The real AI race isn't about building the models. It's about controlling what feeds them.

Most people assume the race for artificial intelligence will be won by whoever builds the technically superior model.

I think that’s only part of the story.

The real contest is over what those models feed on.

Whoever owns the model, the interface and the data doesn’t just decide what a licence costs. They decide which sources appear in an answer, what gets cited, what becomes relevant — and they can influence which companies gain visibility and which fade from view. That’s the power at stake: not building artificial intelligence, but controlling what sustains it, and everything that follows from that.

And that resource is splitting into two layers.

On one side, infrastructure: content, data, distribution channels. It can be bought, sold, acquired. On the other, trust: the relationship with the people who read, listen, watch. It can’t be bought in an acquisition, and it can’t be trained on a dataset.

AI is concentrating the first in very few hands. And in doing so, it puts pressure on the second. This holds for print, for radio, for television — in different ways.

What’s happening to newspapers: the deals, the logic, and why it should concern everyone

According to the Wall Street Journal, Meta has put up to fifty million dollars a year on the table for News Corp’s content: a deal of at least three years, covering US and UK content, whose existence the two companies confirmed without disclosing its terms. Nearly four hundred American newspapers, meanwhile, accuse OpenAI and Microsoft of using their content without authorisation or compensation, and have taken them to court.

On this front, Italy’s two largest publishing groups have taken different routes. GEDI — publisher of Repubblica and La Stampa — signed with OpenAI: under the deal, its content enters ChatGPT with citations and links.

RCS chose another path: a technology partnership, not a content deal — using AI to build an assistant for Corriere della Sera, while keeping its content in-house.

GEDI is betting on visibility inside the AI systems. RCS is using AI to offer a proprietary service. Two different approaches, in the same market.

But what if the material we train these models on is decades of journalism? Years of authored work, political positions, information about readers, about mass preferences, about what consumers think.

What follows from that? And what if we add the day’s fresh news, which these engines can draw on directly?

What are content, historical archives, authors’ voices and audience data actually worth?

It’s already happening, and the picture is split in two. The Reuters Institute captures it in its 2026 report: licensing deals of various kinds are multiplying — model training, visibility inside generated answers, access to the technology — while lawsuits mount against AI developers accused of using content without authorisation.

A coalition of nearly four hundred publishers has sued OpenAI Inc. and Microsoft Corp., alleging that their content was used without permission or payment to train generative AI products such as ChatGPT and Microsoft Copilot.

The publishers argue they invested billions to protect their content — including behind paywalls — and that this was rendered futile because, according to the complaint, the defendants appropriated it anyway. The suit contends that these generative AI products, made possible in part by publishers’ work, generated billions in market value for the two companies, while the publishers received nothing.

The case is the latest in a growing wave of litigation against AI developers, and probably the broadest legal action ever brought by local and regional newspapers.

The billion-and-a-half-dollar settlement between Anthropic and a group of authors likely encouraged these moves: it set a precedent that many read closely.

Local newspapers remain the most trusted source of information for Americans, far more than national outlets or social media. And it’s local journalism — the most fragile democratic institution of all — that finds itself on the front line of this case.

Chartbeat data cited by the Reuters Institute shows Google traffic falling by a third globally between November 2024 and November 2025, across more than 2,500 websites. The publishers surveyed expect search-engine traffic to collapse by more than 40% over the next three years.

Another striking figure from the same report: among 280 editorial executives surveyed across 51 countries, only one in five expects these deals to become a substantial source of revenue — and those are mostly premium titles. Half expect a marginal contribution. And another fifth — local publishers, public service media, small countries — expect nothing at all.

So why do those who sign, sign?

Put crudely: for the money, and to avoid disappearing. To stay on the radar of the places where people now go to inform themselves — ChatGPT, Claude, Gemini, Copilot.

What’s at stake is the management of information, and the distribution of power

Whoever controls the interface through which millions of people receive answers inevitably acquires a new kind of editorial power: deciding which sources are shown, which are cited, and which remain invisible.

A newspaper archive isn’t a warehouse of text. It’s half a century of editorial judgement: what interests people, how a fact must be told to be understood, which language persuades and which repels. Whoever trains a model on that material isn’t buying words. They’re buying the craft of speaking to an audience, distilled at industrial scale.

Then there is the data that never comes back. ChatGPT has passed 800 million weekly users, and OpenAI’s filings ahead of its listing already speak of more than 900 million: the largest observatory ever built on what people want to know and what convinces them. In a typical deal, the publisher hands over the content and gets paid. The platform keeps every scrap of interaction data — including how people consume that very publisher’s work.

This is editorial power on a scale no newspaper editor has ever held, without the ethical and moral responsibilities we demand of an editor.

What counterweights exist — and who guarantees genuine freedom of information?

Radio: AI clones the voices

Not just newspapers. Radio, which looks like a digression, sits at the centre of the AI question. Because here AI doesn’t buy archives. It replaces voices.

“AI Ashley”, “DJ Tori” — entirely synthetic presenters, on air on real stations. The CEO of iHeartMedia told investors that AI could radically change the industry’s cost structure.

And there’s a second pressure, different but pushing the same way: presenters who trained for the job are giving way to those who bring an audience — influencers, public figures. It’s happening in Italy too. Reach in place of craft.

Two pressures compressing the same thing: the voice you know. The one that’s there, live, when the building shakes and tells you what’s happening.

World Radio Day 2026 had a single theme: Radio and Artificial Intelligence. With a slogan that reaches far beyond radio — AI is a tool, not a voice. And a coda: technology alone doesn’t build trust. The people who speak do.

You can clone a voice. You can’t clone the relationship that voice has built. Or can you? The answer depends on how sophisticated the models become — and it’s the question the next decade will be decided on.

Television: buying the infrastructure, then building AI on top of it

While publishers weigh whether and how to enter the world of AI, others are buying entire infrastructures and merging worlds.

In July 2026, Sky announced the acquisition of ITV’s media and broadcast division — the channels and the ITVX streaming platform, not the production arm — for 1.6 billion pounds.

Same principle, seen from the buyer’s side: in the AI economy, advantage comes from combining content, distribution and data. Whoever holds them together sets the terms; whoever gives them away one piece at a time ends up working for someone else.

And there are those who already own the infrastructure and are building the machine on top of it. In June 2026, Warner Bros. Discovery announced, with Amazon Web Services, an agentic advertising platform: AI agents that plan, optimise and sell inventory across linear and digital television. WBD’s library no longer exists only to be watched. It trains the machine that sells advertising.

This isn’t an isolated case. In the consolidation now under way, the prize isn’t the studios — it’s the content libraries. Because a library today is two things at once: a catalogue to show, and a dataset to feed the algorithms.

The invisible wholesaler paradox

The way we look for information is changing fast. From keywords on Google, to full phrases, to asking an AI directly. No longer only SEO, but increasingly GEO and AEO — I’ve written about it here.

For anyone who produces content, this means a deeper dilution: you create something of value, it’s fed to a model, and the interface that summarises it erases you from the scene. The user gets the answer they came for. But they never see your site.

I call it the invisible wholesaler paradox: you supply the market, but nobody knows the goods are yours.

This is why the things no interface can intermediate become priceless. Attribution: your name inside the answer, not buried in a footnote. The bylines people search for. Newsletters, events, communities. A contract that pays you well today but makes you invisible tomorrow isn’t revenue. It’s liquidation.

Two games

In the end, among many, two great games are being played.

One is played to prevail in the world of AI: it’s won by whoever owns the most advanced model, the simplest interface, the most useful data. Whoever has that holds the power to decide which sources appear, what gets cited, which companies become relevant and which vanish without anyone noticing.

The other is the game of visibility on the web, which increasingly means visibility inside AI-generated answers. It’s played by optimising your site, by rethinking strategies, formats, timing and channels. But it’s won by building and tending relationships of trust — because whatever can be summarised will be summarised, and then forgotten. What someone searches for by name will not.

This holds for newspapers. It holds for businesses.

Perhaps, some years from now, we won’t remember this moment as the one in which new AI models were born.

We’ll remember it as the moment in which control was redrawn — over the information, the data and the relationships from which those models draw their value.

Which is why deals, acquisitions and lawsuits are not isolated episodes. They are the signals of a far larger transformation.