A report that Nvidia is quietly preparing a China comeback had a short shelf life. On Thursday, The Information, citing two employees, said the company planned small year-end shipments of an AI inference chip tailored for Chinese buyers, with several customers already placing orders. Within hours, Nvidia denied it, telling Reuters it has no LPU sales in China and no China-specific version of the product on its roadmap. The stock, already down for a third straight day, slipped about 0.5%.
That contradiction is the story here, so it is worth laying out what was claimed and what was disputed rather than treating the report as settled fact.
What the report described
According to The Information, the chip would be a variant of Nvidia’s language processing unit, a co-processor built on technology Nvidia licensed from Groq last year in a deal worth roughly $17 billion to $20 billion. Structured as a non-exclusive licensing-and-talent arrangement rather than an outright acquisition, it handed Nvidia the IP without a lengthy antitrust review. The LPU is designed to work beside GPUs and speed up the token-generation part of running a chatbot, and Nvidia unveiled it at its GTC conference this year as part of the Vera Rubin system.
The reported workaround was clever. Because Vera Rubin cannot be exported to China, engineers were said to have modified the LPU’s software so it could pair with other processors available there, letting the fully compliant chip stand on its own without dialing down its performance. That is the part Nvidia now says is not happening, at least not as a planned product.
Why the idea is plausible even if the report is wrong
Denials aside, the strategic logic is real, which is probably why the report gained traction. Chinese firms are short on inference capacity right now. Moonshot AI recently paused new sign-ups over bottlenecks, DeepSeek raised developer pricing under load, and much domestic chip output is reportedly spoken for through 2027. That leaves a gap even as local suppliers race to fill it.
Nvidia’s own position on China has been muddled by policy whiplash. Washington blocks its top accelerators but has cleared limited sales of the older H200, while Beijing has alternately discouraged and, more recently, signalled tolerance for buying cleared US chips. Jensen Huang has said Nvidia largely ceded China to Huawei, whose Ascend 950DT inference chip is the domestic option any Nvidia product would have to beat. Huawei’s silicon still trails Nvidia’s best on raw performance, but it is far cheaper and, crucially, actually available.
So a compliant inference chip that keeps Chinese customers on Nvidia’s software before they migrate to Huawei would make obvious commercial sense. Whether Nvidia is actually building one is the question its denial leaves open. On the manufacturing side, at least, the machinery is humming: The Information reported that Amkor packaged more than a million H200 chips last quarter using TSMC-made components, a reminder of how much Nvidia hardware is still flowing toward China through approved channels.
So is this a preview of Nvidia’s next China move or a report that got ahead of itself? For now the honest answer is that a credible outlet and the company itself are telling different stories. The commercial incentive is unmistakable. The confirmation is not, and until Nvidia says otherwise, the year-end China chip belongs in the rumour column.
This website uses cookies.