The United States and China are building AI in two very different ways. The U.S. is focused on raw power, scaling up massive frontier models that require enormous amounts of compute, water, and energy. China is focused on efficiency, using techniques like distillation to compress large models into smaller ones that run cheaply on domestic hardware. These two strategies were shaped by the chip embargoes. The U.S. tried to limit China’s access to high‑end GPUs, and China responded by optimizing around the hardware it already had. The result is a “bigger is better” approach in the U.S. and a “smaller is smarter” approach in China.
Distillation is the key to China’s strategy. It takes a large model and trains a smaller model to mimic its behavior. The smaller model doesn’t need the same compute, yet it can perform many of the same tasks. This means China can scrape outputs from Western models, feed them into its own training pipeline, and produce a distilled version that is cheaper to run. In practice, China can use Western compute as a starting point, then optimize it until it fits within its own hardware limits. The U.S. pays full price for innovation. China pays a fraction of the cost to replicate the capability.
The chip embargoes pushed China even further down this path. High‑end GPUs became harder to acquire, so China invested in software efficiency, model compression, and distributed training. At the same time, China’s energy and water infrastructure is spread across many regions, which makes it expensive to run giant centralized data centers. Smaller, more efficient models fit better into China’s resource map. They can be deployed in more places, at lower cost, and with fewer infrastructure bottlenecks. China’s AI ecosystem ends up being more nimble because it has to be.
Meanwhile, the U.S. is racing toward superintelligence by scaling up. Each new model requires more compute, more cooling, more water, and more money. The cost per token is heavily subsidized, which hides the true expense of running these systems. If users had to pay the unsubsidized price, the market would look very different. The U.S. approach assumes that bigger models will unlock new capabilities. China’s approach assumes that efficiency will win in the long run. These assumptions lead to very different AI landscapes.
China’s efficiency advantage means it can train and deploy models at a much lower cost. A ten‑to‑one efficiency ratio means China can afford to run AI in places where the U.S. cannot. It also means China can scale AI across more industries, more cities, and more devices. If China can distill Western models into compact versions that run cheaply, it can effectively use Western compute against the West. Every time a U.S. company trains a massive frontier model, China can scrape the outputs, distill them, and deploy a cheaper version at scale.
For beginners, the core idea is simple. The U.S. is building giant, expensive models that push the limits of global compute. China is building smaller, efficient models that learn from the giants. The chip embargoes pushed China toward efficiency, and China turned that constraint into a strategy. The U.S. is trying to brute‑force its way to superintelligence. China is trying to efficiently get there. Both paths lead forward, but one costs a lot more than the other.
