AINEWS
Ziff Davis's Vivek Shah calls the cost a "rounding error" beside OpenAI's compute plans, reframing the publishers' copyright fight as a market design problem rather than an affordability one.
Ziff Davis CEO Vivek Shah's new Fortune op-ed, as reported by CNET, does not contest whether AI companies should pay news publishers for the journalism their models were trained on. He reframes the question. A working licensing market for U.S. news would cost roughly $20 billion a year, Shah writes, a figure he anchors to the country's existing music-licensing market. That sum is "a rounding error" beside the $750 billion OpenAI has told investors it expects to spend on computing infrastructure by 2030. Once the price is set against the buyer's other costs, the live question is how the market should be structured, not whether the buyer can pay.
That argument lands as the consolidated copyright cases against OpenAI move from discovery into initial fair-use briefing in the Southern District of New York, the procedural phase where the court will be asked to rule on whether training large language models on publishers' archives is legally permissible. The plaintiffs include The New York Times, Ziff Davis (the parent of CNET), and the New York Daily News, with other suits consolidated as a single multidistrict case, MDL 1:25-md-03143. Initial summary-stage briefs were filed this month asking the court to treat the training use as infringement rather than fair use.
The political intervention came a week earlier. The Department of Justice filed a statement of interest in the case arguing that mandatory licensing regimes would burden AI companies, and that disputes over training data should be routed through Congress rather than the courts. Shah's op-ed is the most public rebuttal to that position from a named plaintiff, and it leans on a market-design argument rather than a fair-use one. If the price is small relative to the buyer's other costs, the question shifts from whether the buyer can pay to how the market should be structured.
The publishers' joint summary-stage brief makes a different case. It argues that OpenAI's training use of their archives actively harmed them in two ways: it siphoned referral traffic away from publishers' own sites, and it allowed OpenAI to flood the open web with AI-generated "pink slime" content that diluted the market for original reporting. The brief characterizes the harm as ongoing substitution, not a one-time copying event, a distinction that matters because fair-use doctrine typically treats transformative reuse more leniently than market-substitution claims.
The cost-asymmetry argument borrows from a real benchmark. Shah anchors his $20 billion ask to U.S. music-industry licensing, which he describes as a working market under $20 billion a year. The structural comparison is uneven, though. Music licensing was built around a narrow, auditable use (a song played on the radio, sampled in an ad) and a small set of rights holders. News licensing would have to cover a far broader set of uses (training, retrieval, summarization, output attribution) across models that ingest the same source material in different ways. The plumbing for that market does not exist yet.
The caveat on the numbers is the numbers themselves. Shah is the CEO of a named plaintiff, and his figures are calibrated to a market he wants the court to recognize, not one that already exists. The $750 billion figure is OpenAI's own investor communication about long-run compute spend, not a current revenue line, and the $20 billion is a suggested price tag, not a transaction. Both should be read as anchors in a negotiation, not as a settled market.
The next watch item is the briefing schedule. The court has set initial summary-stage filings for the consolidated multidistrict docket; a fair-use ruling at this stage is the next material event. A ruling against fair use would force OpenAI into settlement talks on the same terrain Shah is sketching. A ruling for fair use would put the question back to Congress, which is the outcome the DOJ is already asking for. Either way, the debate that starts with "can AI afford to pay?" is about to be replaced by a different one: what would a working market actually look like.