TheChek News Flow

The Takeaway

Nvidia plans to begin small-batch shipments of an AI chip specially tailored for Chinese customers by the end of the year, according to two employees, in what would represent a new route to China for the company.

Several Chinese customers have already placed orders for the chip, which is a variant of Nvidia’s language processing unit. The LPU is a new chip developed by Nvidia with technologies licensed from Groq that works alongside Nvidia’s graphics processing units to speed up an AI chatbot’s response times.

The chip complies with U.S. export control rules, according to the employees, but it remains to be seen whether Beijing will allow these orders to go through. The new chip would compete with multiple homegrown chips designed for inference.

Nvidia’s development of a China-specific LPU shows the company hasn’t given up on the China market, despite restrictions imposed by both the U.S. and Chinese governments. The Trump administration has barred Nvidia from selling its most advanced chips to China, though it now permits sales of an older-generation chip, the H200, under strict conditions. Beijing, meanwhile, has blocked or discouraged purchases of even some Nvidia chips Washington has cleared for export. In a May interview with CNBC, Nvidia CEO Jensen Huang said the company had “largely conceded” China’s AI chip market to local chip company Huawei Technologies.

Nvidia unveiled the LPU at the company’s annual GTC conference in March. Huang said then Nvidia would begin shipping LPU to customers in the third quarter. Customers for the LPU include OpenAI and some neocloud providers, or companies that rent computing power built around AI chips.

Nvidia originally designed the LPU to work alongside Rubin graphics processing units as part of Vera Rubin, its next-generation system for running AI models. The GPUs handle most of the model’s workload, while the LPU speeds up parts of the process used to generate each response. Software determines how that work is divided between the two types of chips.

Because Vera Rubin cannot be sold in China, Nvidia modified that software so the LPU can instead work alongside other processors available in the country. The LPU itself didn’t need to be redesigned or have its computing power reduced because it already complies with U.S. export control rules, the employees said.

Before this story published, Nvidia did not respond to repeated requests for a comment, over several days. After the story was published, Nvidia issued a statement saying: “We have no LPU sales in the China market today, and no China-specific LPU product in our roadmap.”

Reuters reported in March that Nvidia was preparing LPU chips for China. Huang subsequently denied the plan in an interview with Punchbowl.

Inference Shortage

The LPU would compete head-on with Huawei’s Ascend 950DT, which is designed for inference and is due to ship in large volumes later this year. Still, demand for inference chips is so strong that there is plenty of unmet demand, which means Nvidia could still seize a share of the market if Beijing allows it.

More importantly, the current supply shortage gives Nvidia a critical window to win back some Chinese customers before they switch entirely to building their AI systems around Chinese silicon.

The inference chip shortage has been affecting Chinese AI labs. Moonshot AI halted new customer subscription sign-ups shortly after releasing the Kimi K3 model as demand overwhelmed the company’s computing capacity. DeepSeek this month significantly raised the prices developers pay to access its latest V4 models.

China’s largest tech companies have tied up much of the production available from domestic chip designers this year and are already securing orders for 2027, according to two people familiar with the situation.

The H200 Flip-Flop

Beijing previously blocked domestic companies from buying Nvidia’s H20, a less powerful processor designed for China, and restricted access to the more capable H200. Chinese officials recently told companies they plan to allow some H200 purchases. Some Chinese companies have since begun receiving the chips in small batches, with deliveries arriving gradually, according to one of the employees.

Yet the blockade has proved porous. Some Chinese companies have obtained Beijing-banned H200s and Nvidia gaming chips through resellers and other intermediaries, according to three people with direct knowledge of the situation, underscoring how difficult it is to suppress demand while the country remains short of computing power.

Nvidia has kept a substantial supply ready. Amkor Technology, a U.S. company that finishes and assembles chips for installation in computers, completed that work for more than 1 million H200s in the second quarter, according to one of the two Nvidia employees. Amkor will continue receiving the main parts of the processors from Taiwan Semiconductor Manufacturing Co., the person added.

Nvidia is separately developing at least two new gaming chips for China, according to one of the employees. To stay within U.S. export limits, the chips will offer less computing power than its leading products. Though intended for gamers, such chips are often modified by Chinese companies to run AI models, The Information reported.