The AI race: the day a free app erased a trillion dollars
Google invented the blueprint and hesitated, OpenAI sprinted, and a Chinese startup crashed the market in a weekend. The story of the race behind every chatbot you use.
On Monday, January 27, 2025, the American stock market lost roughly a trillion dollars in value in a single session, and the trigger was not a war, a bank failure, or a pandemic. It was a free app from a Chinese startup most people had never heard of. DeepSeek's chatbot had climbed to the top of the US App Store over the weekend, and its makers claimed they had trained their best model for a few million dollars — pocket change in an industry where the going assumption was that staying at the frontier cost billions. Nvidia, the company selling the shovels of this gold rush, fell about 17 percent that day, one of the largest single-day value losses of any company in history.
To understand why one app release could do that, you need the story of the race it interrupted.
How did the race start?
The strange part is that the key invention was given away. In 2017, researchers at Google published the transformer architecture in an open paper, and it turned out to be the design that makes modern AI work. Google had the blueprint, the money, the data, and most of the best researchers. What it did not have was the appetite to ship something as unpredictable as a chatbot to billions of users, because unpredictable answers are a small embarrassment for a startup and a headline disaster for a company that runs the world's search engine.
OpenAI had no such weight to carry. It was founded in 2015 as a nonprofit worried about AI safety, restructured itself around a Microsoft partnership worth billions, and in late 2022 released ChatGPT — built on Google's own published architecture. The launch forced everyone's hand. Google declared an internal emergency and rushed out its own chatbot, whose first public demo contained a factual error about the James Webb telescope; the mistake was spotted within hours and Google's parent company shed around a hundred billion dollars in market value that day. The lesson everyone took: this race would be sprinted, polish optional.
Who is actually racing?
Fewer players than the noise suggests, because the entry fee is brutal: chips, electricity, data, and a small pool of researchers who command athlete salaries.
OpenAI, wrapped in Microsoft's money and data centers, runs the most famous product. Google regrouped around its Gemini models and has one advantage nobody can buy — it already owns the search box, the phones, the email, and the office apps where AI gets used. Anthropic was founded in 2021 by researchers who left OpenAI over safety disagreements, and built Claude with a reputation for care; its money comes largely from Amazon and Google, because in this race even the referees bet on multiple horses. Meta took the most disruptive path: it gives its Llama models away with open weights, on the theory that if intelligence becomes cheap and everywhere, the value flows to whoever owns the social platforms. And Elon Musk, who co-founded OpenAI, left after a power struggle, then sued it, started xAI to compete with it.
Underneath them all sits Nvidia, which sells the specialized chips every single one of them needs. For most of the race it has been the only arms dealer in town, which is how a company that made graphics cards for gamers crossed a three trillion dollar valuation in 2024.
What did DeepSeek actually change?
The assumption it broke was that the frontier belonged only to those who could spend the most. DeepSeek, spun out of a Chinese hedge fund and working under US export restrictions that limited its access to the best chips, published models close to the American frontier, claimed startlingly low training costs, and released the weights openly. Skeptics argued the true costs were higher than advertised. It did not matter much. The point was made: efficiency could substitute for a meaningful part of raw spending, and clever engineering was not an American monopoly.
The market panic faded within weeks — Nvidia's chips remained in shortage, the big labs kept building — but the strategic picture had permanently shifted. The race was no longer just OpenAI versus Google. It was also open versus closed, and America versus China, with export controls on chips becoming as much a part of AI news as model releases.
Why should someone who just uses this stuff care?
Because the shape of the race decides what shows up in your pocket, at what price, with what strings attached.
Competition has made frontier AI absurdly cheap for users — capabilities that cost twenty dollars a month in 2023 are largely free in 2026 because nobody can afford to lose you to a rival. Open-weight models mean smaller companies, hospitals, and governments can run AI privately instead of renting it from three landlords in California. And the sprint pace explains the rough edges you have personally experienced: features ship because a competitor shipped last week, not because they were finished.
It also explains the stakes behind the headlines. When you read about chip export bans, data center power deals, or a lab poaching another lab's researchers with nine-figure offers, it is all the same story: a handful of organizations spending like nations because each believes that whoever builds the smartest machine gets to set the terms for everyone else. Nobody in the race believes it is optional to run. That, more than any single model release, is the fact that will shape the next decade of your software.