Nvidia Plots China Comeback With Groq-Derived Inference Chip as Huang Plans Beijing Visit

Nvidia is preparing a return to the Chinese market it has all but lost. The company plans to ship small volumes of a specialized AI chip to Chinese customers by the end of 2026, according to reporting from The Information — a product aimed squarely at China's fast-growing AI inference market.

A different kind of chip

The new part is not a cut-down GPU like the ill-fated H20. It is a modified version of Nvidia's language processing unit, built with technology licensed from Groq, and designed to work alongside existing processors to accelerate AI chatbot responses. Inference — running models, rather than training them — is where Chinese demand is exploding, and where US export rules leave Nvidia the most room to maneuver.

CEO Jensen Huang is planning to visit Beijing ahead of the chip's launch, a signal of how much the company still values a market that once accounted for roughly a fifth of its data center revenue.

Racing a closing door

The comeback attempt faces headwinds from both sides of the Pacific. Washington has spent the year tightening the screws: after moving in May to halt shipments to Chinese firms' overseas subsidiaries, the administration has been working to close the cloud-access loophole that let Chinese labs rent restricted silicon abroad — a route that drew White House accusations against Moonshot AI over alleged GB300 access via Thailand.

Beijing, meanwhile, is pulling in the opposite direction. China aims to triple domestic AI chip output, and Huawei is scaling its Ascend 950 line to fill the vacuum. Bernstein forecasts Nvidia's China AI chip share will collapse from about 40 percent to just 8 percent in 2026, with Huawei climbing toward 50 percent.

Whether a Groq-derived inference part can claw back meaningful share against that tide — and survive both governments' scrutiny — may be the defining question of Nvidia's next fiscal year in Asia.