Chinese AI companies are breathing down the necks of the US frontier labs at benchmarks, but Google Deepmind CEO Demis Hassabis believes that they haven’t yet made any novel innovations beyond what the US has already achieved.
Demis Hassabis, the CEO of Google Deepmind and a Nobel laureate who co-leads one of the world’s most influential AI research organizations, offered a nuanced assessment of the global AI race during a recent interview. While acknowledging China’s rapid progress in matching Western benchmarks, Hassabis drew a critical distinction between execution excellence and fundamental innovation—a difference he believes still separates the two technological superpowers.
“The US and the West is still in the lead in AI if you look at the latest benchmarks and the latest systems,” Hassabis said. “But China is not far behind. If you look at the latest Deepseek, or the latest models, they’re very good. There are some very capable teams there. So maybe the lead is just a matter of months as opposed to years at this point.”
When pressed on whether China might already be winning if semiconductor constraints were removed from the equation, Hassabis pushed back firmly. “No. I think chips is one thing, but algorithmically, innovation-wise, the West still has the edge,” he explained. “I don’t think any of the Chinese models or companies have shown that they can innovate algorithmically on something new beyond the state-of-the-art. They’ve been very good at fast following the current state-of-the-art.”
The distinction Hassabis draws speaks to a fundamental tension in the current AI landscape. Chinese companies like Deepseek, Alibaba, and Baidu have indeed demonstrated remarkable engineering prowess, often achieving comparable performance to Western models while working under significant hardware constraints imposed by US export controls on advanced chips. Deepseek’s recent releases, for instance, have garnered attention for their efficiency and competitive benchmark scores, accomplished with less cutting-edge hardware than their American counterparts typically employ.
However, Hassabis’s comments suggest that replicating and optimizing existing architectures—however skillfully—differs meaningfully from pioneering entirely new approaches. The major algorithmic breakthroughs in recent years, from transformer architectures to reinforcement learning from human feedback (RLHF) to chain-of-thought prompting, have predominantly emerged from Western research labs. Google Deepmind itself has been responsible for landmark innovations including AlphaGo, AlphaFold for protein structure prediction, and contributions to the transformer architecture that underpins modern large language models.
The implications of Hassabis’s assessment extend beyond technical considerations into geopolitical strategy. If China’s strength lies primarily in rapid implementation rather than foundational innovation, it suggests that maintaining the West’s lead depends less on restricting access to current technology—an approach that breeds resourcefulness and optimization—and more on sustaining the research environments that produce paradigm-shifting breakthroughs. This would favor open scientific collaboration, substantial research funding, and the concentration of top AI talent that Western institutions have historically attracted. As the gap narrows from years to months, the question becomes whether incremental optimization can eventually catch and surpass innovation, or whether the next major algorithmic leap will once again widen the divide.