On December 26, the Chinese AI lab DeepSeek announced their v3 model.
Deploying underpowered chips designed to meet US-imposed restrictions and just US$5.6 million in training costs, DeepSeek achieved performance matching OpenAI’s GPT-4, a model that reportedly cost over $100 million to train.
Like most Chinese labs, DeepSeek open-sourced their new model, allowing anyone to run their own version of the now state-of-the-art system.
The announcement came amidst growing concern in Silicon Valley that the massive progress in AI capabilities has already reached an end. Had DeepSeek released their model four days earlier, it would have seemed that the future of AI lay in optimization and cost reduction rather than capability breakthroughs.
Instead, the announcement came within a week of OpenAI’s demonstration of o3, a new model that would rank in the 99.9th percentile of all competitive coders and could correctly solve the world’s hardest math problems at 10 times the rate of its predecessor.
The two events together signal a new era for AI development and a hotter race between the United States and China for dominance in the space. Chip export restrictions have not only failed to keep China significantly behind the US but have also failed to address the next frontier for AI development.
That frontier is reasoning – teaching AI to think step-by-step as humans do. While earlier models excelled at conversation, o3 demonstrates genuine problem-solving abilities, excelling not only at tasks that humans find simple, which often confounded AI, but also on tests that many AI leaders believed were years away from being cracked.
Microsoft CEO Satya Nadella has described the reasoning method as “another scaling law”, meaning the approach could yield improvements like those seen over the past few years from increased data and computational power.
Improvements following this path are less likely to strain the limits of chip capacity. Rather, talent, energy efficiency and cheap power will be key.
In Virginia, a major US data center hub, new facilities can wait years just to secure power connections. After two decades of flat demand, US utilities and regulators are scrambling to adapt to the massive power requirements of advanced AI.
Meanwhile, China is rapidly expanding its power infrastructure, with new integrated computing networks being built across regions like Beijing-Tianjin-Hebei. China’s electricity generation has increased 64% in the past decade, while the United States’ has stalled.
But viewing the race at the country level alone can be misleading. Rather than an established tech giant with significant government ties like Tencent or Alibaba or ByteDance releasing the country’s best model, it was a lab of perhaps 200 people behind DeepSeek and a culture that made the most of that talent.
The United States remains a hub for global talent, but, according to a recent PNAS publication, Chinese researchers are ditching America to return home in greater numbers than ever before.
Outgoing US Secretary of Commerce Gina Raimondo called attempts to hold back China a “fool’s errand” in an interview with the Wall Street Journal late last month.
Ten days later, researchers at China’s Fudan University released a paper claiming to have replicated o1’s method for reasoning, setting the stage for Chinese labs to follow OpenAI’s path.
“The only way to beat China is to stay ahead of them,” Raimondo continued. “We have to run faster, out innovate them. That’s the way to win.” In the race to lead AI’s next level, that’s never been more clearly the case.
Ben Dubow is a senior fellow at the Center for European Policy Analysis and Chief Technology Officer at the AI and intelligence firm Omelas
