彭博：中国AI模型在成本与使用量上领先美国

Well, it's China versus The United States in the battle for AI dominance, and China is closing the gap. In fact, when it comes to usage and cost, China is now in the lead. Of course, this is causing concern in Washington, Wall Street, and in Silicon Valley where US companies are, including and OpenAI, are chasing trillion dollar valuations. Luz Ding is on a team of Bloomberg reporters who have done a deep dive into this fierce competition from The US and China on AI.

Luz joins us now. I wanna do justice to all the data driven research you guys did. It's really, really cool. Presented as well. But first, I wanna start with this chart, which basically shows how close these two countries are in equivalencies with these AI models. I know you broke it down into different sectors and different types and what which country is better at, but can you just give us the top lines what you found when you dug into capacity and usage of these models and cost?

I know cost is a really big issue. Yes. Of course. Thanks for having me. So, so, yeah, as we can see from the graph that you just showed before, that's the gap between The US and Chinese top models have been narrowing. It was, it it was narrowed from over a year to now just a few months, especially after the recent breakthrough from China's moonshots k m k three model release in July. We are also seeing the usage of Chinese models around the world's increasing.

One of the data points that we use was from Open Router, one of the major AI access platforms. We're seeing the use of Chinese models reached over 60% in July. And the and in the month before, the use of Chinese models surpassed The US models for the first time on the platform. And then it's worth noting that so you don't the users, they don't have to send data to China or Chinese models to use these Chinese to Chinese companies to use Chinese models because most of the Chinese top models are open weight, meaning that these models can be downloaded, modified, and deployed anywhere anywhere needed.

So a lot of the use that we're talking about, especially the use of the models in The US, we're talking about Chinese models deployed in data centers in The US or Canada and run by US companies. You're putting all of these models through their paces, and and my favorite test you did was asking them to create a fictional ecommerce site called Bloomberg, a coffee site, fake ecommerce company. Just trying to explain the difference between how each of these models approach that maybe in terms of what they were able to create and also the cost associated with that when you compare kind of The US models to the Chinese models?

Yeah. So, yeah, we see a lot of benchmarks and rankings and scores, but it's difficult to see, like, how actually, they perform. So we gave, this one prompt, as you said, to build a coffee shop website to some of the leading models. And then what we find, first of all, getting a ecommerce website seems not to be a difficult task for most of the models available right now. It's not challenging for the top ones, but it's not also not challenging for the lower tier ones.

Most of them, we have found are able to build functional websites that we need. But then now we're also seeing, for example, the prices are quite different. For example, for Anthropix model, Fable five, when we asked it to be the website, it was able to deliver a very sophisticated website within an hour, and the bill was about $50. That's already pretty good for a website. But then if you use China's Kimi k three model, then you can do the similar website for just a quarter of the price.

And then if you wanna sacrifice a little bit of the, you know, graphic design and then you wanna wait for a longer hours, you can go even below $5 for the website, and they can still work. So you're gonna have, like, this this sheen version of the website, but it sounds like the sheen version is, like, pretty okay. Yeah. Yeah. Yeah. Yeah. I think in the end, it really up to, like, how how fast you want the results in the first take and how good you want it to be and how much you wanna pay for that perfection and speed.

There are also places that it seems like the Chinese models are actually outdoing, The US models. Can you walk us through the different categories? You looked at coding. You looked at finance. You looked at banking support. Where are the Chinese models really excelling? Yes. So it it depends on how you use it and which industry that, you know, are more adapted to the models. We I don't think from from the graphics that you saw in our story, there are rankings of different ways.

And then if I could remember it correctly, on the finance service sides, there are more Chinese models leading in the top three, whereas in some other categories, are only one Chinese models in the top top top 10. But then Chinese models, it's it's real real advantage is in the price because it can be capable enough to address a lot of the a lot of the questions and tasks that we encounter in real life and also in working places.

So with the capable enough capability and then also the very low price, it's being it's been it's become very irresistible for a lot of companies to to adopt, including some of the top US tech companies like Airbnb, DoorDash, and Coinbase. Lisa, I wanna go back to something that Christina mentioned at the top that this is kind of in the crucible of a couple of these AI companies poisoned to to go public, so Anthropic and and OpenAI.

And it makes me wonder just sort of how important this metric of being the best is. That that if you look at these in side by side, they're both quite good, the Chinese models and and The US models. How important is it from a business perspective, a financial perspective that Anthropic and OpenAI maintain this kind of vanguard position here as they go into their into their IPOs? Yeah. I mean, as of now, OpenAI and and for are still the best building the best models, most capable models, and there's still a lot of money being made by them in the current in the current market.

And then also the data that we saw, they're not a full capture of the all the all of the glow global AI use because OpenRotter is just one platform. There are a lot of other platforms that offer these services and the use of Anthropic and OpenAI's API the the use of Anthropic and OpenAI's service is not trackable because it's with their company. And then I would say that it's hard to say, like, how how, for example of course, if you were the best in the category, you are always you're attracting more investors.

But at the same time, there are also another market that's to be tapped is that those companies in smaller companies or companies in developing countries, they may not be able to pay that high price that Anthropic and OPIA are charging. And then now they look at and then they can choose Chinese models where which which are, like, increasingly capable, and then they offer much lower price. And they they are open ways so you can deploy it locally.

So that's a market that Chinese models are trying to trying to conquer, and they're already succeeding in a lot of places.