The content licensing market now has enough disclosed deals to compare, and the comparison produces an odd result. There are two prices for training material, they differ by an order of magnitude, and which one you pay depends less on what you used than on when you decided to ask.

The license price is visible. OpenAI has signed roughly two dozen publisher and data agreements, the largest reported at two hundred and fifty million dollars over five years with News Corp. Reddit disclosed two hundred and three million dollars in aggregate data licensing contract value in its listing documents. Perplexity took a different route entirely, running a revenue-share program for publishers whose material it cites, backed by a reported pool in the tens of millions.

Current guidance across the largest cloud and AI infrastructure providers puts combined 2026 capital spending somewhere around six hundred and sixty to seven hundred billion dollars, roughly double last year. Google in the region of two hundred billion, Amazon similar, Microsoft not far behind, Meta well over a hundred, Oracle around fifty. Tracked AI-attributable spending has gone from about thirty billion in 2020 to several hundred billion now.

Revenue at the pure-play model developers is growing fast and remains a fraction of the infrastructure being deployed on their behalf. This is the gap everyone points at, and the way people point at it is usually wrong.

Both halves of that sentence come from the same dataset, and publishers keep quoting only the half that suits the argument they are making that week.

Growth of two hundred percent is real. So is the base. Referrals from AI chat interfaces remain under one percent of all traffic to publisher sites, against a search channel that has fallen by roughly a third globally and by more in the United States. A channel tripling from nothing is still nothing. Tripling again gets it to three percent, which does not replace a thirty-eight percent decline in the channel that used to carry the business.

Lawmakers and lobbyists have started using generative tools to draft legislation. Reporting on the practice found that the tools are currently poor at the technical mechanics of federal law and regulation, which is an awkward finding, because the technical mechanics are most of what a bill is.

A statute is not an argument. It is a set of instructions for amending other statutes. The operative text of a typical bill consists of directions to strike a phrase in a specific subsection of an existing code title and insert different words in its place, plus definitions, effective dates, severability, and appropriations language. Getting it right requires knowing what the current code says, what pending bills would change about it, how courts have construed the terms of art, and which committee has jurisdiction over the resulting text.

Measurement across thousands of publisher sites put the year-over-year decline in Google search referrals at about sixty percent for small publishers, forty-seven percent for medium sites in the ten thousand to hundred thousand daily pageview range, and twenty-two percent for large ones. Globally, search traffic to publishers fell roughly a third in the year to November 2025, with American sites down closer to thirty-eight percent. Discover referrals fell about twenty-one percent.

A securities filing disclosed on Monday that Nvidia will provide up to one hundred and five billion dollars in financing for an OpenAI data center in Ohio. The facility starts at 4.25 gigawatts of compute with an option on a further 3.75. Nvidia supplies the hardware. SB Energy builds and operates the site at the PORTS-Pike campus in Pike County under a twenty-year lease to OpenAI. Capacity phases in from 2028.