Stay up so far with free updates
Simply log in Artificial intelligence myFT Digest – delivered straight to your inbox.
Chinese artificial intelligence corporations are cutting costs to create competitive models as they contend with U.S. chip limitations and smaller budgets than their Western counterparts.
Startups like 01.ai and DeepSeek have lowered their prices by adopting strategies resembling specializing in smaller data sets to coach AI models and hiring low cost but expert computer engineers.
Larger technology giants resembling Alibaba, Baidu and ByteDance have also waged a price cutting war to scale back “inference” costs, the worth of using large language models to generate a solution, by greater than 90 percent and to a fraction of the worth offered by US -Colleagues.
That's despite Chinese corporations having to contend with Washington's export ban on the highest-quality Nvidia AI chips, that are seen as crucial to developing probably the most advanced models within the United States.
Beijing-based 01.ai, led by former Google China chief Lee Kai-Fu, said it has reduced inference costs by constructing a model trained on smaller data sets that require less computing power , and optimized its hardware.
“China’s strength is in making really inexpensive inference machines after which enabling applications to spread,” Lee told the Financial Times.
This week, 01.ai's Yi Lightning model tied with x.AI's Grok-2 for third place amongst LLM corporations, but behind OpenAI and Google in a rating published by researchers at UC Berkeley SkyLab and LMSYS.
The evaluations are based on users evaluating the responses of various models to queries. Other Chinese players, including ByteDance, Alibaba and DeepSeek, have also made it to the LLM rankings.
The cost of inference for 01.ai's Yi-Lightning is 14 cents per million tokens, in comparison with 26 cents for OpenAI's smaller GPT o1-mini model. Meanwhile, the inference cost for OpenAI's much larger GPT 4o is $4.40 per million tokens. The variety of tokens used to generate a response will depend on the complexity of the query.
Lee also said that Yi-Lightning cost $3 million for “pre-training,” initial model training that may then be refined or customized for various use cases. This is a small fraction of the price that OpenAI quotes for its large models. He added that the goal isn’t to have the “best model,” but quite a competitive model that’s “five to 10 times cheaper” for developers to construct applications.
Many Chinese AI groups, including 01.ai, DeepSeek, MiniMax and Stepfun, have adopted a so-called “expert model” approach, a method first popularized by U.S. researchers.
Instead of coaching a “dense model” suddenly on an enormous database that has aggregated data from the Internet and other sources, the approach combines many neural networks trained on industry-specific data.
Researchers view the expert model approach as a key method for achieving the identical level of intelligence as a dense model, but with less computing power. However, the approach will be more vulnerable to errors because engineers must orchestrate the training process across multiple “experts” quite than in a single model.
Faced with the problem of ensuring a gradual and sufficient supply of high-end AI chips, Chinese AI players have been competing over the past yr to develop the very best quality data sets to coach these “experts” and differentiate themselves from the competition.
Lee said 01.ai has approaches to data collection that transcend the standard approach to searching the Internet, including scanning books and crawling articles on the messaging app WeChat that aren’t accessible on the open web.
“It's loads of thankless work for engineers to label and classify data,” he said, but added that China – with its vast pool of low cost engineering talent – is healthier suited to do that than the US.
“China's strength isn’t in doing the very best groundbreaking research that nobody has done before, with no limits on budget,” Lee said. “China’s strength is to construct well, construct fast, construct reliably and construct low cost.”
