China to get improved AI chips from Nvidia, finding a new loophole


Despite new US restrictions on the exports of high-end AI chips to China, there are reports that Nvidia has developed an improved series of chips for AI computing that will meet the new rules.

According to the China Star Market report, the three latest Nvidia chips, called HGX H20, L20 PCle, and L2 PCle and intended for the Chinese market, are an improvement on the H100, one of the world's most advanced chips ever built.

Nvidia may announce the new chips after November 16th at the earliest, when the new licensing requirements in the US will take effect. That’s less than a month after the new rules were introduced.

The news came from people in the industry chain familiar with the matter. According to Chinese sources, domestic manufacturers may get the new chip variants in the upcoming days.

The new restrictions by the US Department of Commerce were supposed to close the loopholes that allowed tech giants to offer China slightly modified high-end chips used for AI computing. Nvidia, dominating the market with its most advanced chips, has been instructed to stop shipments of A800 and H800 chips immediately.

Another source also said that Intel plans to launch an improved version of Gaudi 2 accelerators, which also falls under new licensing requirements.

Semiconductor research and consulting firm SemiAnalysis was surprised that Nvidia still found a way to slip through US restrictions and claims that Nvidia already has product samples that will go into mass production soon.

Detailed specifications reveal that “one of the China-specific GPUs is over 20% faster than the H100 in LLM inference and is more similar to the new GPU Nvidia is launching early next year than to the H100!”

Wells Fargo analysts expect investors to question whether Nvidia is being too aggressive in its efforts to circumvent US restrictions, which could result in further US government moves. According to Reuters, Nvidia has commanded more than a 90% share of China's $7 billion AI chip market.