Global Markets: The Atlantic and OpenAI

Unit 3, Professor Carrie Elkins, Global Markets: The Atlantic and OpenAI.

The last piece asked whether the silicon shield is a protection or a weak spot. This one swings the camera from wafers to a magazine. What gets taken for training here isn't production capacity — it's sentences.

Here goes — enjoy:

Case Summary

In March 2024, The Atlantic announced a two-year partnership with OpenAI. OpenAI got a licence to use the magazine's content to train its models and answer queries; the magazine got licensing fees, priority access to the technology, and a place inside the tech giant as a "premium news source." CEO Nicholas Thompson cared about more than the money: he saw a potential traffic channel, and a chance to "help shape the future of AI." The newsroom exploded. Staff wrote an open letter to Thompson and editor-in-chief Jeffrey Goldberg: AI cannot replace journalists, and no outlet that still wants to live up to this magazine's century-and-a-half of standards can allow that to happen. Editor Damon Beres called it, in the magazine's own pages, "a devil's bargain." Jessica Lessin, founder of The Information, used the same pages to argue that these deals amount to settling without suing — trading hard-won credibility for a bit of cash, with a counterparty that undervalues you while building products plainly meant to replace you.

From day one, the magazine was more an American idea than a good business. In 1857, a circle of abolitionist intellectuals — Emerson, Longfellow, Holmes — founded it in Boston to talk about slavery, and about the tensions pushing the nation towards civil war. It went on to publish Hemingway and Einstein, and early work by Mark Twain and Henry James. Former editor Cullen Murphy put it this way: it advances something you could call "the American idea." Culturally indispensable, it rarely won on the books. In the 1990s it could lose US$500,000 a year. In 1999 David Bradley bought it for US$10 million, and reportedly poured in more than US$100 million after that. The magazine still lost US$4–8 million a year. Only when he started running it like a startup — digital advertising rising from 9% to 45% of revenue, expansion into events and consulting — did it turn a US$1.8 million profit in 2010. His children didn't want the magazine. He began looking for an exit.

In 2017, Laurene Powell Jobs, Steve Jobs's widow, bought 70% through Emerson Collective for more than US$100 million. Her net worth at the time was about US$18.2 billion. Bradley told staff she had always been at the very top of his list: Emerson was already investing in serious journalism for its own sake. She described the magazine as a critical mission at a critical moment — equality, the American idea, American culture, and this wonderful and often messy experiment in democracy. The collective is named after Emerson. The organisation itself, though, is pure Silicon Valley: an LLC rather than a traditional foundation, with venture capital, activism and philanthropy at the same table. She promised not to touch the editorial side, while hiring heavily: headcount grew by about 30%, a paywall went up, the typography was redone, and the whole company was pushed to use technology more deeply. Bradley said at the time that The Atlantic's core strength was not operating at the technological frontier. She said: it should be. Expansion was expensive. Profits of roughly US$10 million a year when she took over turned back into losses. In March 2020, even as subscriptions rose, the magazine laid off 17% of its staff.

In late 2020, former Wired editor-in-chief Nick Thompson became CEO. His mandate: dig out of the hole and build a business that could sustain itself. He inherited a US$20 million deficit. He immediately raised subscription prices by 50% and launched a dynamic paywall that targets readers by their likelihood to convert. Goldberg moved the magazine away from daily news towards fewer, more deeply reported stories. By 2024 the magazine was profitable again, with more than 1 million subscribers and US$100 million in revenue, nearly two-thirds of it from subscriptions — up from less than half. Under Thompson it won three Pulitzers. The October 2024 issue, its cover almost free of headlines, won ASME's best cover the following year. One story led to the resignation of the then-head of CNN; one cover story helped exonerate a wrongly convicted man. In Goldberg's words:

We treat it almost as an ideology, an idealism — what we make is worth paying for.

Over the same stretch, OpenAI turned the tech industry upside down. ChatGPT launched in 2022 and passed 30 million users in two months. Microsoft's investment reached US$14 billion. Models need oceans of text. Believing that scale converts directly into performance, AI companies tested the limits of copyright and swallowed the internet whole. In one study, ChatGPT-4's hallucination rate approached 30%. In December 2023, The New York Times sued OpenAI and Microsoft: in Common Crawl, a "copy of the internet," Times content alone accounted for 16 million records. GPT-3 through GPT-4 were reportedly trained on 45 TB of data, roughly a 3.7-billion-page Word document. The Times demanded the models be destroyed and sought statutory damages of up to US$150,000 per article — for pieces that typically license for about US$10. OpenAI pleaded fair use and argued that no single source contributes meaningfully to training. By May 2025, some 40 copyright cases were pending. The rulings split: one judge held that training a competing product on another party's case summaries was not transformative; another held that Anthropic's use of 7 million books was. Anthropic still settled in September 2025 for US$1.5 billion, about US$3,000 per book — the largest settlement in US copyright history.

Thompson chose a different road from the Times. He acknowledged that the large models had plainly scraped without permission, and then did the maths: there's a whole range of ways to get paid, so put what the other side offers on the same scale as the cost of suing and the damages you might win. The exact terms were not disclosed. OpenAI was signing deals at a furious pace, and industry estimates of annual fees ran from several million to US$50 million — News Corp at US$250 million over five years, the Financial Times at US$5–10 million a year, Axel Springer at US$25–30 million over three years. By April 2025, OpenAI had partnered with 20 news publishers covering more than 160 outlets; across publishers and AI companies there were roughly 72 licensing deals in total. Thompson said the magazine would not become overly dependent on AI traffic; it would keep telling stories in the best way it could, and then distribute them as hard as possible. He also said: if content like The Atlantic's goes into these models, the world will be a better place.

Powell Jobs has never spoken publicly about the deal. She kept her promise not to interfere with editorial. When Goldberg was mistakenly added to a classified Signal group discussing war plans, she stood by her editor. But her interest in AI is public. Emerson has invested in several AI startups, and she played a key role in the roughly US$6.4 billion hardware partnership between OpenAI and Jony Ive, a longtime friend of the Jobs family. In 2025 she and Ive made a rare joint appearance. Ive said that innovation has unforeseen consequences, some wonderful, some harmful, and that he still feels responsible. She picked up the thread: we now know unambiguously that certain technologies have dark uses — but that was a sideways result.

The case stops at the table, just before renewal. Google's AI Overviews have already cut click-throughs to outside sites by more than 34%; chatbots have replaced search for more than 25% of Americans. The news industry is planning for "search zero." Yet The Atlantic is hiring: 30 editors in the first half of 2025, at salaries of US$200,000–300,000. Most of its peers are cutting. Is this just one of the few self-sustaining outlets hiring by the old logic, or has she seen another layer — that the people building models need writers? Trained only on synthetic data, models may collapse. Is the partnership between writers and AI companies the key to both sides surviving economically, or is it accelerating the replacement of journalism and pushing creativity towards something flatter?

The four questions in front of Carrie Elkins:

1. What is the ideas economy? How does it reshape the world business leaders operate in?
2. Why did Laurene Powell Jobs and Emerson Collective buy The Atlantic?
3. Should The Atlantic partner with OpenAI?
4. What would continuing the partnership mean for the magazine's journalism, and for the wider landscape of ideas?

My Thinking

The ideas economy puts a price on an original sentence

The ideas economy is not a new industry. It is the market that produces, distributes and prices original thought. Journalism, literature, science, argument — and everything a society uses to decide what is true, what is worth paying for and what is permitted — all live inside it.

It has three layers. Production is the work of scarce people: a magazine 168 years in the making, a novel thirty years in the writing, an investigation that can make someone resign and bring a wrongly convicted man home. Distribution used to be print, then the open web, then the search box. Now it is also a chatbot that can answer without ever sending the reader back to the source. Capture is who gets the money. For most of The Atlantic's life, capture was weak. In 2024 the income statement finally held up — US$100 million in revenue, nearly two-thirds from 1 million paying subscribers — and just then a new buyer appeared: model trainers who treat the archive as fuel.

Business leaders already live in this market. Regulation, talent, customers, even whether they are allowed to keep operating — all of it sits downstream of other people's sentences. Consultants who use models outperform those who don't overall, but the gains go mostly to those with low baselines, and the variance in team output gets flattened. The case cites a 2025 MIT study: ChatGPT users showed the lowest brain engagement and later slid into copy-and-paste. People who treat the model as a free intern are quietly thinning out their own judgement.

My own reading is that what's truly new in this Global Markets class is not that a magazine and a lab signed a deal. It's that the way ideas cross borders has changed. They used to travel as printed matter and as graduates. Now they also travel as tokens in a training run. Whoever sets the price of those tokens — licensing, litigation, or fair use marked down to zero — is writing the trade rules for knowledge work over the next decade.

What she bought was an American idea, not an income statement

The 2017 acquisition can't be explained by the US$1.8 million profit of 2010, nor refuted by the US$20 million hole that came later.

Bradley paid US$10 million in 1999, poured in more than US$100 million afterwards, and was still losing US$4–8 million a year before he got to that US$1.8 million. For someone then worth about US$18.2 billion, paying more than US$100 million for 70% was a stewardship cheque. Her words at the time: a critical mission at a critical moment. The collective is named after Emerson. The first page of the case is Emerson himself — the poets and inventors unlock our chains. Buying the magazine that Emerson's circle founded is the most literal way to fund that sentence.

She did two things that don't easily sit in the same room. She promised not to touch editorial, and kept that promise. At the same time she ran Emerson like a startup, told each unit to disrupt its own field, and, when Bradley said The Atlantic wasn't a tech company, answered: it should be. In 2020 her orders to Thompson changed again: dig out of the hole, build a company that can sustain itself.

A mission without a going concern is a subsidy. A going concern without a mission is just another website. She has always wanted to hold on to both. The OpenAI deal is where those two wishes can no longer live in separate rooms.

My own reading is that what she bought was never an IRR. She bought the right to have a voice at the level of ideas. The renewal question isn't whether this deal makes money. It's whether the voice is still there, and at what price it is being sold.

Sign — but don't renew on the same contract

The first deal is a pricing exercise, not a marriage.

Thompson's sentence is the right frame. The models plainly scraped without permission. There's a whole range of ways to get paid. Weigh what the other side offers against the cost of suing and what you might win. With the archive already inside GPT-3 and beyond, that is the only grown-up arithmetic.

The financial case for signing is clearer than the open letter admits. The Atlantic had only just climbed out of a US$20 million hole. The cluster in Exhibit 3 that looks most like it — the FT at US$5–10 million a year, Axel Springer at US$25–30 million over three years — amounts to 5–10% of annual revenue. Enough to pay for most of the 2025 hiring round. Litigation is not free. The Times spent US$10.8 million in the first year, and still has no ruling. A magazine that has only just turned profitable, with 1 million subscribers, shouldn't volunteer to be the test case that funds a decade of discovery.

The "don't sign" side is real too. A hallucination rate near 30% can stain a brand that has only just made "worth paying for" true. Cheap licences may be funding the very product that means readers never have to arrive. Google's summaries have already taken more than 34% of referral clicks. Sharper still is the owner's own position: Emerson's AI portfolio and Ive's hardware mean she has to be able to show that the renewal terms were negotiated as if she had no other chip on the table.

My own choice: renew, on terms written by the seller. Sell training and retrieval separately. Ban verbatim regurgitation; accept only clickable citations. If the Times or the next big settlement raises the floor, reopen the deal. Never let AI referrals become a significant acquisition channel. Keep suing those who don't pay — they later sued Cohere, which shows they can already tell a counterparty from a free rider. Take the owner's other bets out of the terms.

The cost of renewing blindly is the sentence Lessin has already written, in The Atlantic, to The Atlantic. The cost of walking away is spending money the way the Times does, without the Times's balance sheet.

Sell the past to pay for the next investigation

Goldberg has in fact already built the strategy along the way. They left daily news for fewer, deeper stories. Daily news is the first thing chatbots replace. Investigations that make someone resign, that bring a wrongly convicted man home, are what models can't originate and what subscribers will still pay for.

Licensing out the archive while thickening the newsroom is internally consistent. Sell the past; raise the price of the next original piece. Thirty editors in the first half of 2025 at US$200,000–300,000 is not a billionaire's mood of benevolence. It's a bet that human reporting is a scarce input. The question at the end of the case and this bet are the same sentence — whether labs need writers to keep their models from collapsing.

The medium-term risks are on the same page. If models answer with a smooth paraphrase and a dead link, subscribers won't come. Incentives inside the magazine will quietly tilt towards getting cited by models: shorter explainers, fewer year-long investigations. The hiring wave becomes a content farm with nicer furniture. A hallucinated "Atlantic" claim isn't a traffic problem; it's a reputational problem dressed in the magazine's own typeface.

My own leaning: the money has to go back to the writers for this deal to be worthy of the reason she bought the magazine. A quiet, permanent feed for the models is not.

Models will flatten the marketplace of ideas — unless someone can say no

By April 2025, OpenAI had 20 publishers covering more than 160 outlets. Across the industry there were roughly 72 licensing deals. That sounds like a market. It's only a market for those who can afford a Thompson.

The passage in the case on novelist Min Jin Lee is the other half. She spent thirty years writing books, gave up a legal career, and then discovered that models had stolen her stories. If the going rate is a seven-figure cheque for the mastheads and zero for everyone else, there will only be fewer of the next books that take thirty years to write. Adobe's verifiable attribution on creative work helps with credit. It doesn't create a living.

Creativity research is the layer business leaders will feel even if they never buy a magazine. AI beats all but the very best humans at divergent thinking. Stories written with it help weaker writers more, yet as a group they come out more alike. Consultants get better together, and more alike together. ChatGPT users get lazier. The Four C model separates everyday little-c creativity from the steam-engine kind of Big-C creativity. The former will be automated and homogenised. The latter will become more valuable, and harder to sustain — if the distribution layer no longer sends readers or royalties back.

My own reading is that Powell Jobs doesn't owe the labs a cheap corpus. She owes the next Emerson a world in which new ideas still have someone who pays for them. Renewing without terms betrays that sentence. Walking away, and letting fair use stay marked at zero for everyone smaller than the Times, betrays it too. Someone who sits in both rooms at once has to set a price the other rooms can copy.

Takeaways for Arcus42

Reading this case, what comes to mind isn't one magazine's renewal, but any company that writes judgement into systems: the archive is already within a model's reach, and the pricing rules don't exist yet.

First: once knowledge is in the system, the live question is licensing, not purity. Meeting notes, family archives, proposal drafts — all of it will become trainable text. Training, retrieval and generating sentences for clients to read are three different things; don't fold them into one checkbox.

Second: people own the judgement, models handle the rest — and the rest can't face the client directly. The Atlantic used its paywall to protect those 1 million subscribers. What a family office has to protect is that layer in the room. The model can draft. What the family is still buying is a judgement a person is willing to sign.

Third: don't hand distribution to a platform a second time. Publishers first handed their readers to search, then watched summaries take more than 30% of their clicks. Internal systems can be queried. They can't become the place where a client "discovers they have a problem."

Fourth: whoever sits on both sides owes the weaker side a process. An owner can love technology and also love an institution of ideas. When both sit in the same chequebook, a renewal can't run on vibes. Recusal, disclosure and an independent person reviewing the terms are cheaper than explaining afterwards.

Fifth: the new buyer's money has to go back to the scarce people. The deal only holds up because The Atlantic turned the partnership fees into a round of well-paid editors. If platforms pay for process knowledge, the fees should flow back into the quality of judgement, not into a leaner team.

In Closing

Before class, the line I agree with most is: "Sign, and renew — but not on the same contract."