In 1986 Howard Moskowitz helped to make $600 million by proving that there is no single perfect spaghetti sauce. In 2026 Sam Altman believes the same about AI: it doesn't have a singular purpose, singular killer application, singular winning formula. His investors may disagree.

The blank text box

We are accustomed to a clean text box interface of AI chatbots like ChatGPT. But it was not a design choice. It was an admission of ignorance. In November 2022, when OpenAI released ChatGPT, nobody at the company knew which use case would win. A blank input field was the only honest interface for a technology without a known purpose.

Today, 900 million people use ChatGPT regularly. That number is easy to say and hard to grasp. It means that roughly one in eight humans on earth opened the same app in the past seven days. They used it to write emails, explain medical test results, help their children with homework, summarize legal contracts, generate birthday card text, debug Python code, and ask questions they were too embarrassed to type into Google. The range of uses is so wide that it resists any single description. And the interface that made this possible is a blank text box. You type whatever you want. The machine tries to help.

45 sauces

Altman took this ubiquity and turned it into strategy. Through 2022-2025, OpenAI maintained an attitude of research laboratory. They were throwing things at the wall to see what would stick, and given that nobody knew where AI belonged, this was the right call. You cannot focus when you do not know what to focus on. Altman called the approach "betting on a series of startups." In the last few years, the company launched a video generator called Sora, a web browser called Atlas, a hardware project with Jony Ive, shopping features inside ChatGPT, and an erotica mode.

The logic behind this strategy has a name and a history. In 1986, Campbell Soup hired a food scientist named Howard Moskowitz to fix its struggling Prego spaghetti sauce. The company wanted him to find the one perfect recipe. Moskowitz tested forty-five varieties on thousands of people across the country. The data refused to converge on a single winner. Instead it clustered into three groups: people who liked their sauce plain, people who liked it spicy, and people who liked it chunky. Nobody in the industry had ever sold chunky pasta sauce. Moskowitz told Prego to sell all three. The chunky line alone earned $600 million over the following decade. Malcolm Gladwell turned this story into one of the most watched TED talks in history and drew a simple conclusion: the food industry had spent years searching for one perfect product when the answer was to stop searching and sell several. There are no perfect sauces. There are perfect sauces, plural.

Altman's OpenAI is running the Moskowitz experiment on artificial intelligence: try everything, measure what clusters.

The chunky sauce

Then Anthropic found "the chunky sauce" – a popular and heavily used application of AI as a coding agent – and everyone forgot the lesson.

While OpenAI scattered its chips across a dozen projects, Anthropic did one thing well. It built Claude Code, sold it to enterprises, and ignored everything else: images, audio, video, hardware. It just kept making the text model better for people who write software. Results came fast. In mid-2025, Anthropic was at $4 billion annualized sales; by early 2026 it reached $19 billion. Claude Code alone hit $2.5 billion. Among US companies tracked by Ramp, Anthropic's share of enterprise AI spending went from 10% to over 65% in about a year.

Hemingway wrote that things happen gradually, then suddenly. People watched Anthropic's numbers and concluded that the "suddenly" had arrived: coding and enterprise is the answer, the one perfect recipe, and the search is over.

Some of OpenAI's new top managers agreed. On March 16, Fidji Simo, OpenAI's CEO of applications (she came from Meta), told staff that the company would drop its "side quests." Her words: "We cannot miss this moment because we are distracted by side quests. We are very much acting as if it's a code red." CFO Sarah Friar, who came from Nextdoor, backed the pivot. Both are managers, not founders, and they did what managers do when a competitor gains ground: cut experiments and double down on what works today.

Eight days later, Sora was cancelled. The app had launched six months earlier, peaked at a million users, then collapsed to under 500,000 while burning $1 million a day in compute. Disney had signed a $1 billion deal to bring two hundred characters to Sora, and found out about the shutdown less than an hour before the public did. The deal collapsed with it. Altman called the new Disney CEO to say he felt "terrible." But compute is compute, and Sora was eating chips that Codex needed.

So far, this looks like Hemingway's "suddenly." A company finds focus, kills its distractions, pivots to what works. You could teach it at HBS.

The organism rejects the transplant

Except the company didn't agree with itself.

On April 1, Brad Lightcap sat down for a podcast interview. Lightcap is OpenAI's COO, the first business hire, the person who joined when the company had forty people and no product. The host was Jack Altman, Sam's brother. And Lightcap spent the conversation talking about what excites him: the diversity of uses, how many different jobs the tool fits, the breadth of what people build with it. Two weeks after the "no side quests" memo, the COO was on tape celebrating side quests.

That same day, Altman sat with Laurie Segall on her podcast "Mostly Human." Segall asked about the biggest opportunity ahead. His answer had zero to do with enterprise sales or coding tools. He called it "this sort of cluster of automating researchers and companies and the super personal assistant for people's life." Three bets packed into one sentence. Not the enterprise supplier that Simo had described to the staff. Altman still thinks like Moskowitz: he is not looking for one recipe, he is building a shelf.

The next day, OpenAI bought TBPN, a daily tech talk show with eleven employees and about 70,000 viewers per episode, for a reported "low hundreds of millions." The show will report to Chris Lehane, OpenAI's chief of global affairs. Quite in irony: a company that just killed a video app to free up compute spent hundreds of millions on a media property. CNN compared it to RCA creating NBC in 1926 to sell radios. Jessica Lessin of The Information put it simply: Musk has X, and now Altman has TBPN.

Still in the gradually

Meanwhile, on March 31, investors handed OpenAI $122 billion at an $852 billion valuation. They are not pricing a business. They are pricing optionality: the bet that nobody knows where AI lands, and the company with 900 million people poking at a text box every week will find out first.

The people who watched Anthropic's numbers and concluded "coding and enterprise, that's the answer" could be making the same mistake the food industry made before Moskowitz showed up. They found one sauce that sells and decided it was the only sauce. Moskowitz tested forty-five varieties before the data told him the truth. We have tested maybe four or five so far. Coding works, chat works, search sort of works, image generation remains a toy, and video generation just got killed. That leaves a lot of shelf space empty.

Hemingway's "suddenly" has not arrived. We are still in the "gradually." Simo and Friar look at Anthropic and saw a reason to focus. Altman looks at the same numbers and saw one recipe out of forty-five.