China gets closer to finding its own Nvidia

A NEW China-buys-China narrative is taking shape as Beijing steps up its tech rivalry with the US. The world’s second-largest economy not only wants to build generative AI models, but power them with its own hardware, redrawing a supply chain dominated by Nvidia.

There certainly is an opening after Nvidia reportedly halted production related to its H20 AI chip tailor-made for the Chinese market. Beijing had told tech companies to stop buying them due to national security concerns. Last week, DeepSeek released an upgrade to its flagship V3 model to accommodate the next generation of homegrown chips.

Nvidia still dominates the supply chain with chips essential for the complicated work of training AI models. But as the industry evolves, a bigger market is in selling chips used for inference, which creates responses from pre-trained models and produces texts as well as images for people who use generative AI tools.

These chips are less complex. Nvidia’s H20 model sold in China, which US Commerce Secretary Howard Lutnick said was “not even our third best,” is arguably more suited for this purpose. So it is not hard to imagine that over time, Chinese companies that design application-specific integrated circuits, which are less powerful but more cost-efficient and specialised, can close the performance gap with Nvidia.

The question is how. Beijing has tried for years to fund tech advancement and foster national champions. But past practices have led to waste, debt accumulation and scandals. The bankruptcy of Tsinghua Unigroup in 2021, the closest thing China had to Samsung Electronics, and the subsequent corruption investigation at the China Integrated Circuit Industry Investment Fund, known as the “Big Fund,” come to mind.

These past failures only make the AI chip designer Cambricon Technologies’ success all the more heartening. Founded in 2016, it went public four years later on the Shanghai Stock Exchange STAR board, a pilot programme championed by President Xi Jinping to speed up the listing process of tech startups.

Cambricon has become a stock market darling, precisely because investors recognise China has a good chance to take market share in inference chips. As a result, the company can rely on stock sales to finance future R&D spending. It recently announced a plan to raise nearly 4 billion yuan (S$720 million) for AI chip and software development. Despite a sharp jump in sales, the young startup is still burning cash.

In retrospect, fundraising via the stock market is a lot simpler than the labyrinth of financial support privately-held Unigroup received from various state backers, including the Big Fund. It was never clear who was really in charge and which entity would ultimately pick up the hefty bill of building next-generation memory chip plants. It’s no surprise that Unigroup went bust.

China is rethinking how it will fund the next round of AI development. Increasingly, the government is going for “early, small, hard,” favouring venture capital in small companies working on core tech, according to Gavekal Dragonomics’ Tilly Zhang. Cambricon is a good example. As of June, 40-year-old founder Chen Tianshi had a 29 per cent stake, while an affiliate of the Chinese Academy of Sciences, where he used to work as a researcher, held 16 per cent.

Meanwhile, less prosperous regions are being discouraged from funding hard tech. Last month, Xi issued an unusually blunt warning to local governments against over-investment in AI.

These are all good developments. But most importantly, young startups should be encouraged to go public as soon as possible. As investors look for their own version of Nvidia, they will naturally do due diligence for the government. They will find and foster real national champions. BLOOMBERG