The SPUR Coalition Wants AI to Pay for Content. Nobody Has Built the Meter.

The SPUR Coalition Wants AI to Pay for Content. Nobody Has Built the Meter.

SPUR added three members. The New York Times. The Economist. Agence France-Presse. The announcement moved through the tech and crypto press with one clean framing: media organizations forming a bloc to force AI companies to pay for the content their models consume. I read the release twice. Then I did what I do with every protocol announcement — I went looking for the execution layer. The mechanism that meters usage, attributes a price, and settles the transaction. It is not in the release. It is not in any of the follow-ups. There is a demand letter dressed as a pricing regime, and the press is reporting the letter as though the regime already runs.

That gap — between the demand and the meter — is the story. The rest is positioning.

I have spent twenty years reading code and contracts, and I have learned to separate the two. A contract is an intent. Code is an outcome. SPUR's intent is unambiguous: make AI companies pay for the journalism their models ingest. The outcome depends on a metering layer that does not exist at production scale. So before anyone prices this event, before any fund rotates into "AI content" as a theme, the honest question is not whether media should be paid. It is whether anyone can prove, today, which article was used, how many times, and in what context. Right now the answer is no.

The code executes, not the promise. And nobody has shipped the code.

What SPUR Actually Is

SPUR is a coalition of content rights holders seeking collective terms for the use of their material in AI training and generation pipelines. The pitch is simple. Individual publishers negotiating alone against OpenAI or Google have no leverage — the buyer has thousands of substitute sources. A coalition changes the arithmetic by bundling high-value corpora into a single negotiating counterparty.

The three new names matter because of what they represent. The New York Times is the reference corpus for investigative journalism and the plaintiff in the most-watched copyright suit in the field. The Economist is dense, structured, premium analysis — the kind of text that improves a model's reasoning on economics and geopolitics. Agence France-Presse is one of the three global wire services, a machine that produces factual, timestamped, globally distributed reporting in volume.

That combination covers three distinct classes of training material: investigative depth, analytical density, and real-time factual wire. It is a deliberate bundle. It is meant to look unavoidable.

The coalition's stated goal is to establish norms and terms for paid AI content use. That is the language of a standards body. What SPUR is building, functionally, is a collective licensing entity — the same structural form as a performing rights organization in music. The question is whether the form can be filled with an enforceable mechanism. Music figured this out over a century of litigation and legislation. SPUR is trying to compress that into a few negotiation cycles, against a buyer that can walk away from the table and train on a different corpus entirely.

The Timeline Nobody Is Reading Correctly

To understand where SPUR sits, you have to lay out the licensing war in order. It did not start with SPUR. It started with a series of bilateral deals that quietly set the price anchors.

News Corp signed a deal with OpenAI reported in the range of roughly $25 million per year. Axel Springer, the Associated Press, and Le Monde signed similar arrangements. These are the reference prices. They are also the reason SPUR exists — once a handful of large publishers had signed, the remaining publishers faced a binary: accept a similar bilateral deal, or organize to demand better terms collectively.

The New York Times chose a different path. It filed suit against OpenAI and Microsoft, alleging that the models were built on its copyrighted work without license. That suit is the leverage behind the coalition membership. A publisher that is simultaneously negotiating through SPUR and litigating on its own is running a classic two-track strategy. Litigation raises the cost of the negotiation for the buyer. Membership in the coalition raises the reputational and regulatory cost of refusing.

Then came the coalition expansion. And here is the detail that the press skipped. The announcement did not include a complete member list, a founding date, a legal structure, or any statement of whether a single AI company has agreed to negotiate with SPUR as a bloc. Those omissions are not incidental. They are the difference between a functioning cartel and a press release.

The Metering Problem

This is the part that decides everything, and it is the part nobody wants to discuss because it is boring and technical and it does not fit a headline.

To charge for content use, you need to measure content use. Measurement requires attribution: the ability to trace a specific model output or training update back to a specific piece of source material. Attribution at scale is an unsolved problem.

Consider what "using" content means in a modern pipeline. There are at least four distinct stages where copyrighted text enters the system. First, pre-training — the article is in the corpus, and its statistical fingerprint is smeared across billions of parameters. Second, fine-tuning — the article may be in a supervised dataset that shapes instruction-following. Third, retrieval — the article may be fetched at inference time by a retrieval-augmented system and quoted directly. Fourth, generation — the model may reproduce the article's structure, facts, or phrasing without ever retrieving it.

Each stage implies a different licensing regime and a different price. Pre-training is a bulk, one-time, diffuse use. Retrieval is a discrete, countable, per-query use. Generation is a derivative-work question. SPUR's public framing collapses all four into "paid AI content use." That collapse is convenient for a press release and useless for a billing system.

I have audited systems where the gap between the advertised mechanism and the executed mechanism was the entire risk surface. In 2021 I reviewed the ERC-721 implementations of ten trending NFT marketplaces. I found a common flaw in royalty enforcement that put roughly $5 million in creator revenue at risk. The platforms advertised royalties. The contracts did not enforce them. Transfers routed around the royalty hook. The code executed, not the promise — and the promise was worth nothing until the code changed.

The AI content licensing debate is the same failure mode at a larger scale. The publishers are promising a billing regime. The metering infrastructure does not exist. Until it does, "paid AI content use" is a slogan, not a settlement layer.

What Attribution Technology Actually Does Today

I want to be precise here, because the gap between research papers and deployed systems is where most retail analysis goes wrong. There are five families of attribution technology in play. None of them supports a per-article billing meter today.

Influence functions. These estimate how much a single training example changed a model's output on a given input. The math is elegant and the compute cost is brutal. Influence estimation typically requires gradient computations that approach the cost of retraining. You cannot run influence functions across a corpus of trillions of tokens and produce a per-article invoice. It is a research instrument, not a metering service.

Watermarking. This embeds a detectable signal in generated content. It answers "was this text generated by a model" — it does not answer "which source articles did this model learn from." Watermarking is provenance for outputs, not attribution for inputs. The two are constantly confused in public discussion. They are different problems with different maturity curves.

Content credentials (C2PA-style). These attach signed metadata to a piece of media, recording its origin and edit history. This is provenance for the content itself — useful for tracking where a photo came from, useless for proving that a language model ingested a specific article. Metadata is not the asset; the token is.

Retrieval-based attribution. For systems that fetch documents at inference time, you can log the retrieval and bill the retrieval. This is the only one of the five that produces a clean, countable event. But it only covers retrieval-augmented systems. It says nothing about the trillions of tokens absorbed during pre-training. It meters the smallest slice of usage.

Training-data lineage tracking. This is the hard one. You would need to tag every document at ingestion, propagate that tag through every transformation — deduplication, filtering, tokenization, mixing — and preserve a usable link to the model weights. No major model has a public, auditable lineage graph of this kind. Deduplication alone destroys most of the signal, because the pipeline's job is to remove near-duplicates, which is exactly the information you would want to keep for attribution.

So the metering stack, honestly assessed, is this: one narrow mechanism that works for a fraction of usage, and four research directions that do not scale to invoices. That is the foundation under a demand for payment.

The Cartel Economics

The structural comparison everyone reaches for is music. Performing rights organizations — ASCAP, BMI — pooled millions of songwriters' rights and sold blanket licenses to radio, venues, and later streaming platforms. The model worked because three conditions held: the rights were concentrated in a few clearinghouses, the users could not easily substitute away from the licensed catalog, and the law backed the right to collect.

SPUR satisfies the first condition. It is trying to build the clearinghouse. It does not satisfy the second or the third, and that is where the analogy breaks.

Start with the price anchor. If News Corp's deal is roughly $25 million a year, that is the market's revealed price for a top-tier news corpus. Multiply that across the major publishers and you get a content licensing spend for the largest AI labs in the tens to low hundreds of millions of dollars annually. Compare that to compute budgets measured in billions. Content licensing lands somewhere under one to five percent of the cost base. It is a rounding error against capital expenditure.

This is the single most important number in the entire debate, and it cuts against the hype. Payment, by itself, is not a financial threat to AI companies. A five percent cost line does not change a business model. What changes a business model is scope — what the license covers, what it forbids, and what happens when a model violates the terms. Price is negotiable. Scope is existential.

The coalition understands this. The press does not, because "AI pays for news" is a cleaner headline than "AI must define the derivative-work boundary of statistical learning." The real negotiation is about the second sentence.

The Royalty Parallel I Have Seen Before

I audited royalty enforcement in the NFT boom, and the pattern is instructive because it is the closest precedent we have for a digital-first content royalty.

NFT royalties were supposed to pay creators on every secondary sale. Marketplaces implemented a royalty hook in the transfer function. Then marketplaces that wanted volume routed around the hook, or moved to zero-royalty trading, and the enforced royalty collapsed. The lesson was not that royalties are impossible. The lesson was that a royalty is only as strong as the enforcement point, and if the enforcement point is discretionary, the royalty is discretionary.

AI content royalties have the same structural vulnerability, but worse. In NFT, the transfer was on-chain and observable. In AI, the "use" is off-chain, distributed across training runs, and unobservable by default. There is no transfer function to hook. There is no event to index. The publisher cannot see when its article contributes to a model update any more than a songwriter can see a radio play without a reporting system — except the music industry built the reporting system, and the AI industry has not.

This is why I keep returning to the same discipline. Audit first, invest later. Before you believe a payment regime exists, find the enforcement point. In SPUR's case, the enforcement point is either a court order, a statute, or a technical metering layer. Two of those three are external to the coalition. The third does not exist yet.

The On-Chain Provenance Narrative — And Why It Is Oversold

Here is where I have to be direct with my own industry, because this is the trap.

SPUR is being covered by Crypto Briefing — a blockchain-vertical outlet — even though the story contains no blockchain element whatsoever. That is a signal. It suggests the topic is being positioned as narrative fuel for on-chain content provenance, tokenized copyright, and creator-economy protocols. I have watched this pattern before. A real-world problem emerges, a crypto narrative attaches to it, and a wave of tokens launches before a single working product exists.

The on-chain provenance pitch is seductive. Register content on-chain. Mint a rights token. Let smart contracts meter usage and stream royalties. The pitch fails on the same two points as everything else in this space: the oracle problem and the attribution problem.

An on-chain registry can prove that a rights token exists and who holds it. It cannot prove that a specific model used the underlying content. That fact originates off-chain, in a training pipeline, and any bridge from that pipeline to the chain is an oracle — a trusted reporter. If the reporter is the AI company, the publisher is trusting the counterparty to self-report usage. If the reporter is a third party, you have reintroduced the metering problem you were trying to solve, just with a blockchain bolted on.

This is the same error I see in the data availability debate. Every rollup wants a dedicated DA layer, and ninety-nine percent of them do not generate enough data to need one. The narrative outruns the load. Provenance is the DA of the content-licensing conversation — an infrastructure story sold ahead of the demand that would justify it.

Zero knowledge, infinite accountability. If you want a cryptographic settlement layer for content rights, you need a proving system that can attest to usage without revealing the model or the corpus. That is a genuinely hard ZK problem — circuit design for attribution proofs, witness generation at scale, verification cost that does not exceed the value of the license. I reviewed an institutional-grade ZK-rollup in 2025 and found the circuit overhead running fifteen percent above the advertised figure. And that was for a comparatively well-scoped proving problem. Attribution proofs are harder, because the statement being proven is statistical, not deterministic. A ZK proof can attest that a computation ran. It cannot cleanly attest that a model "used" an article in the sense a copyright claim requires.

I am not saying on-chain rights management is impossible. I am saying it is being sold as a solution to a problem it cannot yet address, by parties with a token to launch. Be skeptical of anyone who claims the provenance problem is solved on-chain. The chain is the easy part. The attribution is the hard part, and it is off-chain.

The Buyer's Substitution Set

The coalition's leverage depends on how badly the buyers need its content. This is the variable that decides the outcome, and it points the wrong way for SPUR.

An AI company facing a licensing demand has five substitution options. First, train on public-domain material — centuries of text, free and unencumbered. Second, train on user-generated content, which is voluminous and often already licensed by the platform. Third, train on synthetic data generated by existing models, which sidesteps copyright entirely. Fourth, license from the subset of publishers willing to sign bilaterally, which fragments the coalition. Fifth, accept the license cost, which — at under five percent of compute spend — is affordable.

The critical asymmetry is this: the seller has one monetization path for AI use (license it), while the buyer has five sourcing paths. Buyer substitution options exceed seller alternatives. That is the textbook condition for weak seller pricing power, and no amount of coalition branding changes the underlying economics.

The selective-signing strategy is the buyer's weapon. OpenAI and Google have already signed bilateral deals with News Corp, the AP, Le Monde, Axel Springer, and others. Each bilateral deal does two things. It secures the corpus the buyer actually wants. And it creates a free-rider and first-mover incentive inside the coalition — the member who signs first gets paid, and the member who holds out gets nothing while the corpus gets devalued by the model's continued training on substitutes.

A coalition is only as strong as its exclusivity clause. If members can sign bilaterally, the coalition is a pressure group, not a cartel. SPUR has not published its exclusivity terms. That omission is the whole ballgame.

The Legislative Lever Is Stronger Than The Coalition

If I were advising a content rights holder, I would tell them the coalition is the weaker of two levers. The stronger lever is legislation.

The European Union's AI Act imposes transparency obligations on training data. The United States Copyright Office has issued guidance on the copyrightability of AI outputs and is weighing positions on training use. Japan has taken a comparatively permissive stance on training data, explicitly carving out broad uses. These are the rules that will actually bind AI companies, because a statute is enforceable in a way a coalition is not.

Here is the mechanism. A coalition negotiates price. A statute defines scope. Price is a five-percent cost line. Scope determines whether a model can be built at all. If the EU mandates disclosure of training corpora and the US courts narrow fair use, the compliance cost dwarfs any licensing fee SPUR could negotiate. The coalition's real value may be as a lobbying and publicity instrument — manufacturing the reputational cost that pushes legislators toward the scope rules the publishers actually want.

This reframes the event. SPUR is not primarily a pricing play. It is a political play wearing pricing clothes. The membership announcements from NYT, Economist, and AFP are signaling — to regulators and to the public — that content owners will not passively accept unilateral terms. That signal is worth more in a legislative hearing than in a negotiation room.

The Public Media Blind Spot

There is a class of content owner that SPUR's model handles badly, and it exposes the ethics of the whole enterprise.

Public broadcasters — the BBC, NPR, and their peers — produce content with a public-goods character. Their output is funded by license fees or public money and is meant to circulate freely. If AI companies pay commercial publishers for training rights but public broadcasters are excluded from the paid regime, the public content is either used for free or removed from training. If public broadcasters are included, you create a strange outcome: publicly funded content is licensed as private property, and the public pays twice — once to produce it, once to buy AI services trained on it.

This is the distributional-justice core of the debate, and it is where the "AI pays for news" narrative is weakest. The strongest argument for payment is sustainability: high-quality journalism is expensive to produce, and if AI absorbs it without compensation while eroding its monetization, the supply of quality reporting shrinks. That argument is real. But it is an argument about sustaining a public good, and the coalition's mechanism — a commercial licensing cartel — is not obviously the tool that serves it.

If payment only reaches the head of the distribution — the Timeses and the Economists — the tail is worse off than before. Small and independent publishers cannot afford to litigate, cannot command bilateral deals, and may be excluded from coalition terms. The result is an AI content oligopoly: a handful of large publishers get paid, the long tail gets scraped for free, and the "fair compensation" framing masks a consolidation of the content economy into fewer, larger hands.

The Antitrust Trap

Here is the contrarian angle, and it is the one that could kill SPUR outright.

Collective price-setting among competitors is, in most jurisdictions and especially in the United States, a per se antitrust violation. A group of publishers agreeing on the price at which they will license their content to AI companies is, on its face, price-fixing. The music industry avoided this through a combination of statutory exemptions and consent decrees that governed the collecting societies for decades. SPUR has no such exemption, and there is no sign it has sought one.

The irony is sharp. The coalition's strength — collective negotiation — is also its legal vulnerability. If SPUR functions as a genuine cartel, it invites antitrust scrutiny that could dissolve it. If SPUR functions as a loose pressure group with bilateral escape hatches, it lacks the pricing power that justified forming it. The coalition has to thread a needle between being too coordinated (illegal) and too loose (ineffective).

There is a second-order risk. If SPUR triggers antitrust action, the case becomes a precedent about how content-licensing collectives may operate, and the outcome could constrain the very model SPUR is trying to establish. A coalition that overplays its hand could set back collective content licensing by a decade.

The blind spot in the bullish coverage is that it treats SPUR as a pure positive for content owners. It is not. It is a legal structure with an unresolved antitrust question at its center, and the question is not cosmetic. It is load-bearing.

What Actually Threatens The AI Companies

Let me be precise about the risk vector, because the retail framing gets it backwards.

A licensing fee of tens to low hundreds of millions per year does not threaten an AI lab with a multi-billion-dollar compute budget. That cost is absorbable. What threatens the labs is a court ruling that narrows fair use in a way that invalidates the training of existing models, or a statute that imposes disclosure and consent requirements retroactively. A ruling of that kind triggers two exposures: a one-time damages award on the scale of the NYT suit, and a forward-looking constraint on how future models can be built.

This is why the NYT litigation matters more than the SPUR coalition. Litigation can change the legal baseline. A coalition can only change the price. The price is small. The baseline is everything.

Watch the fair-use rulings. Watch the EU AI Act's implementing rules. Watch whether any AI company negotiates with SPUR as a bloc — because a bloc deal would validate the collective model, while continued bilateral signing would confirm that the coalition is a pressure group, not a counterparty. The coalition's own behavior under exclusivity pressure is the cleanest signal of whether it has real leverage.

The Investment Layer, Stripped Of Narrative

If you are looking for tradeable exposure, be disciplined. The direct beneficiaries — large media groups receiving licensing income — will see contributions well under one percent of total revenue. A $25 million annual deal against a multi-billion-dollar media group is not a re-rating catalyst. Treat any framing of licensing income as "AI strategy progress" with the skepticism it deserves; strip out one-time payments before believing in a durable line item.

The more interesting — and more dangerous — layer is the infrastructure. Copyright management, data provenance, and compliance tooling are real needs. The "shovel sellers" of a compliance regime are where durable value could accrue, because they sell to both sides. But this is also where the crypto narrative does the most damage, because provenance protocols and tokenized-copyright projects will trade on the theme long before they ship a working meter.

My rule stands. Audit first, invest later. For any provenance or rights-management protocol pitching itself on the SPUR theme, ask three questions. Where does the usage data come from — is it self-reported by the AI company? What is the enforcement point if the meter is wrong? And what is the cost of the proof — does the cryptographic overhead exceed the value of the license it secures? If the answers are vague, the token is a narrative instrument, not an infrastructure investment. The code executes, not the promise, and in this sector the code is mostly absent.

The Real Blind Spot: Scope, Not Price

Everyone is arguing about whether AI should pay. Almost nobody is arguing about what the payment buys. That is the blind spot.

A licensing fee is a rounding error. A license's scope is a constraint on product design. If a publisher licenses its content only for retrieval-augmented use and forbids pre-training, the AI company must segregate its corpora and its pipelines. If a license requires that any model output substantially similar to licensed content trigger a royalty, the company must run attribution on every generation — the unsolved problem again. If a license forbids training on content behind a paywall entirely, the company's corpus shrinks and its model quality degrades.

These are not pricing questions. They are architecture questions. And they are the questions SPUR's public materials do not answer, because answering them would reveal that the coalition is asking for a scope regime it cannot yet enforce. The press reports the price because the price is legible. The scope is where the actual negotiation will be decided, and the scope is invisible to anyone who is not reading the terms.

The Takeaway

The SPUR coalition is a real signal wrapped in an unproven mechanism. It marks the shift from bilateral licensing to collective bargaining, and it puts pressure on legislators at exactly the moment the rules are being written. That is strategically meaningful. But the payment regime it demands rests on a metering layer that does not exist — attribution technology is years from supporting per-article billing, and the on-chain provenance narrative is selling a solution to a problem it cannot yet address.

The forward question is not whether AI companies will pay. At under five percent of compute spend, they can afford to. The forward question is who defines the scope of use, and who can prove a violation when it happens. The coalition that answers those two questions first wins. The one that only negotiates a price wins a headline and loses the war.

Immutability is a feature, not a flaw — and so is measurement. If you cannot meter it, you cannot bill it, and if you cannot bill it, you are not being paid. You are being promised.