
"Half a decade of US trade policy aimed at denying China access to America's most potent semiconductor tech has only served to spur China to develop homegrown alternatives. The Trump administration's decision on Monday to reverse course and allow the sale of Nvidia's H200 - still one of the GPU giant's most potent AI chips - to Chinese customers, in exchange for a 25 percent cut of revenue from sales, is unlikely to change that."
"As evidence of this, on Tuesday the Financial Times reported that the Chinese government planned to restrict access to imported H200s. Chinese leaders have already been moving in this direction for months. After the Trump administration reversed a sales ban on Nvidia's made-for-China H20 accelerators, Beijing accused the chipmaker of conspiring with Uncle Sam to plant backdoors in its chips. Nvidia has vehemently denied the insinuation."
"Nvidia's technical superiority has led to market dominance among developers of AI models, including those in China. Following the success of DeepSeek R1 earlier this year, the Chinese AI darling reportedly faced pressure to train its next model using Huawei-designed accelerators. However, unstable chips, glacially slow interconnects, and a dodgy software stack ultimately derailed the effort, and DeepSeek has reportedly re-tasked Huawei's Ascend accelerators for inferencing workloads."
U.S. efforts to block China's access to advanced semiconductor technology over the past five years have prompted China to accelerate development of domestic AI accelerators. A recent U.S. decision to permit sales of Nvidia's H200 to Chinese customers for a revenue cut is unlikely to curb China's drive for technological independence. Beijing has moved to restrict imported H200s, accused Nvidia of collusion over backdoors, and directed firms to pursue homegrown alternatives. Nvidia's technical lead has posed challenges for domestic substitutes, but Chinese accelerators and systems, including offerings from Huawei, are improving rapidly.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]