Saturday, November 21, 2026

Chinese AI and its challenges

There's an interesting article in today's New York Times about AI financing in China. 

Here in the West, we blithely assume that Chinese industries just get everything they need from Big Brother at central government. But, in reality, it seems like Chinese AI companies may be struggling even more than American ones right now.

The shine may be off AI in America to some degree, but that didn't stop Anrhropic from raising over US$80 billion in an IPO this last week. Chinese companies can only dream of such riches. The three largest investors in Chinese AI development - Alibaba, Baidu and Tencent - are all struggling financially, and the STAR 50 Index (China's equivalent to the Nasdaq) is even more awash on red ink than the Nasdaq, down 43% since the end of July.

Nvidia chips are still the bedrock of American AI development, and the gold standard for training AI models. China, however, has been denied access to the most advanced semiconductors by US export controls. Despite work arounds like renting datacentres in other Southeast Asian countries, it is still mainly using inferior homegrown chips. Even with this risky smuggling operation, Chinese AI developers still have access to about a tenth of the top Nvidia chips that America does, and Donald Trump is also moving to close down even that loophole (Trump views AI as a kind of arms race that he is unwilling to lose, or be seen to be losing). Chinese companies pay a huge premium for advanced computing power on the open market, as much as eight times the American price. 

But even if they could find better chips, the money to buy them is very tight. American AI companies are spending five to six times what their Chinese counterparts are able to do. And, crucially, Chinese companies are making a tenth times the income of the big American companies. 

Furthermore, China's embrace of open-source AI, while perhaps laudable in theory, means that new ideas spread globally quickly, and Chinese companies are often unable to profit from them. Rivals may offer the same technology at lower prices, squeezing Chinese profits still further. Chinese startups have struggled to figure out how to actually make money off consumers.

Huawei's semiconductor technology is about two years behind Nvidia's, and lacks sufficient production capacity to compete seriously. Four times as many Huawei 950 chips are needed to do what a Nvidia GB300 chip can do. That said, China's DeepSeek startup stunned the Americans recently by developing a powerful AI model at a fraction of the cost of US ones (using Huawei chips).

China does seem to be able to do more with less (of necessity). They have developed techniques to train and run powerful AI models more efficiently than their American counterparts. Chinese engineers are good and, behind the scenes, the Chinese government is shouldering some of the financial burden. Just how much they'll be prepared to shoulder remains an open question.

No comments: