Three years ago on Sunday, OpenAI released ChatGPT and accidentally started an arms race. Chatbot hit million users in five daysforced Google to declare a “code red” and convinced Silicon Valley that artificial general intelligence was just around the corner, and whoever got there first would win big. Microsoft has invested $10 billion in OpenAI. Google made every effort to free Bard. Chinese tech giants have rushed to create their own chatbots.
It looked like OpenAI had lost the race. Now that ChatGPT's third birthday is approaching, the lead has changed hands so many times that the race is starting to resemble musical chairs.
This year alone, the leader changed hands several times.
OpenAI entered 2025 as the undisputed leader. By the end of January it was already under pressure. Chinese startup DeepSeek released models matching OpenAI performance for a fraction of the cost, sending Nvidia shares plummeting and raising questions about whether the US tech giants were spending too much money on the wrong approach. DeepSeek said it trained its model for less than $6 million. Even if it was understated, the cheaper approach shook ideas about what it takes to compete.
Then summer came. GPT-5, released in August after nearly two years of development, was intended to deliver what CEO Sam Altman called “PhD-level intelligence.” Instead of, users received the model who marked Oklahoma on maps as “Gelabrine” and couldn't solve basic algebra. Prediction markets, which gave OpenAI a 75% chance of creating the best AI model, collapsed to 14% in one hour.
Now Google is using its advantage. Gemini 3, released this month, has received rave reviews. Salesforce CEO Marc Benioff said after two hours of testing that it was not “I'm coming back.” Analysts say that integration of the model into the Google search engine gives it a distribution advantage that OpenAI cannot match.
But Google's triumph may be just as temporary. Llama models from Meta fuel countless startups. Alibaba just announced major update to Qwen chatbot. Claude from Anthropic competes with ChatGPT about coding tests.
Google itself just published a paper on “Nested Learning,” which some researchers have compared to the 2017 Transformer breakthrough that underpins all of its large language model technology. If the new approach takes hold, the deck may be reshuffled again.
3 years is an eternity and a moment in AI
The speed of change makes predictions foolish. Three years ago, ChatGPT couldn't browse the web, remember previous conversations, or create images. Now he can do all three tasks, as well as analyze spreadsheets, write code, and have voice conversations that sound surprisingly human. Eight hundred million people use it every week.
But the same pace means that today's leader is constantly in danger, and even analysts can't keep up. In October, CNBC announced OpenAI's dominance.”unlike anything Silicon Valley has ever seen“, with one seasoned investor telling CNBC that this is the “fastest period of startup creation and disruption” he's seen in nearly two decades. By November, Altman advised employees to prepare for “rough vibrations” and “temporary economic obstacles” After the release of Gemini, OpenAI took a backseat.
And yet, despite all the madness, at least one person warned it was coming. In May 2023, six months after ChatGPT launched, an internal Google document was leaked that predicted the ups and downs to come. “We don't have a moat, and neither does OpenAI.“, the researcher wrote.
Unlike search algorithms or social networks, where data, infrastructure and network effects create long-term benefits, AI models can be freely copied, improved and distributed. The researcher argued that open source developers will eventually eclipse the big players, that the industry's obsession with scale is misguided, and that trying to control AI development is a losing game. Google ignored this. OpenAI did the same.
After what seems like a thousand news cycles later, the note reads less like internal dissent and more like prophecy. ChatGPT has changed the world. But when breakthroughs can take months to achieve, any benefit is temporary. Three years later, OpenAI's musical chairs may not find a place when the music stops.
