In this article:
Frank, talking about AI and China, in the West there's a lot of anxiety about the arrival of AI and robots taking over, that kind of thing. But I don't detect the same level of apprehensiveness in China.
I think you're right. So right now, a bigger source of anxiety for the Chinese companies and the Chinese government is less about, you know, AI taking over the world or getting too advanced in China. Actually, on the contrary, the worry is that it is less advanced when compared with US peers.
AI models in China are actually trained in both the Chinese language and the English language. So that leads to a problem, where if you input the prompt of, ‘Give me a picture of a patriotic cat,’ in Chinese, because the language model generates the picture using both Chinese and English language, what you will get is actually a cat wearing the national flag of the United States.
So right now, both the Chinese government and Chinese companies are working to try to narrow the gap with US AI technology. And the Chinese government has recently launched some policies to coordinate the efforts of Chinese companies to try to promote technology development and get ahead in this field.
If China is trying to close the gap with the US, what does that mean for how heavy-handed or light-handed the government will be when it comes to regulating AI?
I think they will be accommodative. We just saw regulators in China announce the first batch of 41 approved AI algorithms. The approval came faster than I expected and included algorithms from most of the major local Chinese Internet companies. I think this shows that authorities are working fast to give people the necessary approvals. I don't see this as a barrier to entry, I think it’s a supportive signal from the government.
What about the US banning exports of advanced chips? How big a challenge is that for Chinese companies doing AI?
It is surely hurting the companies and slowing them down. For example, the same algorithm that takes US companies one week to train may take Chinese company, you know, two or three weeks. When I talk to experts, it’s one of the biggest challenges they mention. Most of the chips used by Chinese companies developing AI algorithms were made by Nvidia. And either the chips were advanced chips purchased before the export restrictions or they are a newer, reduced-capability chip model called the 800 series that are OK to export directly to China. It costs less but is a lot slower. And for those more basic chips there’s a huge backlog because everybody is placing orders, so the wait time for Chinese companies can be several months or even several quarters.
What about workarounds, like offshoring the process of training algorithms so that part is done in the US?
That only works for a small handful of Chinese companies that have a US subsidiary or sizeable team in the US that can train the AI. And it’s also not ideal because it physically separates the team training the AI from the users, who are in China. The solution for most Chinese companies will be doing this at home. Already more than 20 companies have already launched large language models in China that have been trained domestically with existing GPUs (graphics processing units).
What do you think about the massive run up in valuations of Chinese AI stocks in recent months? Is it a bubble?
I do think there is a bit of a bubble, at least for certain stocks. For example, there was a company claiming it would do its own LLMs (large language models) that saw its stock price roughly triple in a matter of weeks before its largest shareholder sold off a bunch of the stock. That’s just one big sign of a bubble but there are others in the market. Don't get me wrong, I think there will be some real winners out there, amid all these so-called ‘AI theme stocks’ in China. But expectations overall are just pretty high at the moment.
So in a nutshell, what’s the investment case for you?
From what I'm observing, many Chinese companies are not trying to be very innovative in terms of inventing a different type of algorithm and are instead adapting existing work and trying to replicate the success of GPT, but by doing it domestically in China. At least, that’s the current focus for a lot of them. Often this comes down to enhancing existing verticals, for example making software easier for users to interact with by enabling better speech recognition.
But I think we need to be very selective in this environment. Because while we can see lots of use cases for AI in China, what’s less clear is the ability to monetise those capabilities. It’s really hard to predict whether enhancing functionality means you can actually charge a customer more for something. So at the moment in China I am focused on a few companies such as in the software space, and only where I see a very concrete use case with a clear path to monetisation.