[News] NVIDIA’s China-Ready LPUs Could Launch as Early as May; H200 Production Restarts

Fresh off unveiling its Groq 3 LPU—built on Samsung’s 4nm for high-speed inference—at GTC, NVIDIA is already eyeing its next move. Reuters reports the chip giant is preparing a China-compliant version of its Groq AI chips, aiming to navigate export curbs and re-enter the Chinese market.

According to one source cited by Reuters, the chips being prepared for China are not downgraded or uniquely designed for the market. Instead, the variant is expected to be adaptable to different system configurations, with Groq-based chips likely to become available as early as May.

As noted by Reuters, while NVIDIA continues to dominate AI training, the inference segment is far more competitive. Major Chinese players, including Baidu, have already developed their own in-house inference chips, the report adds.

According to South China Morning Post, Baidu—which announced in January plans to spin off its Kunlunxin chip unit for a Hong Kong listing—has highlighted the growing importance of the AI inference market. The company’s Kunlunxin, as per the report, is best known for its M100 chip, engineered for large-scale inference optimization.

Meanwhile, competition is heating up, with Chinese players such as Huawei Technologies and Cambricon Technologies also rolling out their own inference-focused chips, the report notes.

Against this backdrop, Reuters reports that NVIDIA is looking to deploy Groq’s chips for inference workloads—powering tasks such as answering queries, generating code, and executing user-driven AI functions.

H200 Production Reignites

Notably, the push to tailor a China-ready version for its LPUs comes as CEO Jensen Huang revealed the company has resumed production of its H200 chips after obtaining export licenses from the Trump administration and securing fresh orders from Chinese customers.

After a long pause in China, NVIDIA is gearing up to deliver its H200 processors to select customers, CEO Jensen Huang told CNBC on Tuesday at GTC. “We have received purchase orders, and we’re in the process of restarting our manufacturing,” he said, adding that “our supply chain is getting fired up.”

CNBC explains that China once accounted for at least one-fifth of NVIDIA’s data center revenue, but the company was effectively blocked after the Trump administration required export licenses for chip sales to China and several other markets—a move that forced NVIDIA to take a $5.5 billion charge.

However, in December, Trump reversed course, allowing NVIDIA to resume H200 chip exports to China under a framework that grants the U.S. a 25% share of revenue, the report notes.

Notably, potential sales to China are not factored into Huang’s projection of more than $1 trillion in revenue from NVIDIA’s Blackwell and Rubin AI chips by the end of 2027, according to South China Morning Post.

Read more

• [News] Key GTC Takeaways: How Samsung, Micron, Intel & More Power NVIDIA’s Vera Rubin and Feynman
• [News] NVIDIA GTC Could Spotlight CPUs for Agent AI with Intel Partnership and CPU Racks Highlighted

(Photo credit: NVIDIA)