Chinese AI companies are reducing costs by optimising hardware and using smaller data sets.,Strategies like the "model-of-expert" approach help achieve competitive models with less computing power.,Companies like 01.ai and ByteDance are making significant strides despite US chip restrictions.
In the rapidly evolving world of artificial intelligence (AI), Chinese companies are making waves with innovative strategies to drive down costs and create competitive models. Despite facing challenges like US chip restrictions and smaller budgets, these companies are proving that creativity and efficiency can overcome significant hurdles.
The Cost-Cutting Revolution
Chinese AI start-ups such as 01.ai and DeepSeek are leading the charge in cost reduction. They achieve this by focusing on smaller data sets to train AI models and hiring skilled but affordable computer engineers. Larger technology groups like Alibaba, Baidu, and ByteDance are also engaged in a pricing war, cutting "inference" costs by over 90% compared to their US counterparts. This focus on cost-efficiency is a key trend in the broader APAC AI in 2026: 4 Trends You Need To Know landscape.
Optimising Hardware and Data
Beijing-based 01.ai, led by Lee Kai-Fu, the former head of Google China, has successfully reduced inference costs by building models that require less computing power and optimising their hardware. Lee emphasises that China's strength lies in creating affordable inference engines, allowing applications to proliferate. This approach contrasts with some Western firms that are Running Out of Data: The Strange Problem Behind AI's Next Bottleneck.
"China’s strength is to make really affordable inference engines and then to let applications proliferate." - Lee Kai-Fu, former head of Google China
"China’s strength is to make really affordable inference engines and then to let applications proliferate." - Lee Kai-Fu, former head of Google China
The Model-of-Expert Approach
Enjoying this? Get more in your inbox.
Weekly AI news & insights from Asia.
Many Chinese AI groups, including 01.ai, DeepSeek, MiniMax, and Stepfun, have adopted the "model-of-expert" approach. This strategy combines multiple neural networks trained on industry-specific data, achieving the same level of intelligence as a dense model but with less computing power. Although this approach can be more prone to failure, it offers a cost-effective alternative.
Navigating US Chip Restrictions
Despite Washington's ban on exports of high-end Nvidia AI chips, Chinese companies are finding ways to thrive. They are competing to develop high-quality data sets to train these "experts," setting themselves apart from the competition. Lee Kai-Fu highlights the importance of data collection methods beyond traditional internet scraping, such as scanning books and crawling articles on WeChat. The broader implications of such restrictions have been discussed in relation to Huang's dire warning on US-China tech war.
"There is a lot of thankless gruntwork for engineers to label and rank data, but China — with its vast pool of cheap engineering talent — is better placed to do that than the US." - Lee Kai-Fu
"There is a lot of thankless gruntwork for engineers to label and rank data, but China — with its vast pool of cheap engineering talent — is better placed to do that than the US." - Lee Kai-Fu
Achievements and Rankings
This week, 01.ai’s Yi-Lightning model ranked joint third among large language model (LLM) companies, alongside x.AI’s Grok-2, but behind OpenAI and Google. Other Chinese players, including ByteDance, Alibaba, and DeepSeek, have also made significant strides in the rankings. This competitive environment is also spurring innovation in areas like Free Chinese AI claims to beat GPT-5.
Cost Comparisons
The cost for inference at 01.ai’s Yi-Lightning is 14 cents per million tokens, compared to 26 cents for OpenAI’s smaller model GPT o1-mini. Meanwhile, inference costs for OpenAI’s much larger GPT 4o are $4.40 per million tokens. Lee Kai-Fu notes that the aim is not to have the "best model" but a competitive one that is "five to 10 times less expensive" for developers to use. This focus on cost-effectiveness is a significant driver in the global AI market, as detailed in reports like the Stanford AI Index Report 2024 AI Index Report 2024.
The Future of Chinese AI
China's AI industry is not about breaking new ground with unlimited budgets but about building well, fast, reliably, and cheaply. This approach is not only cost-effective but also fosters a competitive environment that encourages innovation and efficiency.
Comment and Share:
What innovative strategies do you think will shape the future of AI in Asia? Share your thoughts and experiences with AI and AGI technologies in the comments below. Don't forget to Subscribe to our newsletter for updates on AI and AGI developments.
Latest Comments (4)
It's truly impressive to see how our companies are pushing the envelope with such clever strategies. The idea of optimising hardware, rather than just throwing more computing power at the problem, feels very much in line with our ethos of *节约* (frugality) and making the most of what we have. It makes me wonder, though, how these "model-of-expert" approaches are being scaled for really vast and complex applications. Are we seeing these smaller, specialised models communicating seamlessly enough to tackle, say, a country-wide logistics optimisation challenge, or are there still hurdles in stitching them all together effectively for truly grand-scale problems?
This article on Chinese AI's cost-efficient innovations is really timely, even with the shift in the global tech landscape. I'm curious, how exactly are they optimising hardware to achieve these competitive models? Is it more custom chip design or ingenious software-hardware co-optimisation that's making the biggest difference?
This is quite fascinating, eh? I've noticed a few of these Chinese tech innovations popping up on my feeds lately. Need to properly dive into how they're keeping costs down.
It's interesting to see this perspective on Chinese AI. I remember reading about their hardware optimisation years ago, even before some of these "model-of-expert" ideas became mainstream. Frankly, the efficiency they achieve is quite something. We in Europe could surely learn a trick or two from their approach to keeping costs down. C'est vrai.
Leave a Comment