China Leverages Tech Talent to Boost AI Data Center Boom

China is increasingly combining state-owned data centers with private sector expertise to accelerate the development of its AI infrastructure, as reported by the Financial Times. Local governments are partnering with tech companies to improve the efficiency of AI data centers, a crucial element of China's ambitious goal of becoming a global AI leader.

"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 a Beijing-based chip investor to the Financial Times.

This initiative follows the recent success of DeepSeek, a Chinese start-up that has garnered international attention for developing AI models comparable to those of US rivals like OpenAI and Google, but at a significantly lower cost. DeepSeek's innovative approach, which involves optimizing Nvidia GPUs for maximum computing power, has inspired state-owned data centers to seek similar engineering expertise.

By mid-2024, China had completed or was constructing 250 AI data centers, a multibillion-dollar investment driven by local governments issuing bonds to finance construction and secure necessary chips. However, many of these projects have struggled due to a lack of technical expertise, resulting in wasted chip resources.

To address these challenges, local governments are partnering with tech start-ups. Infinigence AI, a Shanghai-based company backed by HongShan and Qiming Venture Partners, is a prominent example. Infinigence AI rents Nvidia chips to developers building AI applications, generating revenue from the difference between its payments to data center operators and its charges to end-users.

In December, the Shanghai government announced a collaboration between Infinigence AI and SiliconFlow to build a platform for creating AI applications using third-party models hosted at a data center operated by a local subsidiary of China Telecom. Both Infinigence AI and SiliconFlow develop inference engines that accelerate LLMs (large language models) through techniques that adapt pre-trained models to specific data or reduce model size.

Zhejiang-based Merit Interactive is another example of a tech company contributing its engineering expertise. The company has partnered with local officials to construct a large AI data center in Wenzhou. Notably, Merit Interactive is associated with the business partner of DeepSeek founder Liang Wenfeng at his quant trading fund High-Flyer.

The US has a similar ecosystem of "neocloud" companies like CoreWeave, Lambda, and Vultr, which offer pay-as-you-go access to computing clusters in data centers. China's AI industry has thrived despite US export controls on Nvidia's high-end chips, with a thriving black market supplying smuggled chips. State-owned data center operators also have preferential access to Huawei's Ascend AI chips, a major competitor to Nvidia in China's growing inference market.

Both Infinigence AI and SiliconFlow are working to reduce inference costs and improve compatibility between Huawei's Ascend chips and leading Chinese AI models. A recent collaboration between Huawei and SiliconFlow has made DeepSeek's AI models available through Huawei's cloud service.

This public-private partnership approach is crucial for attracting AI clients, as technology companies conduct extensive testing before committing to expensive training runs with data centers.