Chinese AI’s Sputnik Moment

This time, the great leap forward had Washington’s help.

This post was also published in National Review

What made the Soviets’ successful launch of the satellite Sputnik in 1957 a major crisis for America was our realization that Soviet technology had leapt ahead of the West in the space race. Nearly 70 years later, a new Chinese AI model should sound similar alarm bells across the West — not least because this time, U.S. policy is partly to blame.

The release of China’s free, open-source DeepSeek R1 in January 2025 showed China was closing the gap with American AI faster than Western analysts expected. Still, it was not quite a Sputnik moment, because R1 was optimized for Nvidia chips. Consequently, even if you wanted to run advanced Chinese AI models, you still needed American-made chips and developer tools.

That advantage has now vanished. DeepSeek V4, released in April 2026, took the world by surprise because it was optimized for Huawei chips. Even before DeepSeek V4, Chinese open-source models were already popular worldwide — including in the United States. Now, through the production of chips advanced enough to train and run those models, Chinese chipmakers have broken the overwhelming competitive advantage that U.S. chipmakers maintained in the AI race just a few months ago.

In practice, DeepSeek V4’s release means that demand for Chinese AI models will now inure to the benefit of the Chinese AI ecosystem rather than the American one. This, in turn, has reinforced a self-sustaining flywheel of investment, research, and development that positions China to leap ahead of the U.S. in the AI race.

The release of Z.ai’s free, open-source GLM-5.2 a few weeks ago — a direct result of that self-sustaining cycle — is a true Sputnik moment. Shockingly, GLM-5.2 exceeds OpenAI’s best paid model, GPT-5.5, on key engineering benchmarks, with performance very close to Anthropic’s flagship Claude Opus 4.8. Even more concerningly, GLM-5.2 was reportedly trained entirely on Huawei Ascend chips, with no U.S. chips involved at all. GLM-5.2 has industry-leading technical features long thought the domain of U.S. AI labs — for example, long-running agentic tasks such as coding, the AI feature that turned Anthropic into a trillion-dollar company almost overnight.

To underscore the pathbreaking nature of this development, an agentic AI model is one that can work autonomously toward a user’s defined goal on its own, devising complex plans with feedback loops, choosing what apps or tools to use, launching large numbers of other AI agents that it controls, correcting for mistakes, and continuing to completion with limited human direction. Introduced just a few months ago, agentic AI already accounts for more than half the traffic on the internet. And it is laying the foundations of the next phase of the AI revolution: robotics, where China has already developed a huge lead over the rest of the world.

Part of the thinking behind controls on exporting AI technology to China was that, deprived of American technology, China would fall even further behind in the AI race. Late last year, prominent export control proponents Jeff Smith and Bryan Burack of the Heritage Foundation confidently asserted, “For now, banning sales of advanced AI chips to China won’t lead Chinese companies to turn to a domestic competitor for a simple reason: there aren’t any. They don’t yet have the quality, scale, and reliability to satisfy Chinese domestic demand, much less offer a competitive ‘AI stack’ globally.” According to Smith and Burack, even inferior U.S. technology — such as Nvidia’s H20, which the Trump administration cleared for export to China last year — was “superior in performance, reliability, or availability over China’s homegrown alternatives.”

As with many similarly sanguine assessments, this turned out to be wrong in virtually every respect. Just two months later, Huawei released the new Ascend 950PR. The new chip has comparable inference performance to the Nvidia H100 — which still powers frontier models in the U.S — and is three times more powerful than the H20. Just as alarming, the new chips are almost certainly built using Chinese semiconductors. That means that China has broken the dependency on Taiwan’s semiconductor industry, something not even the United States has been able to do. It is now clear that China’s labs, cloud providers, chipmakers, and software engineers can replace U.S. technology at every layer of the AI stack, all while catching up in both price and performance.

It’s not just that U.S. export controls have failed to stop China’s ascent in AI. Our restrictions have accelerated China’s rise by compelling the country’s entire developer ecosystem to rely on domestically produced chips — thus forcing its chipmakers to rise to the challenge. Indeed, the main impact of U.S. export controls has been to help ensure the success of China’s industrial strategy for AI. Even in terms of total AI data center compute power, where the U.S. is still far ahead, Chinese data centers have more than enough compute power to train world-beating models (as GLM-5.2 shows), to say nothing of military applications.

Claims by export control proponents that Chinese AI companies are nowhere near competing with America for global market share have been shown to be mere wishful thinking. The first exports of Huawei’s Atlas 850E supercluster for use in AI data centers are slated to arrive in South Korea this year; South Korea will not be the only U.S. ally buying Chinese AI. The company is reportedly marketing systems that combine its new 950DT chips with DeepSeek V4 as a “full stack” solution for overseas customers in the Middle East and Central Asia. Furthermore, Huawei is reportedly considering deploying its newest chips inside existing cloud and AI servers in Latin America.