Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
Chinese state-owned data centres flush with artificial intelligence chips are tapping the expertise of technology groups to help bolster their multibillion-dollar investments as part of a nationwide effort to increase adoption of the fast-developing technology.
Local governments have enlisted Merit Interactive and start-ups Infinigence AI and SiliconFlow to develop technology for the surge of AI data centres being created across China, according to people with knowledge of the matter and public documents.
It shows how China is increasingly pooling resources from the private and state sectors — combining local government spending power with start-up tech talent — as part of a push to accelerate its network of AI computing infrastructure.
“The bottleneck at the moment isn’t getting the chips but figuring out how to make them work in a cluster. This is really complicated work,” said one Beijing-based chip investor.
The move comes after Chinese start-up DeepSeek captured the world’s attention last month with the latest in a series of AI model releases that showed similar performances to those of US rivals such as OpenAI and Google, but achieved on what appeared to be a bootstrapped budget.
The Zhejiang-based company used engineering prowess to squeeze as much computing power as possible from its Nvidia GPUs, driving down the cost of model training and so-called inference, the process of calling upon large language models.
State-owned data centres are leveraging this type of private sector engineering talent amid an AI infrastructure boom across China, as the country invests billions in the race against the US to take the lead in the rapidly developing technology.
By mid-2024, 250 AI data centres had been completed or were under construction, according to official data. Local governments have played a central role in bankrolling these multibillion-dollar projects, issuing bonds to finance construction and secure the chips, according to people familiar with the deals. However, they added that many of the projects had suffered from a lack of technical knowhow and had wasted critical chip resources.
Infinigence AI has emerged as one of the main beneficiaries of the state’s push to improve computing performance, signing deals with Ningxia and Tianjin state-run AI data centres.
The Shanghai-based start-up, backed by HongShan and Qiming Venture Partners, rents Nvidia chips to developers building AI applications. It makes money between the difference it pays the data centre operator and what it charges the end customer access to the chips.
In December, the Shanghai government announced a collaboration with Infinigence AI and start-up SiliconFlow to build a platform for customers to create AI applications on third-party models hosted at a data centre operated by a local offshoot of the state-owned China Telecom.
Both companies build inference engines that speed up LLMs through techniques that take pre-trained models and better adjust them to data or by compressing the size of AI models.
One investor said such collaborations were crucial to courting AI clients. Technology companies run tests before signing contracts with data centres to ensure no disruption during expensive training runs.
Zhejiang-based Merit Interactive has also been tapped for its engineering expertise. A Zhejiang media outlet publicised a meeting between Merit Interactive executives and local officials, where the company was singled out for its role in constructing a large AI data centre in the city of Wenzhou.
The group is associated with the business partner of DeepSeek founder Liang Wenfeng at his quant trading fund High-Flyer, according to a person with knowledge of the matter.
According to local media reports, Xu Jin, who co-founded High-Flyer with Liang, used to work at Merit Interactive as a technical director. Merit Interactive has business interests spanning an import and export business for tech hardware to investments in high-tech start-ups.
The US has a similar set of “neocloud” companies, including CoreWeave, Lambda and Vultr, which manage and rent out access to computing clusters in data centres, giving companies the ability to use the chips on a pay-as-you go basis.
China has been able to develop its AI industry despite Washington tightening its export controls on Nvidia’s high-end chips. A thriving black market has emerged for smuggled chips, with industry insiders noting that prices of highly sought-after H100s have decreased in recent months as the number of illicit chips brought into China has swelled.
State-owned data centre operators have also had preferential access to Huawei’s Ascend AI chips, the primary challenger to Nvidia in China for the growing inference market.
Like DeepSeek, both Infinigence AI and SiliconFlow have been working on driving down inference costs to make it cheaper to use LLMs. They are also working with Huawei to make the Ascend chips compatible with leading Chinese AI models for inference.
This month, Huawei and SiliconFlow announced a collaboration to make DeepSeek’s AI models available through the telecom giant’s cloud service. According to a person familiar with the matter, DeepSeek collaborated with the two companies during the Chinese new year holiday to deploy its models on Huawei chips.
Additional reporting by Michael Acton in San Francisco
