NVIDIA China Inference Chip Uses Groq LPUs to Bypass Export Rules

NVIDIA is preparing to re-enter the Chinese market with a specialized AI inference chip that bypasses strict US export controls. This move matters to buyers because it offers a compliant path to high-performance AI computing in a region where standard data center GPUs are banned. The company aims to restart hardware sales in China by late 2025 without violating Washington's restrictions on high-bandwidth memory and advanced packaging.

Groq LPUs hard-bake weights into SRAM for fast inference

The new hardware relies on Groq's Language Processing Units (LPUs) rather than traditional graphics cores. NVIDIA licensed Groq's technology and assets in December 2025 to build this inference-focused architecture. The chip targets large language model workloads and uses a design that hard-bakes AI model weights directly into SRAM. This approach completely bypasses memory cache, resulting in extremely fast inferencing speeds as described by The Information.

• Architecture: Groq Language Processing Units (LPUs)
• Target Market: China
• Release Timeline: End of year
• Compliance: US Export Controls

Manufacturing for this chip appears tied to SMIC, the only Chinese foundry with approximately 7nm process capability. SMIC recently reported a 93.7% capacity utilization rate and $3.01 billion in quarterly revenue. The foundry informed customers it would adjust to fairer pricing as it scales production. This manufacturing base supports the claim that NVIDIA can produce the chip within China's borders.

Chinese tech firms like Alibaba and DFSX are also developing alternative AI chips to compete in this space. The new NVIDIA product aims to comply with US export controls on HBM and packaging. We looked at earlier NVIDIA China-specific products while tracking these regulatory shifts. The chip is expected to be ready by the end of the year according to current reporting.