Tools: America Spent $100 Billion Trying to Stop Chinese AI. It Didn't Work.

The Hardware Wall That Wasn't ## 80% of Silicon Valley Runs on Chinese Code ## The Paradox the Pentagon Didn't Plan For ## What $100 Billion Bought The most downloaded AI system in the world isn't made by OpenAI, Google, or Meta. It's Alibaba's Qwen. Seven hundred million downloads on Hugging Face as of January 2026, overtaking Meta's Llama in October 2025. December downloads alone exceeded the combined total of the next eight leading models. The second most important AI release of February wasn't GPT-5.3-Codex or Claude Opus 4.6. It was GLM-5, a 744-billion-parameter model from the Chinese lab Zhipu AI — trained entirely on Huawei Ascend chips. No Nvidia. No TSMC. No American silicon anywhere in the stack. Three years and multiple rounds of export controls later, the numbers tell a clear story: the controls slowed China's chip production. They did not slow China's AI. US export controls achieved exactly one thing well: they crippled China's ability to manufacture advanced chips. ASML's EUV lithography tools never reached SMIC. Huawei produced only 200,000 AI chips in 2025 — a fraction of what it could have built without restrictions. Commerce Secretary Howard Lutnick cited this figure in congressional testimony as proof the strategy worked. But manufacturing chips and building AI models turned out to be different problems. GLM-5 runs 744 billion parameters across 256 mixture-of-experts modules, with 44 billion active per inference. It was trained on 100,000 Huawei Ascend 910B chips using 28.5 trillion tokens. The Ascend 910B doesn't match an H100 in raw compute. Zhipu compensated with software optimization and sheer cluster scale. The result competes with GPT-5.2 on Humanity's Last Exam (50.4% vs 47.8%) and SWE-bench Verified (77.8% vs 76.2%). It costs roughly six times less per token. MIT license. Full weights on Hugging Face. Free chat tier at chat.z.ai. The model achieved the industry's lowest hallucination rate — 34%, compared to Claude Sonnet 4.5's 42% — through a reinforcement learning technique called Slime, which uses asynchronous active partial rollouts to eliminate sequential training bottlenecks. The name is odd. The results are not. Martin Casado, a general partner at Andreessen Horowitz managing a $12.5 billion infrastructure fund, made a quiet observation during pitch meetings: among startups presenting with open-source stacks, roughly 80% are running on Chinese models. DeepSeek. Qwen. Kimi. Zhipu. Software engineer Rohan Paul noticed something sharper: the top 16 open-source models on global leaderboards are all Chinese. The highest-ranked non-Chinese open-source model sits at number 17. Chinese open-source AI went from 1.2% of global usage in late 2024 to nearly 30% by late 2025. That's not a trend. That's a migration. Why? The economic argument is blunt. Startups burn cash. Chinese open-source models are free, permissively licensed, and increasingly competitive. American alternatives from Meta came with restrictive licenses and uncertain futures. When your runway is eighteen months and your inference bill determines whether you survive, you pick the model that's free and works. Alibaba has open-sourced nearly 400 models across the Qwen family. More than 180,000 derivative versions exist. The ecosystem isn't just big. It's self-reinforcing. Export controls created a specific paradox: China can build frontier AI models but struggles to deploy them at scale. DeepSeek had to restrict API access after releasing R1 because it lacked sufficient inference compute. Up to 80% of China's AI chips may sit unused in data centers that can't maintain the stability needed for production serving. China leads in model capability. America leads in deployment infrastructure. The controls successfully prevented China from exporting AI compute — one announced deal for 3,000 Ascend GPUs in Malaysia was retracted by the Malaysian government under pressure. China's influence in global AI cloud infrastructure remains negligible. But here's the problem with that framing: the models travel freely. Code doesn't need a shipping lane. When Zhipu publishes GLM-5 weights under MIT license, the entire world gets access to Chinese-trained AI — including the 80% of American startups already building on it. Stanford's AI Index Report 2025 concluded that Chinese labs are, at worst, "fast followers" in model capabilities. The gap between leader and follower now measures in weeks, not years. And DeepSeek's founder Liang Wenfeng told researchers: "Money has never been the problem for us. Bans on shipments of advanced chips are the problem." He said this while his team released models that matched GPT-4.5 on standard benchmarks. The US has spent enormous political capital, disrupted billions in semiconductor trade, and alienated allies over chip export controls. The bet was that restricting hardware would restrict AI capability. For three years, that logic held. It doesn't hold anymore. Zhipu's Hong Kong IPO raised $558 million in January 2026. Its stock has surged over 300% since listing. The company is already planning a secondary float in Shanghai. It designed compatibility not just with Huawei but with Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame, and Hygon — building a parallel hardware ecosystem that doesn't need a single American component. The export control regime successfully delayed China's chip manufacturing by several years. It failed to delay China's AI by several months. The models got built anyway, on inferior hardware, with better software. And now those models power the majority of new American startups. That's not a failure of execution. It's a failure of theory. The assumption was that AI follows hardware. In 2026, hardware follows software — and the best open-source software is Chinese. Templates let you quickly answer FAQs or store snippets for re-use. Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse