The Export Controls Worked. That's Why Chinese AI Is Everywhere.

On Saturday, DeepSeek suspended a planned $71 billion fundraising round. The cause was a leaked private admission from founder Liang Wenfeng: China still lags behind the United States in AI capability and remains heavily dependent on Nvidia chips. Investors who had been pricing in Chinese dominance pulled back the moment they heard the founder's own assessment ^[1]. The admission is the strongest single proof that US export controls achieved what they were designed to do. The capability gap the restrictions set out to preserve is real, and the man who runs China's most celebrated AI lab concedes it in the room where the money is. The public evidence on that gap is contested — Stanford's 2026 AI Index reported in April that the US-China performance gap has "effectively closed," with Anthropic holding only a 2.7 percent lead — but a private admission to investors carries a weight public benchmarks do not: it is what the founder told the people writing the checks. The paradox is that the same export controls that preserved the capability gap created a different one. Each step in the sequence is documented. The restrictions cut DeepSeek off from Nvidia's most advanced chips. The company optimized its V4 model for Huawei's Ascend 950 instead, excluding Nvidia and AMD from early access entirely ^[2]. That optimization enabled DeepSeek to permanently cut V4-Pro prices by 75 percent — a reduction the company explicitly tied to the Ascend 950. The result is that Chinese frontier models now cost 12 to 19 times less than GPT-5.5 or Claude Opus 4.7 for equivalent tasks. Nvidia CEO Jensen Huang predicted exactly this. When the restrictions were being debated, he warned that export controls must be "commercially practical, else the market will continue to move to foreign alternatives." Then he went further.

Although KYC is important, KYC is not the issue. For American industry to make any sales, the conditions need to be commercially practical, else the market will continue to move to foreign alternatives. — Nvidia

The competitor now costs one-twelfth as much. The downstream consequences are no longer speculative. Roughly 80 percent of American AI startups now use Chinese open-source base models, with Alibaba's Qwen surpassing Meta's Llama in downloads. Chinese models process 12.96 trillion tokens globally against 3.03 trillion for US models — a four-to-one margin — and all six of the world's most-used AI models are Chinese. Pinterest integrated DeepSeek R-1 for recommendations and reported 30 percent higher accuracy than off-the-shelf proprietary alternatives; Airbnb uses Alibaba's Qwen for customer service. The Trump administration has no coherent response to any of this because it has five positions on Chinese AI and no mechanism to reconcile them. The president's own policy is détente: he authorized Nvidia H200 chip sales to China with a 25 percent government revenue cut and withdrew Commerce Department plans to restrict Chinese drone imports ^[3]. The House Foreign Affairs Committee unanimously passed a bill to sanction foreign entities that distill capabilities from US models ^[4]. The House separately launched investigations into Cursor and Airbnb for using Chinese models. White House AI adviser David Sacks offered a different explanation for why anyone would want to restrict Chinese open-source models.

the leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition. — David Sacks

And under a federal law targeting foreign adversary-controlled applications, Apple geoblocked ByteDance's apps — Douyin, Doubao, CapCut, Lark — for US iOS users, even as the open-source models 80 percent of startups actually use remained untouched. Five positions, no center. Microsoft operates on both sides of the line. The company built a business worth more than a billion dollars a year selling OpenAI models to Chinese firms — ByteDance, Tencent, Ant Group, Meituan — through Azure, routing via Singapore. Its own executive Judson Althoff was candid about the arrangement.

The one company bringing those two places together is Microsoft. It’s pretty awesome. — Microsoft

At the same time, Microsoft ended its exclusive cloud partnership with OpenAI in April, relinquishing its exclusive IP license, and is now exploring self-hosting DeepSeek-V4 to power a lower-cost Copilot tier ^[5]. The company that profits from selling American AI to China is hedging toward Chinese AI for its own products. The structural irony is not subtle. US export controls achieved their stated goal: the adversary's own founder privately concedes the capability gap. And the mechanism that preserved that gap — forcing Chinese labs onto domestic chips — is the same mechanism that produced models now being adopted at scale inside the American economy. Jensen Huang warned that restrictions would create a competitor the US would not otherwise have. That competitor now costs one-twelfth as much, and four out of five American AI startups are using it.