Enjoy complimentary access to the Editor’s Digest! Each week, Roula Khalaf, the Editor of the Financial Times, handpicks her favorite stories for this newsletter.
For months, people have anticipated that the inflated valuations of American tech stocks were ripe for a market correction, yet the catalyst has emerged from an unlikely origin. China’s AI start-up, DeepSeek, has rolled out a cutting-edge language model that, while not quite on the level of a “Sputnik moment,” draws interesting parallels to the early space race era. Back then, the Soviet Union’s ingenious engineering efforts sometimes allowed it to surpass the US competitors, all despite having less computing power and fewer advanced technologies. While the US ultimately triumphed in the race to the Moon, establishing dominance in space exploration, DeepSeek’s strides challenge the assumptions behind American tech giants’ AI investments. This potential disruptor questions the belief in America’s unchallenged supremacy in AI, built on substantial spending on chips and infrastructure.
DeepSeek’s development revolves around a LLM, or large language model, that competes with US-based OpenAI and Meta, according to AI specialists, yet it purportedly does so using significantly fewer, less sophisticated Nvidia chips, and it claims to have trained these models at a dramatically lower cost. Although some of these claims still await verification, should they hold true, DeepSeek could emerge as a formidable player in the field.
Intriguingly, DeepSeek has chosen to make its model open source, a move similar to Meta’s approach but unlike OpenAI or Google’s Gemini, both of which guard their AI developments closely to maximize financial gain. By opting for open access, DeepSeek makes its technology attractive for developers eager to innovate. It also runs on a much tighter budget and requires less computing power, a factor that led to Monday’s dip in stock prices not only for frontline tech companies but also for those manufacturing chip equipment and supplying the hardware needed for data centers.
The US tech giants have heavily invested in an AI development model centered on large-scale spending on data, computing power, and capital. Yet, the success of companies like France’s Mistral and now DeepSeek underscores how entities with fewer financial resources can craft smarter, more efficient models. Such rising efficiencies are likely to fuel demand for these models’ applications.
Moreover, DeepSeek’s progress emphasizes China’s ability to make significant technical strides in AI, even amidst US-imposed export restrictions aimed at limiting its access to top-tier chips and advanced manufacturing tools. Instead of suppressing innovation, these constraints might have spurred Chinese firms to extract maximum performance from the resources they have. This resilience, demonstrated by highly skilled domestic engineers, challenges the notion that the technological stronghold constructed by well-funded US companies such as Meta, Google, OpenAI, and Anthropic is as secure as previously believed.
While countries with advanced democracies might be hesitant about China, seen as a counter to Western interests, gaining leadership in a transformative technology, certain political figures, consumers, and developers might find a less US-centric marketplace advantageous.
The current point of intrigue is less about who will build the top-performing AI models and more about who will effectively implement them in real-world scenarios. Kai-Fu Lee, a prominent voice in AI from China, has long maintained that China possesses a competitive edge in practical applications, even if it lags in infrastructure. Although the aggressive regulation of tech entrepreneurs in China and the surge of investment in US AI ventures have changed the playing field, DeepSeek’s accomplishments suggest a more leveled battle ahead.
