The empire strikes back

Chinese AI researchers have more ingenuity than US policymakers give them credit for

Last updated: 1/22/2025

For most of the generative AI era to this point, American developers have enjoyed a clear lead over their Chinese peers. Since the launch of ChatGPT, observers have largely assumed that American innovation—fueled by world-class research institutions and abundant venture capital—would outpace China’s efforts for the foreseeable future. This week, a surprising new development from a Chinese AI lab has called that assumption into question.

“What the hell was that?”

The breakthrough that has caught the AI community off guard is the release of Deepseek R1-Distill-Qwen-1.5B (herein referred to as Deepseek-1.5B), an open-source large language model (LLM) developed by the Chinese AI lab Deepseek. Remarkably, this model outperforms OpenAI’s GPT-4o—considered one of the premier all-purpose LLMs—on key reasoning benchmarks. What makes this even more noteworthy is that Deepseek operates under constraints that American labs like OpenAI, Anthropic, Meta, and Google simply don’t face. Despite restrictions on high-end chips and other advanced computing resources, the team behind Deepseek-1.5B has managed to deliver a model that surpasses one of the West’s strongest offerings—until very recently considered the best on the market.

Why smaller is smarter

The “1.5B” in the model’s name stands for 1.5 billion parameters, a fraction of what powers industry-leading LLMs from the United States. Despite its modest size, Deepseek-1.5B manages to exceed or match GPT-4o in key areas like knowledge, reasoning, and prompt alignment.

This is crucial because a smaller model has enormous practical advantages:

• Device compatibility: A 1.5B-parameter model can run on consumer-grade hardware—think smartphones, laptops, and basic cloud setups—without requiring a giant cluster of specialized AI chips. (Note that a low parameter count does not necessarily mean less training compute).
• Cost savings: Reduced hardware demands mean far lower operating costs. This is especially significant in environments without easy access to cutting-edge GPUs. (e.g. the vast majority of start-ups).
• Greater accessibility: Being open-source and portable, the model can diffuse rapidly among developers and entrepreneurs around the world, powering new applications and services with minimal barriers to entry.

From a purely technical standpoint, “leading on AI” is about more than model capability alone. Model capability per unit of compute is also an important metric. At least one group of Chinese researchers appear to be innovating along this efficiency axis more aggressively than many in the West expected.

AI diffusion is a key strategic battleground

American policymakers have focused on limiting China’s access to advanced AI chips, viewing these controls as a strategic priority. However, this approach may have inadvertently spurred Chinese labs to “do more with less.” Forced to cope with fewer resources, these researchers have pushed the boundaries of algorithmic efficiency, creating powerful models that require only a fraction of the hardware their Western counterparts use. Constraint-driven innovation may enable widespread diffusion of AI faster than anything we’ve seen in the United States—an outcome at odds with the intentions behind export controls.

Even if American companies remain at the cutting edge of AI research, there is a second, equally important question: Which country will harness AI’s productivity benefits most effectively? Models like Deepseek-1.5B can be deployed anywhere. They work on local devices without requiring major infrastructure investments or advanced chips. They represent low barriers for new AI-powered tools, services, and business models. American labs are still producing state-of-the-art models, often with higher parameter counts, requiring vast data centers and specialized hardware. These models will likely retain an absolute edge in maximum capability, but they may not spread as widely or as quickly across diverse industries and markets—especially in places without robust cloud infrastructure.

Deepseek-1.5B highlights an important truth: innovation often thrives under constraints. By focusing on the efficiency of smaller models, Chinese researchers have tackled one of the field’s most pressing challenges—how to bring AI to the masses without requiring data supercomputer-level hardware. Years from now, we might look back on this week as a watershed moment in the ongoing US-China AI competition.

Deepseek-1.5B is a reminder that size isn’t everything. A highly efficient, smaller model can outperform larger, more resource-intensive rivals. As the United States continues to regulate high-end chip exports, Chinese researchers are finding ways around those constraints, pushing ahead in the art of building lean AI.

For Western policymakers, this should serve as a wake-up call. Ensuring national competitiveness in AI isn’t just about building the next giant language model. It’s also about diffusion—ensuring that AI-driven productivity gains are both widespread and accessible. If America’s strategy focuses only on stifling China’s access to cutting-edge hardware, it risks underestimating the ingenuity of China’s AI community.