One week ago, a new and formidable challenger for OpenAI’s throne emerged. A Chinese AI start-up, DeepSeek, launched a model that appeared to match the most powerful version of ChatGPT—but, at least according to its creator, was a fraction of the cost to build. The program, called DeepSeek-R1, has incited plenty of concern: Ultrapowerful Chinese AI models are exactly what many leaders of American AI companies feared when they, and more recently President Donald Trump, have sounded alarms about a technological race between the United States and the People’s Republic of China. This is a “wake up call for America,” Alexandr Wang, the CEO of Scale AI, commented on social media.
But at the same time, many Americans—including much of the tech industry—appear to be lauding this Chinese AI. As of this morning, DeepSeek had overtaken ChatGPT as the top free application on Apple’s mobile-app store in the U.S. Researchers, executives, and investors have been heaping on praise. The new DeepSeek model “is one of the most amazing and impressive breakthroughs I’ve ever seen,” the venture capitalist Marc Andreessen, an outspoken supporter of Trump, wrote on X. The program shows “the power of open research,” Yann LeCun, Meta’s chief AI scientist, wrote online.
Indeed, the most notable feature of DeepSeek may be not that it is Chinese, but that it is relatively open. Unlike top American AI labs—OpenAI, Anthropic, and Google DeepMind—which keep their research almost entirely under wraps, DeepSeek has made the program’s final code, as well as an in-depth technical explanation of the program, free to view, download, and modify. In other words, anybody from any country, including the U.S., can use, adapt, and even improve upon the program. That openness makes DeepSeek a boon for American start-ups and researchers—and an even bigger threat to the top U.S. companies, as well as the government’s national-security interests.
To understand what’s so impressive about DeepSeek, one has to look back to December, when OpenAI launched its own technical breakthrough: the full release of o1, a new kind of AI model that, unlike all the “GPT”-style programs before it, appears able to “reason” through challenging problems. o1 displayed leaps in performance on some of the most challenging math, coding, and other tests available, and sent the rest of the AI industry scrambling to replicate the new reasoning model—which OpenAI disclosed very few technical details about. The start-up, and thus the American AI industry, were on top. (The Atlantic recently entered into a corporate partnership with OpenAI.)
DeepSeek, less than two months later, not only exhibits those same “reasoning” capabilities apparently at much lower costs, but has spilled at least one way to match OpenAI’s more covert methods to the rest of the world. The program is not entirely open-source—its training data, for instance, and the fine details of its creation are not public—but, unlike with ChatGPT, Claude, or Gemini, researchers and start-ups can still study the DeepSearch research paper and directly work with its code. OpenAI has enormous amounts of capital, computer chips, and other resources, and has been working on AI for a decade. In comparison, DeepSeek is a smaller team formed two years ago with far less access to essential AI hardware, because of U.S. export controls on advanced AI chips, but it has relied on various software and efficiency improvements to catch up. DeepSeek has reported that the final training run of a previous iteration of the model that R1 is built from, released in December, cost less than $6 million. Meanwhile, Dario Amodei, the CEO of Anthropic, has said that U.S. companies are already spending on the order of $1 billion to train future models. Exactly how much the latest DeepSeek cost to build is uncertain—some researchers and executives, including Wang, have cast doubt on just how cheap it could have been—but the price for software developers to incorporate DeepSeek-R1 into their own products is roughly 95 percent cheaper than incorporating OpenAI’s o1, as measured by the price of every “token”—basically, every word—the model generates.
DeepSeek’s success has abruptly forced a wedge between Americans most directly invested in outcompeting China and those who benefit from any access to the best, most reliable AI models. (It’s a divide that echoes Americans’ attitudes about TikTok—China hawks versus content creators—and China’s other apps and platforms.) For the start-up and research community, DeepSeek is an enormous win. “A non-US company is keeping the original mission of OpenAI alive,” Jim Fan, a top AI researcher at the chipmaker Nvidia and former OpenAI employee, wrote on X. “Truly open, frontier research that empowers all.”
But for America’s top AI companies, and the nation’s government, what DeepSeek represents is unclear. The stocks of many major tech firms—including Nvidia, Alphabet, and Microsoft—dropped this morning amid the excitement around the Chinese model. And Meta, which has branded itself as a champion of open-source models in contrast to OpenAI, now seems a step behind. (The company is reportedly panicking.) To some investors, all those massive data centers, billions of dollars of investment, or even the half-a-trillion-dollar AI-infrastructure joint venture from OpenAI, Oracle, and SoftBank, which Trump recently announced from the White House, could seem far less essential. Maybe bigger AI isn’t better. For those who fear that AI will strengthen “the Chinese Communist Party’s global influence,” as OpenAI wrote in a recent lobbying document, this is legitimately concerning: The DeepSeek app refuses to answer questions about, for instance, the Tiananmen Square protests and massacre of 1989 (although the censorship may be relatively easy to circumvent).
None of that is to say the AI boom is over, or will take a radically different form going forward. The next iteration of OpenAI’s reasoning models, o3, appears far more powerful than o1 and will soon be available to the public. There are some signs that DeepSeek trained on ChatGPT outputs (outputting “I’m ChatGPT” when asked what model it is), although perhaps not intentionally—if that’s the case, it’s possible that DeepSeek could only get a head start thanks to other high-quality chatbots. America’s AI innovation is accelerating, and its major forms are beginning to take on a technical research focus other than reasoning: “agents,” or AI systems that can use computers on behalf of humans. American tech giants could, in the end, even benefit. Satya Nadella, the CEO of Microsoft, framed DeepSeek as a win: More efficient AI means that use of AI across the board will “skyrocket, turning it into a commodity we just can’t get enough of,” he wrote on X today—which, if true, would help Microsoft’s profits as well.
Still, the pressure is on OpenAI, Google, and their competitors to maintain their edge. With the release of DeepSeek, the nature of any U.S.-China AI “arms race” has shifted. Preventing AI computer chips and code from spreading to China evidently has not tamped the ability of researchers and companies located there to innovate. And the relatively transparent, publicly available version of DeepSeek, rather than leading American programs, could mean Chinese programs and approaches become global technological standards for AI—akin to how the open-source Linux operating system is now standard for major web servers and supercomputers. Being democratic—in the sense of vesting power in software developers and users—is precisely what has made DeepSeek a success. If Chinese AI maintains its transparency and accessibility, despite emerging from an authoritarian regime whose citizens can’t even freely use the web, it is moving in exactly the opposite direction of where America’s tech industry is heading.