Blog

By Christine Ji and Britney Nguyen

OpenAI sits at the center of the AI boom that it started. But it faces increasing challenges around both the business and the technology itself.

Excitement around recent Google and Anthropic AI offerings is marring the celebration of ChatGPT’s third birthday.

OpenAI’s launch of ChatGPT on November 30, 2022 was the starting gun for the artificial-intelligence arms race that has lifted up the stock market and unleashed billions of dollars of investment.

Whether OpenAI can preserve its leading position going forward is now the big question.

Within five days of its release, ChatGPT had achieved 1 million users. Three years later, the latest version of ChatGPT – with real-time access to the web, enhanced reasoning, coding skills and image- and video- generation capabilities – makes its initial features appear rudimentary. In September, OpenAI released its Sora video-generation app, which also achieved 1 million users in five days.

But the talk around OpenAI isn’t as exuberant now as it was when ChatGPT first burst on the scene. Competing models from Google and Anthropic have won praise for their technological capabilities, all while OpenAI’s financial commitments have started to worry Wall Street. Some technologists also wonder if large language models will ever have the potential to unlock the most futuristic AI visions.

And in some cases, things haven’t changed much. Gary Marcus, an emeritus professor at New York University and prominent psychologist known for his research on AI, told MarketWatch that he raised issues about OpenAI’s systems in late 2022 that still haven’t been solved. “They hallucinate and have trouble with reasoning,” he said.

The competitive scene

Last December, Marcus also predicted that without a truly differentiated product by the end of 2025, OpenAI would risk losing its moat.

“And now, they have a competitor with a product that’s more or less just as good,” Marcus added, referring to Alphabet’s (GOOGL) (GOOG) Google Gemini 3 AI model, which debuted earlier this month.

The AI space has seen “an explosion of activity” in the past few years, Benjamin Lee, a professor at the University of Pennsylvania’s School of Engineering, told MarketWatch. “We’re seeing a lot of experimentation and adoption by individual users and others.”

Gemini, which many thought had been left for dead at the beginning of the year, has made aggressive strides to increase its market share. Earlier this month, Google’s newest Gemini 3 and Nano Banana Pro updates further impressed investors and users alike.

“Google was fairly far behind, and I think people counted them out, but their most recent model is arguably ahead of GPT-5,” Marcus said.

Gemini 3 outperformed GPT-5 on a majority of key AI benchmarks, or standardized tests used to evaluate models. And Anthropic’s latest model, Claude Opus 4.5, beat both GPT-5 and Gemini 3 on agentic coding benchmarks upon its release last week.

“This is really healthy and, from a technical perspective, super exciting to see the robust competition between these types of models,” Lee said. But for OpenAI, the competitive playing field means it can’t coast on its first-mover advantage.

Also read: How can Anthropic stand out in the AI wars? I went to a Greenwich Village pop-up to find out.

A market linchpin

Concerns about an AI bubble have boiled over this month, leading to a selloff that’s hit tech stocks hard.

OpenAI has come under scrutiny for its central role in a complex web of AI financing. The company has inked over $1.4 trillion in AI infrastructure deals, and investors are skeptical about how the AI lab will be able to pay for all of its commitments.

New features such as OpenAI’s ChatGPT Pulse, Sora and web browser could help scale its business, but AI monetization is less of a concern for Google, which has business segments spanning search, cloud computing, Android and enterprise software. Additionally, these verticals facilitate the distribution and scale of Google’s AI products.

Google’s custom-chip business also gives it a major leg up in the AI race, enabling the company to save on infrastructure costs and opening up another revenue stream in the form of equipment rentals. The tech giant uses its tensor processing units (TPUs) to train its Gemini models. The chips have also been used to power its search and YouTube algorithms.

In October, Anthropic said it plans to use up to 1 million of Google’s TPUs as it seeks to scale up computing – an issue also plaguing OpenAI as AI models become larger and more advanced.

Read: These two ‘Magnificent Seven’ stocks could be the strongest survivors of an AI apocalypse

Are LLMs the path to AGI?

As the chatbot wars rage on, a growing number of AI researchers are questioning if large language models will even be the future of the technology. After all, hallucinations and reasoning gaps aren’t issues unique to ChatGPT. Wall Street shares that worry, as investors question the return on investment of chatbots that still lack the power to fully automate corporate workflows.

LLMs operate by predicting the next word, or token, in a sequence based on statistical probability. They don’t actually “know” facts or understand logic.

“They’re not really abstracting away a stable comprehension of the world,” Marcus said of LLMs. As AI technology trends toward autonomous vehicles and robots, Marcus and others believe the new frontier will be “world models,” or AI with a mental simulation of the real world.

Sergey Gorbunov, a technologist and co-founder of blockchain-infrastructure platform Axelar, also sees world models as perhaps a better path toward artificial general intelligence, or AGI, which is the point at which AI models will be believed to be as intelligent as humans. Admittedly, ideas about what AGI will look like have changed in the three years since OpenAI launched ChatGPT.

Unlike LLMs, world models “interact with physical spaces,” and therefore can “understand a little bit what’s happening in physics, not just in text,” Gorbunov told MarketWatch. For example, world models could help improve self-driving cars, he said, because an autonomous vehicle would be able to predict seconds ahead what another car will do.

Earlier this month, Gorbunov outlined in a blog post “two fundamental limitations” for LLMs: the reliance on preexisting data, and the fact that LLMs essentially navigate probabilities.

“If you look at the math of these models or how they’re constructed, they’re just predictable probability distributions,” Gorbunov told MarketWatch. “There is no sense of artificial intelligence anywhere there.”

What’s next for ChatGPT, and AI in general?

Three years after ChatGPT’s release, Vasant Dhar, a professor at New York University’s Stern School of Business and the author of “Thinking With Machines,” told MarketWatch that the fundamental way AI has changed is that it’s become better at simulating understanding, and is therefore more relatable to people.

ChatGPT’s progress in three years is “astounding,” Dhar said, even as its benefits to business and other aspects of life remain to be seen. He emphasized that the adoption of general-purpose technologies such as electricity and the internet historically took years to unfold.

“Our expectations have gone up so much that we expect this capability to be realized instantly,” Dhar said. “There is so much capability in AI at the moment that is still to be realized.”

While large language models like the ones powering ChatGPT have been at the forefront of the current AI boom, Dhar said he’s seeing incremental research into vision models and other types of models that will be able to integrate more human senses.

In the next year or two, Dhar said he expects to see improvements to ChatGPT and Gemini that, despite seeming incremental, will be “actually very significant” because of how they impact the lives of people using them.

But world models are still far ahead of where the AI industry currently is, Gorbunov told MarketWatch. In the coming months, he expects to see a battle on the front end of user interaction with the web through AI-powered web browsers such as OpenAI’s Atlas and Perplexity’s Comet. Google is also part of that fight, he said, as the dominant search engine incorporates more AI features into its platform.

“I think whoever is going to win that [user experience] in some sense will be able to capture a lot of traffic going forward,” Gorbunov said.

-Christine Ji -Britney Nguyen

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

  (END) Dow Jones Newswires

  11-29-25 0800ET
Copyright (c) 2025 Dow Jones & Company, Inc.

Over the last three years, I have been actively experimenting with AI through a course I created called “AI and Consultative Decision Making.” In parallel, I wrote the book Epic Disruptions, which involved conducting deep historical research into case studies of world-changing innovations ranging from gunpowder to Pampers disposable diapers.

One of the themes that emerged from my research is that disruptive change is predictably unpredictable. There are broad patterns, but because there are humans and complex systems involved, there are unexpected twists and turns in every story.

As the saying goes, history may not repeat, but it certainly rhymes. There are five historical lessons that seem pertinent to how AI could—or could not—drive epic disruptive change.

1. Disruption often starts in unexpected places

In the 1940s, Walter Bradeen, John Brattain, and William Shockley from Bell Labs developed a new technology called the transistor. The intent of their research effort was to develop a technology to replace vacuum tubes that powered communications networks. The transistor had clear benefits. It was small, rugged, and didn’t give off heat. However, early versions were also unreliable and required rearchitecting systems. 

It took decades for transistors to make it into communications networks. The first commercial market was hearing aids. The transistor fit perfectly in the market. Hearing aids were relatively simple, making it easy to incorporate transistors. Vacuum tubes gave off heat, which made battery packs affixed to a belt uncomfortable. Tubes burned out, making the total cost of owning a hearing aid expensive. The transistor-based hearing aid market exploded, supporting further technological development that ultimately ushered in the modern communications and computing age.

We naturally focus on the development and deployment of AI in large, sophisticated markets like the United States or Western Europe. However, one driver of ChatGPT’s rapid growth is usage in emerging markets that lack robust health and education infrastructures. Consumers don’t ask, “How does AI compare to a skilled teacher or clinician?”; they ask, “Is AI better than nothing at all?” History suggests carefully examining emerging market developments to spot disruptive change early. 

2. The secret sauce of disruption is a unique way to create, capture, and deliver value. 

When Mac and Dick McDonald first opened their restaurant, it was unremarkable. The path to disruption started when they shut the restaurant in 1948 and unveiled the “Speedee Service System” that simplified and standardized food production. When Ray Kroc became in essence the master franchisor of the concept in 1954, he and his team architected a unique system that involved close partnership with franchise owners. In the 1960s, Heny Sonneborn perfected a model that allowed the McDonald’s Corporation to profit through real estate. The unique way that McDonald’s created, delivered, and captured value—its business model—allowed it to serve billions profitably.

A unique business model is the secret sauce of disruptive innovation. It is what allowed Amazon.com, Google, and Netflix to emerge as powerhouses three decades ago. Unique business models provide funding for further improvement and frustrate incumbent response. 

Right now, leading labs like OpenAI and Anthropic are following business models that are neither novel nor difficult for technology companies like Amazon, Microsoft, or Google to follow. If the labs don’t develop unique ways to create, capture, and deliver value, history suggests they are likely to have finite lives as standalone providers.

3. Disruption is always messy in the middle.

In the 1920s, a battle broke out for the soul of the streets of many major US cities. Henry Ford had achieved his vision: the car for the “great multitudes.” Perfecting the assembly line brought the cost of Ford’s Model T from $30,000 (in today’s terms) in 1908 to $5,000. Sales soared. 

This was good for some, but less good for others. Cities were designed for people, not for cars. The sharp increase in automobile adoption spurred chaos and carnage. Newspaper cartoons in the 1920s often showed the Grim Reaper driving cars. One in the St. Louis Star showed a man kneeling holding up a platter of children to a car with a humanoid maniacal grin. In 1922 the mayor of Baltimore dedicated a 25-foot wood and plaster obelisk as a monument for the 130 children who died in motor accidents that year.

It is always messy in the middle of disruptive change. Getting out of the automotive’s middle required technologies such as traffic signals, regulations such as the need for drivers to have licenses, and norms, such as right-of-way at intersections.

Through this lens, a push to minimize rules and regulation is misguided as it elongates the time in AI’s messy middle and increases the odds of harm. Futurists Bob Johansen and Jamias Cascio note that it is hard to set precise rules in markets emerging as quickly as AI, so suggest the metaphor of a “bounce rope” in a wrestling ring. There are firm ring posts and boundaries at the edge of the ring, but those boundaries have slack and give in them.

4. There’s often a twist in the story

When Johannes Gutenberg and his team sought an early customer for the printing press, they naturally turned to the Catholic Church. The Church had real problems to solve, such as standardizing missals used for church services and shortening the three years it took to hand scribe a Bible. When Enea Silvio Piccolomini, who went on to become Pope Pius II, saw a Gutenberg Bible in 1454 he praised their “very neat and legible script” and noted how they could be read “without the use of glasses.” 

The Church didn’t foresee what happened next. The printing presses accelerated the ability for people like Martin Luther to spread ideas attacking the Church. A third of the books printed in Germany between 1518 and 1525 were from Luther. The printing press was a boon to some—scientists, revolutionaries, entrepreneurs who built businesses around it—and a curse to others: scribes, cardinals, and anyone else who profited from ignorance.

Management consulting companies have profited handsomely from AI-related work. In early 2024 Boston Consulting Group said that 20 percent of its revenues was AI-related. McKinsey touted how it was using its custom-created AI solution to boost its productivity and accelerate developing unique impact. What if, however, clients learn how to use AI in ways that obviate consultants? Or if AI reliance withered a consulting company’s ability to develop unique talent? Could the major consulting companies look at AI the same way the Church looked at the printing press?

5. It’s all about the people

Singapore’s DBS Bank is a remarkable story of transformation (detailed in my 2020 book Eat, Sleep, Innovate). In 2010, it was a laggard in its local market. In 2025, DBS was widely recognized for its nimbleness and digital prowess.

Its digital transformation involved key strategic shifts and major investments in technology. Those moves were necessary, but not sufficient. The critical unlock came from a set of behavioral interventions to help bankers use technologies in new ways. Paul Cobban, who was DBS’s Chief Data and Transformation Officer from 2009-2022 observed that without a systematic and structured approach to cultural change, adopting digital technologies would be akin to replacing memos with emails or emails with Slack messages. One of Cobban’s mantras was, “Nothing changes unless people’s behavior changes.”

The same is true of AI. Adoption is not a technological problem; it is a sociological and cultural one. Jim Wilson from Accenture estimates that for every dollar companies spend on technology, they should expect to spend six dollars on the human side of change.

* * *

One recurrent lesson that struck me during the research and writing of Epic Disruptions is how history provides a unique way to make sense of a complicated present. Disruption is predictably unpredictable, so AI will surely break from some of these patterns. However, the past provides a guide for where to look and what to look for to make sense of what will happen next.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.