How Is China Keeping Up in AI?

Saudi Arabia has money. Taiwan has chips. India has engineers. Yet only China builds models that rival ChatGPT and Claude. How?

Building frontier AI rests on five pillars that must work at the same time: money, chips, electricity, training data and people. China got all five working together. In January 2025, DeepSeek proved it by matching top US models at a fraction of the cost, and today Alibaba's Qwen and Moonshot's Kimi are close behind.

Were these labs really new?

No. They had been building AI for years before ChatGPT arrived. DeepSeek grew out of High-Flyer, a hedge fund started in 2015 that was already using AI to trade stocks and had built its own AI supercomputers by 2021. Alibaba has run its AI research lab DAMO since 2017, and ByteDance has used AI since 2012 to choose what people see on its apps, and later on TikTok. When ChatGPT arrived, they pointed years of experience at a new goal, and DeepSeek even started with the open design Meta shared for its LLaMA model. Now let's see how China competed on each pillar.

Where did the money come from?

First came private money. DeepSeek's founder paid for the early work out of his hedge fund's profits. Once DeepSeek proved China could compete, big investors rushed in. In 2026, DeepSeek raised about 23 billion dollars, and Alibaba committed about 53 billion dollars over three years to AI and cloud.

How did they get chips when the US banned them?

China competed with a handicap here, and it found three ways around it. Before the US tightened its rules in late 2023, Chinese firms legally stockpiled tens of thousands of Nvidia's China-only H800 chips. They also rented top Nvidia computers in other countries. And US prosecutors have exposed smuggling rings that used fake paperwork to move banned chips into China through brokers in Singapore and Malaysia.

How does China handle the electricity?

China produces more than double the electricity of the US, and it puts many data centers in western provinces rich in wind, water and solar power. Chinese chips like Huawei's Ascend use 30 to 50 percent more power than Nvidia's for the same work, so the state steps in to cover that cost. To support data centers that use Chinese chips, local governments in Gansu, Guizhou and Inner Mongolia cut their power bills by up to half.

Where did the training data come from?

Training data is the examples a model learns from, and creating good examples takes a lot of time and money. Instead, Anthropic says Chinese labs used thousands of fake accounts to pull answers out of Claude, so their own models could learn to copy a stronger one. Anthropic's September 2026 report linked over 151 million of these exchanges to Alibaba. It also says Moonshot and DeepSeek secretly sent their own users' questions to Claude, showed Claude's answers as their own, and saved them for training.

Where did the AI scientists come from?

Mostly from home. In 2025, 57 percent of the world's top AI researchers did their first degree in China, and 69 percent of them now work in China. China also brings in experienced people from the US. ByteDance hired Wu Yonghui, a 17-year Google veteran, and a state program offered overseas chip experts signing bonuses of up to 700,000 dollars.

But why do Chinese labs give their models away for free?

It's a money war. Free models that are nearly as good push down the price of AI, which hurts OpenAI and Anthropic, who are spending hundreds of billions on computers and need high prices to earn that money back. Free models also get developers around the world to improve Chinese AI for no pay. And they place Chinese models, trained to follow Beijing's rules on sensitive topics, deep inside Western software, where US bans can't reach them.

What to remember

Most countries have one or two of these pillars, like money, chips or engineers. China is the only country outside the US that got all five working at the same time, and that is why its models rival ChatGPT and Claude.