More 'DeepSeek Moments' Ahead: Will US Continue to Build Walls or Learn?

By Global Times

Jun 28, 2026 11:01 PM

Illustration: Chen Xia/GT

A Chinese AI model ranks second globally in coding ability - behind only the best American model, at a fraction of the cost. Many media outlets are already calling it "another DeepSeek moment" or "DeepSeek 2.0." 

The "Sputnik moment" describes America's sudden realization, after the Soviet Union launched the first satellite, that it had fallen behind and needed to catch up quickly. The "DeepSeek moment" borrows from that same logic. It refers to a Chinese team building a high-performance model with limited resources, challenging the conventional wisdom about the China-US AI gap. This suggests that something significant has shifted. Chinese AI can now compete where it matters.

DeepSeek first shook the industry with a low-cost, high-performance reasoning model. Zhipu AI's GLM-5.2, released recently, follows the same playbook. It handles long-context coding tasks and supports full software development workflows - from development and testing to deployment - while costs roughly one-sixth as much as comparable US models. 

On Code Arena, a large-scale, user-driven coding benchmark based on blind comparisons, GLM-5.2 scored 1,595 points - ranking second globally and first among all publicly available models. It is also open-source, freely available for anyone to use or modify.

What makes this pattern significant is the logic behind it.

Chinese companies are no longer just stacking more chips to get more power. They are getting more out of what they have - through algorithmic efficiency, smarter architectures, and relentless engineering.

Resource constraints have forced a more pragmatic approach. The IndexShare architecture in GLM-5.2, for example, reduces per-token computation by about 2.9 times at a 1-million-token context. The gap with top US models on long-horizon coding benchmarks has narrowed to 1-4 percent.

The market is already responding. OpenRouter data shows that the share of token requests going to Google, OpenAI and Anthropic models has dropped from 72 percent a year ago to 30 percent. More users are turning to cheaper, faster Chinese open-source models for routine tasks. As one analyst put it: "You don't need a Nobel laureate to fill out a reimbursement form."

The obvious question is: What can stop this from happening again? The answer is: nothing.

US sanctions and export controls have not stopped Chinese AI progress. In fact, they may have accelerated it. When access to advanced chips is cut off, Chinese labs have to find alternative solutions. The AI chip self-sufficiency ratio in China has risen from about 10 percent in 2021 to 41 percent in 2026. 

Former Google chief Eric Schmidt recently admitted that US hardware controls "failed to stop China" and that the gap between Chinese AI models and US state-of-the-art has shrunk to about six months.

Each new Chinese model release resets expectations regarding cost and accessibility, forcing US incumbents to respond. Nvidia, a company whose advantage has long rested on closed hardware and software, released an open-weight model in March 2026 - a notable shift. The competitive dynamic has become a forcing function for both sides.

The next few years will likely bring more such moments, not just in coding, but in other domains as well. Each one will narrow the gap a little further. The sanctions and controls will not stop this; instead, they will produce more pragmatic, efficient solutions.

This brings us to the real question: How can both sides contribute to global AI development within this competition - rather than trying to defeat each other? The evidence increasingly suggests that containment is not working. The more you try to wall off Chinese AI, the more determined and resourceful it becomes.

The "DeepSeek moment" was never a one-time event. It was a preview. More are coming. The only question is whether Washington will continue trying to build walls, or start asking what it can learn from a competitor that has figured out how to do more with less.