The A.I. Industry Is Booming. When Will It Actually Make Money?

Keeping track of all the developments in the A.I. industry can be hard, but it’s helpful to bear in mind two basic questions: What’s the potential size of the market for A.I. products? And who will profit from it? In a recent lecture at Stanford University, where he is teaching a course on the economics of A.I., Apoorv Agrawal, a venture capitalist at Altimeter Capital, said that Alphabet, through its Chrome browser, Android operating system, and YouTube video-sharing service, reaches about four billion people worldwide and generates, mainly through advertising, revenues of about a hundred dollars per user per year. Meta, through Facebook, Instagram, Messenger, and WhatsApp, reaches about 3.5 billion people, and generates about seventy dollars per user per year, Agrawal estimates. OpenAI recently said that it has more than nine hundred million weekly active users, but, according to Agrawal’s calculations, ChatGPT only yields about ten dollars per user per year. In an e-mail, Agrawal told me that he thinks of himself as very much an optimist on A.I.’s future. Referring to users of the technology, he told his students, “The question is how do we get the one billion up to four billion? I’m not sure knowledge work is the answer.” He continued, “The second question is: How do we get the ten dollars per user per year up from ten to one hundred? And I’m not sure subscription is the answer.”

The vast majority of ChatGPT users employ the free version. Agrawal suggested that advertising could eventually provide OpenAI and its rivals with a big revenue stream and help “unlock” the A.I. economic model. Since A.I. firms gather a lot of information about their users, they should be able to target ads effectively. Alphabet has incorporated A.I. into its Google search engine. When it released its latest quarterly results last week, its C.E.O., Sundar Pichai, said that revenues in this part of the business rose nineteen per cent and search queries hit an all-time high, “with A.I. experiences driving usage.”

The other big question is how competitive the A.I. industry will be in the long run. Size doesn’t guarantee riches. The retail sector is a huge industry, but it’s fiercely competitive and has small profit margins. Ditto food and travel. Traditionally, the tech sector has been controlled by a few very large firms that generate big margins. Increasing returns to scale and network effects could well push A.I. in the same direction. If Google ended up dominating consumer A.I. and Anthropic ended up dominating business A.I., the industry would resemble other digital markets that have tipped into monopolies or oligopolies, like search and social media. But what, then, would be left for other players, like OpenAI and Meta, and all the infrastructure they are building?

Right now, there’s no shortage of competition. After Anthropic’s success in attracting business users, OpenAI is beefing up its Codex coding agent and other work tools. Last month, Denise Dresser, the firm’s chief revenue officer, said that it had nine million paying business users, and payments from businesses accounted for more than forty per cent of its revenues. Meta, xAI and others are also investing heavily in their models. There are also countless A.I. startups looking to establish a niche, and cheaper open-source models, like OpenAI’s own GPT-OSS or the DeepSeek-R1, future iterations of which could gain traction in some business realms and territories.

In short, the long-term outcome is far from certain. And at a moment when the A.I.-driven stock market is breaking through the stratosphere, it’s perhaps worth looking at some cautionary tales from the past. In an article from 2src21, the economist and venture capitalist Bill Janeway used the term “productive bubbles” to describe periods like the railway and dot-com boom-busts, which bequeathed to future generations invaluable productive assets, such as rail networks, power grids, and fibre-optic networks. (Unproductive bubbles include tulip mania in seventeenth-century Amsterdam, and the meme-stock phenomenon of 2src21.) Last year, in another piece, Janeway argued that generative A.I. could be the latest productive bubble, and when I corresponded with him last week, I found that recent developments hadn’t changed his thinking much. He pointed out that following the Great Panic of 1873, more than a hundred and twenty railroads defaulted on their bonds. After the dot-com bust of 2srcsrcsrc-src1, countless internet companies went under. So did WorldCom and Global Crossing, two big providers of telecom infrastructure, Janeway reminded me. “Why will it be different this time?” he wrote. “I get that Microsoft and Alphabet have enormous flows of monopoly rents from their established business, but OpenAI and Anthropic as well as CoreWeave et al?”