The Sequence Opinion - Issue 935: Chinese Algorithmic Efficiency vs. American Scale in Frontier AI

How Chinese and American labs pursue the next generation of intelligence Imagine giving two AI teams the same challenge: make the model substantially smarter. One team asks for a larger GPU cluster. The other starts interrogating the architecture. Why are we moving this much memory? Does every token need the same computation? Could a better optimizer teach the model more from each training example? These instincts help explain a fascinating contrast in frontier AI. Chinese labs such as DeepSeek and Moonshot have made algorithmic efficiency unusually visible in their releases. American frontier competition also features enormous infrastructure ambitions, exemplified by OpenAI’s Stargate project. The resulting debate often sounds like a contest between cleverness and purchasing power. That framing misses how both approaches actually produce progress. The useful question concerns the next dollar. Should a lab spend it acquiring more computation, or making its existing computation more productive? The answer shapes everything from model architecture to the price of an agent completing a task. It also changes over time: an algorithmic breakthrough can make a much larger training run suddenly worth attempting. The incentives behind the divide Read more