Can China win at AI without America's best chips?
In January 2025, a Chinese startup called DeepSeek shocked Wall Street with a model that rivaled America's best at a fraction of the cost. Four years into the chip bans, China is close behind, but running short of computers to use what it builds.
Plain-English explainer by Joe Miller, a nuclear engineer. Not investment advice. Figures are the latest public numbers we could find and change fast; sources are at the bottom of each page.
How far behind is China?
Not far, on paper. DeepSeek's own report for its V4 model (April 2026) says it trails the best U.S. models by about 3 to 6 months, and the research group Epoch AI estimates the U.S. lead at about seven months. V4's reasoning matched U.S. models released about six months earlier, but it fell short of OpenAI's and Google's newest.
The bigger gap is computing power. America's companies have far more AI chips in their data centers, so they can train bigger models and serve them to more people. DeepSeek reportedly admitted it couldn't offer its top V4 model to most customers because it didn't have enough chips to run it.
China's leading AI models
| Model | Company | Notes |
|---|---|---|
| DeepSeek V4 (Apr. 2026) | DeepSeek (Hangzhou) | 1.6 trillion parameters, free to download, tuned to run on Huawei chips |
| Kimi K2.6 | Moonshot AI | About as strong as DeepSeek on most tests |
| GLM 5.1 | Zhipu AI | Also close to DeepSeek |
| Qwen | Alibaba | The most downloaded family of free models in the world |
| Doubao | ByteDance (TikTok's owner) | China's most used chatbot app |
One pattern stands out: China gives many of its best models away as free downloads. That spreads Chinese AI around the world, especially in poorer countries, even when it isn't the very best.
China's own chips
- Huawei Ascend 950: Huawei expects about $12 billion in AI chip sales in 2026, up from $7.5 billion, and plans about 750,000 of its 950PR chips. It reportedly performs between Nvidia's H100 and H200, two to three years behind Nvidia's best.
- The catch is the factory. SMIC makes Huawei's chips on a 7-nanometer-class process without EUV machines. Yields are lower and each chip takes about eight months to make, versus about three at TSMC. Demand exceeds supply.
- Big targets: Beijing aims to triple domestic AI chip output in 2026, and SMIC plans large increases in 7- and 5-nanometer capacity by 2030.
Doing more with less
- Smarter software: Chinese labs became experts at efficiency: models that wake up only the parts of the network needed for each question, and long-memory designs that cut computing costs.
- Smuggled and rented chips: U.S. officials say DeepSeek trained V4 on banned Nvidia Blackwell chips; DeepSeek's report doesn't say what it used. See the smuggling section of the chip war.
- Copying answers ("distillation"): Anthropic and OpenAI accuse DeepSeek of using tens of thousands of fake accounts and millions of questions to learn from their models.
- Gaming chips: Chinese tech giants reportedly ordered more than 3 million Nvidia RTX Pro graphics cards, worth over $35 billion, that are still legal to sell.
Are the export bans working?
Yes, partly. They haven't stopped China from building good models, but they have kept China short of computing power, which limits how many people can use Chinese AI and how fast its labs can experiment. Most experts think the U.S. lead in chips is two to three years and growing at the very top, while the lead in models is months.
The risk is that the bans push China to build a complete supply chain of its own, from tools to chips, and then compete with Nvidia worldwide. That's already happening; the question is how fast.
My take
Think of it like power plants. You can have the best reactor design in the world, but if you can only build a few, you can't light up the country. China has the designs; America has the fleet. China is building its own fleet fast, so the U.S. lead is real but not permanent.
Sources
Checked November 16, 2026.
- Council on Foreign Relations: DeepSeek V4 signals a new phase in the U.S.-China AI rivalry (Apr. 29, 2026)
- Tom's Hardware: Huawei expects $12 billion in AI chip revenue as Nvidia's China share hits zero (May 5, 2026)
- China on AI: China sets 3x domestic AI chip output target for 2026 (Aug. 2026)
- Reuters via The Next Web: Nvidia H200 chips pile up in Hong Kong (Nov. 6, 2026)
- CSIS: What to know about Chinese AI models