo, which raises the floor of any settlement. The coalition membership raises the reputational stakes. Both moves serve the same end. It is the most sophisticated play in the entire affair, and it comes from the party with the most to lose.
The contrast is instructive for anyone who thinks coalitions are about solidarity. They are about relative position. The Times can afford to litigate. A regional paper cannot. The coalition gives the small player a seat at a table it could never buy alone, but it also gives the large player a crowd to hide behind while it negotiates its own exit. Both things are true at once.
The music-industry analogy, and where it breaks
Everyone reaches for music. Collective rights management, performance royalties, the whole ASCAP/BMI playbook. It is a good analogy as far as it goes, and it breaks in three places that matter.
First, music has a defined unit of use. A radio play is a play. A stream is a stream. The measurement is external, observable, and auditable by a third party. AI training has no equivalent unit. There is no 'play' of a training document. That single asymmetry guts the analogy's core.
Second, music had a captive distribution channel. Radio stations and streaming platforms needed the catalog to operate. They could not substitute their way out of it. An AI lab can substitute. It can crawl, synthesize, license the public domain, or simply train on a slightly worse corpus and accept a small quality hit. A cartel only works when the buyer cannot walk away. This buyer can walk, and it can walk quickly.
Third, music royalties flowed through a mature collection infrastructure built over decades โ societies, standards, audit trails. AI content licensing has none of that. It would have to be built from scratch, and the measurement problem above says the foundation is not ready.
Speed eats stability for breakfast in this business, and here the incumbents are trying to impose stability on a market that is still forming. The music model took fifty years to mature. SPUR is trying to compress that into an announcement cycle.
The open-source asymmetry nobody is pricing
Here is the blind spot I keep coming back to, and it is the one with the longest tail.
If content-licensing norms harden into law, they hit closed and open models very differently. A closed API provider has a corporate entity, a revenue line, and a legal address. It can sign a deal, absorb a fee, and pass the cost to subscribers. An open-weights model has none of that clarity. Who is liable when a downloadable model trained on copyrighted material is fine-tuned and redeployed a thousand times? The answer is nobody, which is precisely the problem for the open ecosystem.
The practical effect of stricter content licensing is to make open models legally riskier than closed ones, without either side changing a line of code. The closed lab's compliance department becomes a competitive moat. The open project's lack of a legal entity becomes an existential liability. SPUR, whether it intends to or not, is pushing the regulatory thumb onto the closed side of the scale.
I do not think this is the coalition's goal. I think it is a second-order effect that almost nobody in the debate has modeled. But it is the kind of thing that decides which ecosystems survive the next regulatory cycle. When the licensing regime tightens, capital flows to whoever can absorb the paperwork. That is never the open-source project.
There is a parallel story running in China that Western coverage keeps missing. Content cooperation between Chinese AI firms and state media moves through policy guidance, not market negotiation. People's Daily and Xinhua partner with domestic model developers by directive, not by contract. Which means the SPUR model is not portable there, and the global landscape will fork into two regimes โ one negotiated, one directed. A single global 'AI content standard' is a fantasy. Watch the fork, not the framework.
Who actually gets paid โ the distribution question
Now the part that should worry every working journalist, and the part the coalition's press coverage never touches.
Suppose SPUR wins a lump sum. Suppose a pool of licensing revenue lands in a collective account. How does it get divided? By content volume? By traffic? By number of articles a model happened to ingest? There is no agreed formula, and the formula decides everything. A coalition that cannot answer the distribution question is a coalition that will pay its largest members first and its smallest members never.
The structural risk is that AI content money concentrates at the top, exactly where media ownership already concentrates. The Times, The Economist, and AFP have the lawyers, the lobbyists, and the leverage to claim a proportional share. The freelance photographer and the regional newsroom โ whose work was almost certainly in the training corpus too โ have no seat at the table and no way to prove their contribution. The coalition solves the visibility problem for the powerful and deepens the invisibility of everyone else.
I have seen this movie before, and I wrote about it. In 2021 I embedded with Play-to-Earn communities in Jakarta and interviewed fifty scholars and managers. The headline was a game that paid players. The reality was an economy where eighty percent of the revenue went to the administrators, not the players doing the work. The people generating the value were the last to be paid, and the structure was designed so they could not see it. Tokenized content royalties risk the same architecture: a distribution layer that looks fair on the whitepaper and concentrates wealth in practice.
Public broadcasters face a sharper version of this. The BBC and NPR hold content with a public-good character. Paywalling their material behind a licensing regime raises a genuine ethical question โ whether publicly funded journalism should be sold to private models at all, and if so, who captures the proceeds. The coalition's framing treats all content as equivalent private property. It is not.
The antitrust tripwire the coalition is standing on
Here is the risk that the coalition's own press coverage never mentions, and it is a serious one.
Collective price-setting among competitors is, in several jurisdictions โ the United States most notably โ a place where you can walk straight into a lawsuit. The line between legitimate collective licensing and illegal price-fixing is thin, and it is patrolled. A coalition that agrees on a price floor and negotiates as a bloc is doing something that looks, from a certain angle, exactly like the thing antitrust law exists to stop.
The music societies survived this scrutiny because they were granted statutory exemptions and operated under consent decrees. SPUR has no such protection on the record. If the coalition sets a common price and refuses to deal except as a bloc, it invites a counterattack that has nothing to do with copyright and everything to do with competition law. The coalition's greatest legal vulnerability may not be the AI labs. It may be the regulators.
I will say plainly what the coverage will not: nobody has shown me SPUR's legal architecture, its exclusivity terms, or its pricing consensus. Until those documents surface, every claim about the coalition's power is speculation. I am not going to pretend otherwise, and neither should anyone selling you a take.
Verification Protocol
Before any reader acts on this story, here is exactly what I checked, and what I could not.
I could not find a timestamped SPUR announcement with a member roster. The core fact โ that the Times, The Economist, and AFP joined โ traces to a single aggregation, with no link to a primary source. I could not find a filed legal complaint, an agreed pricing schedule, or any statement from an AI lab acknowledging SPUR as a counterparty. I found the existing individual deals โ News Corp, Axel Springer, the AP, Le Monde โ reported independently and consistently across outlets, which is why I treat the anchor number as reliable and the coalition's leverage as unproven.
When a fact appears in one place with no origin, treat it as a claim, not a fact. When it appears across independent sources with documents attached, treat it as data. This story sits in the first bucket, and the difference is the whole ballgame.
The contrarian angle: the point of SPUR is not to win, it is to set the baseline
Let me offer the read that cuts against everything above.
Maybe the coalition does not need to win a deal to succeed. Maybe the victory condition is the norm, not the contract. Every public coalition member makes it slightly more expensive for an AI lab to use content for free. Every headline about paid licensing shifts the default expectation from 'free to scrape' toward 'permission required.' The coalition may be playing a legislative game in commercial clothing โ using the visible act of collective bargaining to move the political baseline, so that when a court finally rules on fair use, the cultural presumption already leans toward payment.
If that is the strategy, it is patient and clever. It also means the coalition's failure to sign a deal this quarter is not failure at all. The metric is the norm, and norms move in years, not earnings calls.
But even here, the engineering bites. A norm without a measurement system is a norm without enforcement. You can legislate that publishers should be paid. You cannot legislate a system that knows how much to pay. Volatility is just liquidity with a pulse, and a right without a meter is just a headline with a pulse. The coalition can move the culture. It still cannot move the number, because nobody has built the scale.
And that is the real investment signal buried under the noise. The money is not in the content, and it is not in the models. It is in the picks and shovels โ the attribution tools, the provenance standards, the compliance infrastructure that lets a lab prove where its data came from. That is a smaller, slower, less glamorous bet than the tokenized-copyright narrative, which is exactly why it is the more honest one. The chart didn't spike on any of it, because the market has not figured out where the real bottleneck is. The bottleneck is the meter.
Takeaway
Watch three things, and ignore the rest. First, whether SPUR publishes an actual member roster and exclusivity terms โ the moment it does, the coalition becomes a counterparty instead of a rumor. Second, the Times litigation, because a ruling on fair use will set the boundary the coalition is trying to negotiate around, and the courts will move faster than the bloc. Third, and most important, watch for the first production-grade attribution system โ the tool that can tell a lab, and a publisher, how much of a specific work entered a specific model. Whoever builds that meter owns the market SPUR is only renting.
The coalition added three famous names this week. It did not add a way to count. Until somebody does, the price of news is still whatever the strongest supplier can prove โ and right now, nobody can prove anything. Speed eats stability for breakfast, and the provenance gap is eating SPUR's leverage while the headlines celebrate its birth.