OpenAI’s First-Mover Disadvantage
My colleague Perry was clutching a Ben & Jerry’s ice-cream bar. Our office, which made websites at the turn of the millennium, had decided to try Kozmo.com, a new site that promised to bring DVDs, books, and treats to your door within an hour. We paid a couple of dollars for the ice-cream bar—delivery was free. Yet Perry’s eyes opened wide, the way they did when something bad was about to happen.
Kozmo.com collapsed a year later, in April 2001, along with much of the rest of the dot-com economy that employed us at the time. Today, of course, the Kozmo.com model is everywhere—DoorDash, Uber Eats, Grubhub, and other services allow you to order almost any foodstuff to your door. Kozmo.com wasn’t a bad idea so much as a badly timed one. The marketplace wasn’t mature, and the business model for delivery hadn’t been established.
The technology industry believes in the gospel of original ideas and disruptive innovation. But Kozmo.com, just one of many examples, shows how likely the first mover in a sector is to founder or fail or fall out of favor: Friendster and Myspace were first in social networking, but Facebook eradicated them. Treo and BlackBerry owned the smartphone market until the iPhone stamped them out. Early winners seem obvious and enduring, until they do not.
Will the AI market be any different? No and yes. A year ago, OpenAI’s ChatGPT was a generic name, the Coke or Kleenex of generative-AI chatbots. Today, its competitors, especially Anthropic’s Claude, are advancing quickly. OpenAI’s fall from favor looks inevitable.
But AI appears to work nothing like delivery apps, smartphones, social networks, or even computer operating systems. If ChatGPT becomes outmoded, it won’t be the result of OpenAI losing ground or failing to innovate. Instead, the entire generative-AI sector will have become a commodity, like soft drinks or facial tissues. That process has already begun.
The logic of first-mover advantage is essentially colonialist: a land grab across a newly discovered commercial wilderness. In the early days of American industry, some of that wilderness really was untamed. DuPont ushered in the modern synthetic-fiber industry with nylon; Bell Labs invented the transistor, the foundation of modern electronics.
A more boring path to success, made from standards, licensing, and distribution instead of frontier-crossing invention, proliferated in the 20th century. In the early 1980s, Microsoft pushed MS-DOS (and then Windows) onto various PC manufacturers to get the software into as many machines as possible. JVC’s VHS cassette beat Sony’s (in many ways superior) Betamax format because JVC licensed its format faster and more broadly, meaning more VCRs used the format and more video stores stocked it. In the early days of cable television, channels such as CNN and MTV sought to be carried on as many cable systems as possible. Retailers such as Walmart started saturating regions with stores, helping solidify supplier and distribution networks while slowly eroding competition from local businesses.
By the 1990s, lessons from Microsoft, JVC, CNN, and others suggested that network effects—increased value derived from the number of users of a product or service—were central to land-grab victories. At the same time, technology shifted economic growth toward immaterial goods, such as software and online services. If a platform could scale fast and first, it could own the market. This rationale drove venture-capital investment during the entire internet era, from the dot-coms through Web 2.0 and the smartphone. User growth, market share, ecosystem lock-in, and brand awareness became more important than profit or even revenue.
But success didn’t always come from being first. In many cases, what we now consider the big winner became the most familiar only after the market was established. Google was not the first search engine—Lycos, AltaVista, Ask Jeeves, and others filled that role years earlier. Google invented a better way to catalog the web, and then it built a viable advertising model atop that success. Facebook did the same with social networking; Apple did the same with smartphones. Zimride, which later rebranded as Lyft, offered ride-share services first, but Uber became dominant by expanding aggressively. Reversals of early success turn out to be far more common than first-to-market entrenchment.
Given that history, OpenAI might seem more likely to cede ground than to retain it. The company’s apparent advantage has turned out to be more fragile than expected. (The Atlantic entered into a corporate partnership with OpenAI in 2024.) But something else has changed too. Just as the online-services market differed from the electronics market that preceded it, so too does the market for AI services differ from what came just before.
A single AI victor may never emerge. Different AI companies are proving to be good at different things. Anthropic’s Claude, for example, is particularly effective at coding and at analyzing long documents. Sophisticated AI end users are using multiple services for different tasks. The various models—ChatGPT, Claude, Gemini, and so on—tend to trade places rapidly on various metrics of performance.
ChatGPT won over many ordinary consumers—college students, say, or your aging parents. But workplaces, schools, and organizations adopt technology as a part of interoperating software suites such as Microsoft Office and Google Workplace. And unlike Uber and Amazon, AI has quickly become embedded in those. The network effects that lock users into an AI service are situated inside these software packages, up the value chain from GPT, Gemini, Claude, and other AI models. The AI is a (supposedly!) thinking utility that powers old business services in new ways, but it isn’t the service—or won’t remain so, at least.
[Read: Even Silicon Valley says that AI is a bubble]
At work, AI is not a choice but a tool—welcome or not—to which employers subscribe that becomes a default. Initially, ChatGPT did the hard work of educating users on what large language models are and how they can be used. Now many consumer and business users understand what AI is and what it can do, and they have moved on to trying to figure out how to use it. Which AI they use has become less important—and perhaps less of a choice, besides.
The technology of LLMs feels disruptive and world-changing—more like nylon or iPhones than like Kozmo.com. But the big AI companies’ models have reached parity fairly quickly. Switching costs are fairly low in the AI marketplace, compared with those of videocassette recorders, smartphones, or even social networks. Whether or not AI is ethical, effective, or even fit for purpose, the act of using it appears to be naturally promiscuous.
Microsoft, Google, Facebook, Uber, and other platform-centric companies relied on winning users and holding them hostage. AI doesn’t seem to work that way. Instead, it looks more like corporate infrastructure or commodities. Companies don’t really care which cloud-computing provider they use, so long as the service is reliable and competitively priced. Firms will procure whatever hard disks, printer paper, or lunchtime slop bowls are convenient, suitable, and priced right. They seem likely to end up doing the same with AI.
OpenAI CEO Sam Altman made the provocative statement last week that in the future, intelligence will be “a utility, like electricity or water.” Instead of taking that claim as hubristic—as Altman claiming that smarts will be owned by OpenAI—consider a far more mundane and probable idea: AI could become, in just a few years, a commodity as invisible and anonymous as power or plumbing. Nobody cares what company makes the lights work or the toilets flush, so long as they do.